Submit to this Journal Review for this Journal Propose a Special Issue

Article Menu

Share Help Cite Discuss in SciProfiles

Open AccessArticle

Peer-Review Record

Detection of Water Hyacinth (Eichhornia crassipes) in Lake Tana, Ethiopia, Using Machine Learning Algorithms

Water 2023, 15(5), 880; https://doi.org/10.3390/w15050880

by Getachew Bayable^1,2,*, Ji Cai³, Mulatie Mekonnen², Solomon Addisu Legesse²

, Kanako Ishikawa³, Hiroki Imamura¹ and Victor S. Kuwahara¹

Reviewer 1: Anonymous

Reviewer 2:

Sk Ajim Ali

Reviewer 3:

Zahid Ullah

Reviewer 4: Anonymous

Reviewer 5:

Nicholas A. Friedenberg

Water 2023, 15(5), 880; https://doi.org/10.3390/w15050880

Submission received: 24 November 2022 / Revised: 29 January 2023 / Accepted: 2 February 2023 / Published: 24 February 2023

(This article belongs to the Special Issue Research on Applications of Remote Sensing and Geographic Information Systems (GIS) in Water Resources)

Round 1

Reviewer 1 Report

Please see the attached file.

Comments for author File: Comments.pdf

Author Response

Dear Editor and Reviewer

Thank you very much for the constructive and useful comments. We have put extra effort to address the comments provided by the reviewer. Below, you can find a direct copy of the comments from the reviewer followed by our response in “Italic”. The “red” colored words, phrases and sentences are those included in the manuscript as a response to the comments. Changes made during revision were also highlighted in red color in the manuscript

Reviewer #1:

Through the comparison of the experiment results, the paper concludes that RF (Random Forest) classification algorithm is an accurate water hyacinth detection algorithm, but the reason why RF is superior to other detection algorithms is not explained from the theoretical level.

Response: Random Forest (RF) is considered to be more robust and less sensitive to hyperparameters, noise, and sample size compared to Support Vector Machine (SVM) and Classification and Regression Trees (CART). RF is capable of handling high-dimensional data, reducing overfitting through averaging predictions of many decision trees, and can also handle missing data, unbalanced classes, and outliers.

RF is capable to handle high-dimensional data, reduce overfitting by averaging the predictions of many decision trees, and it is also capable to handle missing data, unbalanced classes and outliers. RF is superior to SVM and CART. Because RF is less sensitive to hyperparameters and noise in the data compared to Support Vector Machines (SVM). CART is sensitive to sample size. Therefore, the RF classifier is more robust and less affected by parameters, noise and sample size when compared to SVM and CART methods (page 14).

Discussion

----------------------------------------------------------------------------------------------------------------------------------------- Each machine-learning algorithm has its own set of advantages and disadvantages [36]. SVM is hyperparameter sensitive, whereas RF is more robust and less affected by parameters and noise [69,70]. On the other hand, the performance of SVM is good when few training sample points are available [71]. CART is also sensitive to sample size [36]. In both Landsat 8 and Sentinel 2 datasets, RF outperformed SVM and CART whereas CART and SVM algorithms performance were similar in terms of F1-score, overall accuracy, and Kappa coefficient for detection and mapping during all seasons. This difference is attributable to the fact that the RF classifier is less sensitive to noise and parameters(including sample size).--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------.

The experimental data only selects the data of four months of October 2021, January 2022, March 2022 and June 2022 to represent the four seasons of spring, summer, autumn and winter, with a small amount of data and may have accidental errors due to climate and other reasons.

Response: In addition to October 2021, January 2022, March 2022 and June 2022, monthly median composite of Sentinel 2 images from October 2022 and a twenty-day median composite of Sentinel 2 images from January 2023 were also used for comparison, and to quantify the E. crassipes spatial coverage during autumn (2022) and winter (2023) seasons, respectively (page 4).

The main work of this paper is to use random forest (RF), support vector machine (SVM), classification and regression tree (CART) machine learning algorithms to detect and recognize the images of Landsat 8 and Sentinel 2 satellites. At the same time, the paper lacks the description of classification detection algorithm and process, and the innovative point is a little insufficient.

Response: Description of classification detection algorithm and process have been included and improved as per the comments given by the reviewer (Page 5&6).

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Three non-parametric machine learning methods available in the GEE platform were used and chosen for their ability to handle outliers in the training data. A detailed explanation of each classifier and process is given as follows:

Classification and Regression Tree (CART): CART is a binary decision classification tree [39]. The algorithm enables simple decision-making in logical if-then scenarios. It recursively divides the feature space into smaller regions, or nodes, based on the values of one or more input features by selecting the feature and the threshold value that will produce the most "pure" child nodes. CART is simple to interpret and understand, but is heavily influenced by the sample size employed in each class. Its effectiveness is also hindered by high-dimensionality data, which results in complex tree architectures [36]. CART can be prone to overfitting, especially when the tree is deep and complex. To deal with overfitting, the data tree can be pruned by setting a minimum number of observations per leaf. In this study, cross-validation technique was used to determine the optimal number of observations per leaf, and a value of 10 produced high accuracy, resulting in the minimum number of observations per leaf being set to 10. The ee.Classifier.smileCart () function is part of Google Earth Engine (GEE) JavaScript API and creates a CART classifier. The ee.Classifier.smileCart () function was used to train the CART classifier with a given set of training data and feature bands. The classify () function was used to apply the trained classifier to the target imagery to classify each pixel of the image into one of the predefined classes.

Random Forest (RF): RF method was first introduced by Breiman [43]. RF is an ensemble machine learning method that uses multiple CART trees to make predictions. In RF, multiple trees are created by selecting a random subset of the data, and combined to give a more accurate and stable prediction [43]. Because of the randomness of sample and feature selection, RF has the advantages of strong noise resistance, generalization ability, and strong processing ability for high-dimensional data without the need to manually select features[37]. The number of trees in the 50 – 500 range performed better with higher accuracy for RF classification. In the present study, 300 trees produced good results, thus, the number of trees was set to 300 for all datasets. The ee.Classifier.smileRandomForest () function is part of the GEE JavaScript Application Programming Interface (API) and creates a RF classifier. The function ee.Classifier.smileRandomForest () was used to train the RF classifier, and then classify () function was used to apply the trained classifier to the target imagery.

Support Vector Machine (SVM): SVM is a supervised classification algorithm that formulates a classification function based on an optimal hyperplane to separate classes from training data based on the process that is performed [56]. SVM can help to determine the locality of decision boundaries or hyperplanes between classes or land cover types [38,57]. It is a binary classifier that can assess pixels from known training samples into one or two possible classes, where the distance of each class from the data points in the training data to the optimal hyperplane is maximized [58,59]. This method helps to reduce the misclassifications that occur during training [60]. The most representative samples could be obtained by finding a hyperplane between classes. The cost parameter (C), gamma, and kernel functions are the most important parameters for selecting support vectors [36]. A grid search technique was used to determine the optimal values for the C and linear kernel in this method. The C has a significant impact on SVM performance and support vector selection. The linear kernel is recommended for large datasets because it is computationally efficient and can handle high-dimensional data [61]. For SVM classification, the C-SVC method with a cost parameter of 10 and a linear kernel type is used. The function ee.Classifier.libsvm() is used to create a support vector machine (SVM) classifier in Google Earth Engine (GEE) JavaScript API. The ee.Classifier.libsvm() function is a wrapper for the LIBSVM library, which is a widely used library for SVM classification and regression. The function ee.Classifier.svm() was used to train the SVM classifier, and then classify ()function was used to apply the trained classifier to the target imagery.

Response: The innovative points have been improved in the manuscript (page 2).

------------------------Previous studies have utilized different classifiers, such as the maximum likelihood classifier (MLC) [1,14,19], decision tree [19], and random forest (RF) [41] to determine the spatial and temporal distribution of E. crassipes in Lake Tana. However, most studies on Lake Tana focused only northeastern part and used Level 1 Top of Atmosphere (TOA) satellite imagery without any atmospheric correction and have only used spectral bands. For example, Dersseh et al. [41] used the RF method to classify and estimate areal cover of E. crassipes from Sentinel-2 MSI Level-1C TOA data, but focused within the lake shoreline boundary without atmospheric corrections. Similarly, Worqlul et al. [1] used the MLC method on PlanetScope Level 1 Top of Atmosphere (TOA) products without considering atmospheric correlation in their analysis. However, atmospheric corrections are crucial for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Besides, Asmare et al. [19] employed a decision tree algorithm for E. crassipes detection in Landsat 8 images, but did not evaluate the classifier's accuracy, but it is crucial for evaluating the performance of a classifier. To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

-------------------------------

The explanation of the experimental pictures should be further added. For example, the black circle in Figure 3 and Figure 4 of the experimental part does not show the annotation meaning in the legend. The explanation should be further added in the discussion part too

Response: Black oval shape in the figure represents the area of misclassification. The explanations of black oval have been included in the figure caption (from Figure 3-Figure 6) and in the result descriptions too (Page 9 & 10)

The Figure 3 and Figure S1-3 was mentioned by author in line 202 of Section 3.2, but Figure S1-3 did not exist in the paper. In addition, the picture in line 278 was labeled as Figure 7, but Figure 5 and Figure 6 don’t exist in the paper.

Response: Figures S1-3, which were first, included in the supplementary material but have now been included in the main manuscript as per the comments given by the reviewer and changed to Figures 4-6 as shown below (Page 9 &10)

Regarding to Figure 7, we made a mistake, we wrote Figure 7 instead of Figure 4, it was supposed to be Figure 4 in the first paper, but now it should be Figure 7

Figure 4: Machine learning classifier comparison in winter (January 2022): Sentinel 2 (a) RF, (b) CART, (c) SVM; Landsat 8 (d) RF, (e) CART, (f) SVM land use/cover map. Black oval shape in the figure represents the area of misclassification.

Figure 5: Machine learning classifier comparison in spring (March 2022): Sentinel 2 (a) RF, (b) CART, (c) SVM; Landsat 8 (d) RF, (e) CART, (f) SVM land use/cover map. Black oval shape in the figure represents the area of misclassification.

Figure 6: Machine learning classifier comparison in summer (June 2022): Sentinel 2 (a) RF, (b) CART, (c) SVM; Landsat 8 (d) RF, (e) CART, (f) SVM land use/cover map. Black oval shape in the figure represents the area of misclassification.

The details of theory and algorithm need to be described. For example, the RF used to estimate the spatial coverage in Section 3.3 only lists the results and does not introduce the specific method of calculating the spatial coverage of water hyacinth. The formula of the experimental evaluation index mentioned in line 181 of the article is not exist in the paper

Response: The spatial coverage (area) of E. crassipes and other land use/cover in the study area was calculated using grouped reducer and reduceRegion () function on the Google Earth Engine (GEE) platform and the result was exported as a table for further analysis. (Page 7)

The formulas of the experimental evaluation index have been included in the manuscript as shown below (Page 6 &7).

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------Machine learning classifier performance was evaluated by comparing the accuracy of classified images with the most commonly used quality metrics such as overall accuracy (OA) (Eq.1) [63] and Kappa coefficient (K) (Eq.2) [36].

Overall accuracy is a measure of how well a reference sample is classified by comparing the number of correctly classified samples to the total number of samples. It gives a percentage of the reference sites that were correctly mapped out from the entire data set [64].

Kappa coefficient (K) is determined where r represents the number of rows and columns in the error matrix. Xii represents the number of observations in row i and column i, Xi+ represents the marginal total of row i, X+i represents the marginal total of column i, and N represents the total number of observations. The Kappa coefficient measures the degree to which accuracy in a classification system exceeds what would be achieved by random chance.

Furthermore, the user accuracy (UA; Eq.3), producer's accuracy (PA; Eq.4), and F1-score (Eq.5) at the class level were used [63].

The F1-score is the harmonic mean of producer and user accuracy and can be used to assess accuracy at the class level [65]. The F1-score is the best performance metric and widely used, giving equal importance to both PA and UA by combining them into a single model performance metric [63,66,67].

The spatial coverage (area) of E. crassipes and other land use/cover in the study region was calculated using grouped reducer and reduceRegion () function on the Google Earth Engine (GEE) platform and the result was exported as a table for further analysis. Moreover, spectral reflectance curves of Sentinel 2 images were used to determine the best bands for distinguishing E. crassipes from other land use/cover types. Moreover, feature importance of the RF classifier was assessed using the explain () method in GEE for different seasons in Sentinel 2 image to evaluate the influence of spectral bands and indices on the classifier's accuracy. Feature importance values indicate how much each input feature, such as spectral bands or indices, contributed to the predictions made by the RF classifier.-----------------------

In the experimental part of the paper, the author compares the accuracy of RF, SVM, and CART algorithms in four seasons. It would be better to supplement the detection algorithms in recent years for comparison.

As shown in the Table S3 and Figure S&S2 below, the RF algorithm slightly outperformed the SVM and CART algorithms in autumn (2022) and winter (2023). However, the SVM and CART algorithms performed similarly in both the autumn (2022) and winter (2023).

Figure S1: Machine learning classifier comparison in autumn (October 2022): Sentinel 2 (a) RF, (b) CART, (c) SVM. Black oval shape in the figure represents the area of misclassification.

Figure S2: Machine learning classifier comparison in winter (January 2023): Sentinel 2 (a) RF, (b) CART, (c) SVM. Black oval shape in the figure represents the area of misclassification.

Table S3. Performance of RF, CART and SVM methods

Season	Sentinel 2 MSI
	Classifier		Overall Accuracy (%)	Kappa Coefficient		F1 Score (%)
October 2022 (Autumn)	RF	99.7			1	100
	SVM	97.7			0.97	98.9
	CART	97.2			0.96	97.2
January 2023 (Winter)	RF	99.5			0.99	100
	SVM	98.9			0.98	99
	CART	98.4			0.98	98.9

Author Response File: Author Response.pdf

Reviewer 2 Report

Title of the manuscript - "Water Hyacinth (Eichhornia Crassipes) Detection Using Machine Learning Algorithms in Lake Tana, Ethiopia"

I have reviewed the manuscript entitled “Water Hyacinth (Eichhornia Crassipes) Detection Using Machine Learning Algorithms in Lake Tana, Ethiopia”. In this article, the authors highlighted an important aspect i.e. water hyacinths (WH), which is a cause of water degradation and biodiversity loss. So, the authors examined the ability of random forest (RF), support vector machine (SVM), and classification and regression tree (CART) machine learning algorithms and Landsat 8 and Sentinel 2 images to detect WH on Google Earth Engine (GEE) platform. Satellite remote sensing data and machine learning algorithms have efficiency to extract accurate results that the authors have proved. Thus, the article presents an interesting topic. But I have found some serious issue and by considering this issue I am not able to accept this manuscript:

Comments:

1. I have found that there are numerous recent studies on Water Hyacinth (Eichhornia Crassipes) in Lake Tana, Ethiopia, using multiple machine learning models. Some of them are as follows:

Detecting spatiotemporal expansion of water hyacinth (Eichhornia crassipes) in Lake Tana, northern Ethiopia, published in Journal of the Indian Society of Remote Sensing, 48(5), 751-764.

Water Quality and Water Hyacinth Monitoring with the Sentinel-2A/B Satellites in Lake Tana (Ethiopia), published in Remote Sensing, 14(19), 4921.

Water Hyacinth (Eichhornia crassipes) Detection Using Coarse and High Resolution Multispectral Data, published in Drones, 6(2), 47.

So, what is the novelty of this study? As there are multiple studies on the same aspect, so what is the relevance of this study?

2. I have found that the authors only considered the LULC of autumn, winter, spring, and summer seasons for the year October 2021, January 2022, March 2022, and June 2022. Excluding LULC, there is other water and vegetation related parameters, which are completely ignored by the authors and not considered in this study.

3. For showing spatio-temporal differentiation of WH, vegetation indices such NDVI, MNDVI are most significant that LULC.

4. The authors used random forest (RF), support vector machine (SVM), and classification and regression tree (CART) machine learning algorithms, but no explanation about the ML models and how these models were executed in case of the present study.

5. Writing is not good and lack of well framing.

Thank you.

Author Response

Dear Editor and Reviewer

Reviewer #2:

I have found that there are numerous recent studies on Water Hyacinth (Eichhornia Crassipes) in Lake Tana, Ethiopia, using multiple machine learning models. Some of them are as follows:
Detecting spatiotemporal expansion of water hyacinth (Eichhornia crassipes) in Lake Tana, northern Ethiopia, published in Journal of the Indian Society of Remote Sensing, 48(5), 751-764.

Response: The mentioned study (a) was conducted by Asmare et al (2020) to detect E. crassipes in Landsat 8 images using a Maximum Likelihood Classifier (MLC) in 2017 and a decision tree algorithm in 2013 and 2015. However, the study did not evaluate the accuracy of the decision tree algorithm, which is a crucial aspect of evaluating classifier performance. Additionally, the study was limited to the Northeastern region of the lake and did not report any atmospheric correction process of the Landsat 8 image, which is essential for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Asmare et al. (2020) used Maximum Likelihood Classifier (MLC) and only used Landsat 8 bands, however, it should be noted that MLC is not considered a machine learning method and it is more prone to outliers compared to common machine learning techniques. Besides, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies (2013 - 2017), and there may be a need to update the information on the lake's current status.

Water Quality and Water Hyacinth Monitoring with the Sentinel-2A/B Satellites in Lake Tana (Ethiopia), published in Remote Sensing, 14(19), 4921.

Response: The second study (b) was conducted by Mucheye et al. (2022) and only used NDVI to detect water hyacinth in the northeastern region of Lake Tan. It should be noted that NDVI is not considered a machine-learning method. NDVI is a commonly used remote sensing index to detect vegetation. However, using only NDVI for water hyacinth detection can be problematic because NDVI is sensitive to variations in the water surface, which can affect the NDVI values of water hyacinth. Additionally, NDVI values for water hyacinths can be similar to those of other aquatic plants, making it difficult to distinguish water hyacinths from other vegetation using NDVI alone.

Water Hyacinth (Eichhornia crassipes) Detection Using Coarse and High Resolution Multispectral Data, published in Drones, 6(2), 47.

Response: The third mentioned study (c) was conducted by Pádua et al (2022) using machine learning algorithms in Portugal to detect Water Hyacinths. Although Pádua et al (2022) have used machine learning methods to detect this species in Portugal, the unique characteristics of Lake Tana's ecosystem and the presence of other aquatic plants may require another approach for accurate detection. Hence, comparing performance and identifying the most efficient machine-learning method and quantifying spatial coverage of E. crassipes using the most accurate method will provide a more complete understanding of the seasonal variability and spatial coverage of the infestation, inform management and control strategies, and improve decision-making by providing accurate and up-to-date information on the E. crassipes infestation.

So, what is the novelty of this study? As there are multiple studies on the same aspect, so what is the relevance of this study?

Response: Previous studies have utilized different classifiers, such as the maximum likelihood classifier (MLC) [1,14,19], decision tree [19], and random forest (RF) [41] to determine the spatial and temporal distribution of E. crassipes in Lake Tana. However, most studies on Lake Tana focused only northeastern part and used Level 1 Top of Atmosphere (TOA) satellite imagery without any atmospheric correction and have only used spectral bands. For example, Dersseh et al. [41] used the RF method to classify and estimate the areal cover of E. crassipes from Sentinel-2 MSI Level-1C TOA data but focused within the lake shoreline boundary without atmospheric corrections. Similarly, Worqlul et al. [1] used the MLC method on PlanetScope Level 1 Top of Atmosphere (TOA) products without considering atmospheric correlation in their analysis. However, atmospheric corrections are crucial for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Besides, Asmare et al. [19] employed a decision tree algorithm for E. crassipes detection in Landsat 8 images but did not evaluate the classifier's accuracy, but it is crucial for evaluating the performance of a classifier. To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

The spectral reflectance of E. crassipes varies with age and season, and the most suitable machine-learning algorithm for detection during different seasons must be clarified. Non-parametric machine learning methods, such as Random Forests (RF), Classification and Regression Trees (CART), and Support Vector Machines (SVM), can effectively handle outliers in training data [43] and often perform better on complex and non-linear datasets. The purpose of this study is to compare the performance of three non-parametric machine-learning algorithms (RF, SVM, and CART) for detecting E. crassipes, and quantify spatial coverage using atmospherically corrected Sentinel 2 and Landsat 8 images spectral bands and indices in combination through the Google Earth Engine (GEE) platform during different seasons. Comparing the performance and identifying the most efficient machine-learning method and quantifying spatial coverage of E. crassipes using the most accurate method will provide a more complete understanding of the seasonal variability and spatial coverage of the infestation, inform management and control strategies, and improve decision-making by providing accurate and up-to-date information on the E. crassipes infestation.

I have found that the authors only considered the LULC of autumn, winter, spring, and summer seasons for the year October 2021, January 2022, March 2022, and June 2022. Excluding LULC, there is other water and vegetation related parameters, which are completely ignored by the authors and not considered in this study.

Response: Water and vegetation indices are important parameters for land use and land cover (LULC) analysis, particularly for the detection of the invasive water hyacinth plant. These indices directly relate to the growth and survival of the plant, making them essential for understanding and managing the spread of water hyacinth. Other parameters may also be used in LULC analysis, but water and vegetation indices are very important (the most significant) parameters for the detection of this invasive species.

For showing spatiotemporal differentiation of WH, vegetation indices such as NDVI, MNDVI are most significant than LULC.

Response: In addition to the raw spectra bands, classifiers were provided with four derived environmental indices of water and vegetation. Nine spectral indices, such as the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), chlorophyll index green (Clg), normalized difference water index (NDWI), simple ratio index (SRI), modified normalized difference water index (MNDWI), green normalized difference vegetation index (GNDVI), optimized soil adjusted vegetation index (OSAVI), and difference vegetation index (DVI) were generated. To identify less correlated indices, a pair-wise correlation between the bands and indices was calculated using the ee.Reducer.pearsonsCorrelation() function in GEE. Five indices (GNDVI, OSAVI, DVI, EVI, and Clg) were identified as being correlated to each other and were removed. The remaining four indices (NDVI, SRI, NDWI, and MNDWI) (Table 2) were less correlated and used as input for the classifier, along with the spectral bands to improve E. crassipes detection accuracy.

SRI is sensitive to the chlorophyll content of the plant, while NDVI is sensitive to the greenness of the plant, both of which are important when mapping water hyacinths. The use of NDWI and MNDWI in combination is important for detecting water hyacinths in lakes with spatially variable turbidity. In a lake with spatially variable turbidity (like Lake Tana), different areas of the lake may have different levels of turbidity, which can affect the NDWI and MNDWI values. By using both NDWI and MNDWI, it is possible to detect water hyacinths in different areas of the lake, regardless of the level of turbidity. When NDWI values are decreased in areas with higher turbidity, MNDWI values can still provide reliable information about the presence of water hyacinths and vice versa. This can provide a more robust and accurate detection of water hyacinths in a lake with spatially variable turbidity.

The authors used random forest (RF), support vector machine (SVM), and classification and regression tree (CART) machine learning algorithms, but no explanation about the ML models and how these models were executed in the case of the present study.

Response: Explanation of machine learning algorithms and processes have been included as shown below (Page 6).---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Classification and Regression Tree (CART): CART is a binary decision classification tree [39]. The algorithm enables simple decision-making in logical if-then scenarios. It recursively divides the feature space into smaller regions, or nodes, based on the values of one or more input features by selecting the feature and the threshold value that will produce the most "pure" child nodes. CART is simple to interpret and understand, but is heavily influenced by the sample size employed in each class. Its effectiveness is also hindered by high-dimensionality data, which results in complex tree architectures [36]. CART can be prone to overfitting, especially when the tree is deep and complex. To deal with overfitting, the data tree can be pruned by setting a minimum number of observations per leaf. In this study, a cross-validation technique was used to determine the optimal number of observations per leaf, and a value of 10 produced high accuracy, resulting in the minimum number of observations per leaf being set to 10. The ee.Classifier.smileCart () function is part of Google Earth Engine (GEE) JavaScript API and creates a CART classifier. The ee.Classifier.smileCart () function was used to train the CART classifier with a given set of training data and feature bands. The classify () function was used to apply the trained classifier to the target imagery to classify each pixel of the image into one of the predefined classes.

Random Forest (RF): RF method was first introduced by Breiman [43]. RF is an ensemble machine learning method that uses multiple CART trees to make predictions. In RF, multiple trees are created by selecting a random subset of the data and combined to give a more accurate and stable prediction [43]. Because of the randomness of sample and feature selection, RF has the advantages of strong noise resistance, generalization ability, and strong processing ability for high-dimensional data without the need to manually select features[37]. The number of trees in the 50 – 500 range performed better with higher accuracy for RF classification. In the present study, 300 trees produced good results, thus, the number of trees was set to 300 for all datasets. The ee.Classifier.smileRandomForest () function is part of the GEE JavaScript Application Programming Interface (API) and creates a RF classifier. The function ee.Classifier.smileRandomForest () was used to train the RF classifier, and then classify () function was used to apply the trained classifier to the target imagery.

Writing is not good and lack of well framing.

Response: we have improved as per the comments given by the reviewers.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, the authors examine different machine learning techniques to detect water hyacinth (Eichhornia Crassipes) in a lake. The experimental study is interesting information in this paper. However, the main weakness of the paper lies in its lack of originality and novelty. The following suggestions may be considered to enhance the quality and clarity of the manuscript.

1- The motivation is not clear. Why did this work? Is any problem does it address that the previous methods can not?

2- Some papers are outdated, therefore those papers should be replaced with some recent papers i.e.,

https://www.sciencedirect.com/science/article/abs/pii/S0306987720308689

https://www.sciencedirect.com/science/article/pii/S0020025522007332

https://www.sciencedirect.com/science/article/pii/S0957417422024940

3- Related work section is totally ignored. Please include it as a separate section.

4- include a comparison table in the related work that should highlight the strengths and weaknesses of the proposed method as well as previous methods.

5- In the related works, existing studies can also be summarized in a tabular form to improve readability

6- Novelty of the algorithm needs to be incorporated

7- Frankly, using machine learning approaches for brain WH detection is not quite new, e.g., multiple other approaches have been used for comparison by the authors in the case study part. Then, what is revolutionarily new about this study?

8- How to deal with overfitting in your model?

9- The best results should be bolded in the tables

1- Besides language also needs improvement

Author Response

Dear Editor and Reviewer

Reviewer #3:

1) The motivation is not clear. Why did this work? Is any problem does it address that the previous methods can not?

Response: Previous studies have utilized different classifiers, such as the maximum likelihood classifier (MLC) [1,14,19], decision tree [19], and random forest (RF) [41] to determine the spatial and temporal distribution of E. crassipes in Lake Tana. However, most studies on Lake Tana focused only northeastern part and used Level 1 Top of Atmosphere (TOA) satellite imagery without any atmospheric correction and have only used spectral bands. For example, Dersseh et al. [41] used the RF method to classify and estimate the areal cover of E. crassipes from Sentinel-2 MSI Level-1C TOA data but focused within the lake shoreline boundary without atmospheric corrections. Similarly, Worqlul et al. [1] used the MLC method on PlanetScope Level 1 Top of Atmosphere (TOA) products without considering atmospheric correlation in their analysis. However, atmospheric corrections are crucial for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Besides, Asmare et al. [19] employed a decision tree algorithm for E. crassipes detection in Landsat 8 images but did not evaluate the classifier's accuracy, but it is crucial for evaluating the performance of a classifier. To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help to identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

2) Some papers are outdated, therefore those papers should be replaced with some recent papers i.e.,

https://www.sciencedirect.com/science/article/abs/pii/S0306987720308689

https://www.sciencedirect.com/science/article/pii/S0020025522007332

https://www.sciencedirect.com/science/article/pii/S0957417422024940

Response: Dear Reviewer, thank you very much for your constructive and useful comments. We appreciate you for sharing an interesting research work on the use of different machine learning and classification techniques for human brain MRI image classification, automatic brain tumor segmentation, and COVID-19 detection. However, the objective of our research was to evaluate the effectiveness of commonly used machine learning algorithms, such as Random Forest, Classification and Regression Tree, and Support Vector Machine, on Google Earth Engine, and to select the most effective method from the three, as well as to estimate the seasonal spatial coverage of water hyacinth using the selected method. The papers you mentioned used advanced machine learning methods, such as deep neural networks (DNNs) and CNNs, for human brain MRI image classification, automatic brain tumor segmentation, and COVID-19 detection. These algorithms will be helpful for our future work in different environmental management applications, including water hyacinth detection. But for now, the scope and context of the papers you mentioned are different from our work.

3) Related work section is totally ignored. Please include it as a separate section.

Response: Related work section have been included in the manuscript (pages 15&16) as shown below

Machine Learning Algorithms Related Work in Water Hyacinth Detection

In recent years, some studies have employed machine learning methods for the detection of water hyacinth (E. crassipes). A study by Dube et al. [78] used a Discriminant Analysis (DA) and Partial Least Squares Discriminant Analysis (PLS-DA) machine learning classification ensemble for E. crassipes detection, applied to Landsat 8 data. The DA and PLS-DA machine learning classification ensemble was able to detect E. crassipes with an accuracy of 95%. Similarly, a study by Pádua et al. [79] used Random Forest (RF), Support Vector Machine (SVM), Gaussian Naive Bayes (NB), k-nearest neighbors (KNN), and Artificial Neural Networks (ANN) on Unmanned Aerial Vehicle (UAV) and Sentinel 2 data. From these different classifiers, the study reported that RF performed the best, with an overall accuracy of 94%, while SVM performed the worst, with an overall accuracy of 87%. The study by Mukarugwiro et al. [72] used RF and SVM for E. crassipes detection and trained the methods using Landsat 8 images. The study also reported that RF (85%) outperformed SVM (65%) for E. crassipes detection in Rwandan water bodies. Thamaga and Dube [80] used Linear Discriminant Analysis (LDA) applied to Sentinel 2 data to map the seasonal dynamics of invasive E. crassipes in the Greater Letaba river system in Limpopo Province, South Africa. LDA was able to map E. crassipes with an overall accuracy of 80.79% during the wet season, and 79.04% during the dry season. Thamaga and Dube [76] also used Discriminant Analysis (DA) to test the capability of Landsat 8 Operational Land Imager (OLI) and Sentinel-2 MultiSpectral Instrument (MSI) for E. crassipes detection in a river system. They revealed that E. crassipes in small reservoirs can be mapped with an overall accuracy of 68.44% and 77.56% using Landsat 8 and Sentinel-2 data, respectively. In 2022, Ade et al. [81] employed RF to map E. crassipes using Sentinel-2 data and obtained an overall accuracy of 90%, with class-specific accuracies ranging from 79-91%. Moreover, Singh et al. [82] used a hierarchical classification approach on medium-resolution multispectral satellite data for mapping E. crassipes, resulting in an overall accuracy of 93% and an F1 score of 87%. In conclusion, a variety of machine learning methods have been used in recent years to detect E. crassipes with varying degrees of accuracy using various data sources. Overall, these previous studies show that machine learning methods have the potential to aid in the detection of E. crassipes. In the present study, all methods (RF, CART, and SVM) achieved greater than 90% overall accuracy on both Landsat 8 and Sentinel 2 images, indicating that all methods have the potential to aid in the detection of Eichhornia crassipes when compared to previously used methods. However, the Random Forest method had the highest accuracy.

4) include a comparison table in the related work that should highlight the strengths and weaknesses of the proposed method as well as previous methods.

Response: The strengths and weaknesses of previous and present study methods have been included in the manuscript (in the introduction and discussion section of the manuscript (Page2&14)

-----------------------------------------------------------------------------------------------------------------------------------------Machine learning algorithms are increasingly being used to classify and detect ground objects from satellite images due to high efficiency and accuracy when compared to common parametric algorithms such as Maximum Likelihood Classifiers (MLC) [27-35]. When identifying the distribution of E. crassipes from satellite imagery, machine learning algorithms have the potential to significantly improve detection accuracy, enhance clean-up efforts, and forecast potential hotspots [5]. Among machine learning algorithms, Random Forest (RF), Classification and Regression Tree (CART), and Support Vector Machine (SVM), are widely used for remote sensing image analysis due to their accuracy in classification [36]. RF and SVM are known for robustness and accuracy in image classification tasks, and the ability to handle high-dimensional data [37,38]; CART is known for interpretability and simplicity [39]. Additionally, deep learning techniques, such as Convolutional Neural Networks (CNNs), are also commonly used for image classification tasks and have shown promising results. However, they are computationally more expensive than common machine-learning methods [40]. -----------------------------------------------------------------------------------------------------------------------------------Each machine-learning algorithm has its own set of advantages and disadvantages [36]. SVM is hyperparameter sensitive, whereas RF is more robust and less affected by parameters and noise [69,70]. On the other hand, the performance of SVM is good when few training sample points are available [71]. CART is also sensitive to sample size [36].

5) In the related works, existing studies can also be summarized in a tabular form to improve readability

Response: Existing studies have been summarized in a tabular form as shown in Table 4 (Page 16)

Table 4: Related studies with machine learning methods applied to remote sensing data for water hyacinth detection

Literature	Methods	Data sets	Overall Accuracy
Dube et al. (2017)	DA and PDA ensemble	Landsat 8		95%
Mukarugwiro et al. (2019)	RF	Landsat 8		85%
	SVM	Landsat 8		65%
Pádua et al. (2022)	RF	Sentinel 2		90%
	SVM			83%
	NB			87%
	KNN			87%
	ANN			90%
Thamaga and Dube (2019)	LDA	wet season Sentinel 2		81%
		dry season sentinel 2		79%
Thamaga and Dube (2018)	DA	Landsat 8		68%
		Sentinel 2		78
Ade et al. (2022)	RF	Sentinel-2		90%,
Present study	RF	Sentinel 2		98
	CART			97.6
	SVM			97.5
	RF	Landsat 8		97
	CART			95
	SVM			95

6) Novelty of the algorithm needs to be incorporated

Response: The present study objective was to evaluate the effectiveness of existing common machine learning methods such as random forest (RF), support vector machine (SVM), and classification and regression tree (CART) machine learning algorithms for water hyacinth (E. crassipes) detection and estimating seasonal spatial coverage of the weed using the most effective method on the Google Earth Engine (GEE) platform using Landsat 8 and Sentinel 2 images. We have used the existing common machine-learning methods for water hyacinth detection. All methods (RF, CART, and SVM) achieved greater than 90% overall accuracy on both Landsat 8 and Sentinel 2 images. Through the accuracy difference among methods small RF was the most accurate for water hyacinth detection.

7) Frankly, using machine learning approaches for brain WH detection is not quite new, e.g., multiple other approaches have been used for comparison by the authors in the case study part. Then, what is revolutionarily new about this study?

Response: Although few studies have used machine learning methods to detect water hyacinth (WH) in different parts of the world Lake, the unique characteristics of Lake Tana's ecosystem and the presence of other aquatic plants may require another method for accurate detection. Hence Comparing the performance and identifying the most efficient machine-learning method and quantifying spatial coverage of E. crassipes using the most accurate method will provide a more complete understanding of the seasonal variability and spatial coverage of the infestation, inform management and control strategies, and improve decision-making by providing accurate and up-to-date information on the E. crassipes infestation.

To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

8) How to deal with overfitting in your model?

Response: Hyperparameter tuning was used to deal with overfitting. A cross-validation technique was used to determine the optimal number of observations per leaf and the number of trees in CART and RF, respectively. A grid search technique was also used to determine the optimal values for the C and linear kernel in SVM. Detailed processes are shown below.

--------------------Three non-parametric machine learning methods available in the GEE platform were used and chosen for their ability to handle outliers in the training data. A detailed explanation of each classifier and process is given as follows:

Classification and Regression Tree (CART): CART is a binary decision classification tree [39]. The algorithm enables simple decision-making in logical if-then scenarios. It recursively divides the feature space into smaller regions, or nodes, based on the values of one or more input features by selecting the feature and the threshold value that will produce the most "pure" child nodes. CART is simple to interpret and understand, but is heavily influenced by the sample size employed in each class. Its effectiveness is also hindered by high-dimensionality data, which results in complex tree architectures [36]. CART can be prone to overfitting, especially when the tree is deep and complex. To deal with overfitting, the data tree can be pruned by setting a minimum number of observations per leaf. In this study, a cross-validation technique was used to determine the optimal number of observations per leaf, and a value of 10 produced high accuracy, resulting in the minimum number of observations per leaf being set to 10. The ee.Classifier.smileCart () function is part of Google Earth Engine (GEE) JavaScript API and creates a CART classifier. The ee.Classifier.smileCart () function was used to train the CART classifier with a given set of training data and feature bands. The classify () function was used to apply the trained classifier to the target imagery to classify each pixel of the image into one of the predefined classes.

Random Forest (RF): RF method was first introduced by Breiman [43]. RF is an ensemble machine learning method that uses multiple CART trees to make predictions. In RF, multiple trees are created by selecting a random subset of the data and combined to give a more accurate and stable prediction [43]. Because of the randomness of sample and feature selection, RF has the advantages of strong noise resistance, generalization ability, and strong processing ability for high-dimensional data without the need to manually select features[37]. The number of trees in the 50 – 500 range performed better with higher accuracy for RF classification. In the present study, 300 trees produced good results, thus, the number of trees was set to 300 for all datasets. The ee.Classifier.smileRandomForest () function is part of the GEE JavaScript Application Programming Interface (API) and creates a RF classifier. The function ee.Classifier.smileRandomForest () was used to train the RF classifier, and then classify () function was used to apply the trained classifier to the target imagery.

9) The best results should be bolded in the tables

Response: Best results have been bolded in Table

10) Besides language also needs improvement

Response: The grammatical errors and language have been checked and corrected.

Author Response File: Author Response.pdf

Reviewer 4 Report

The manuscript “Water Hyacinth (Eichhornia Crassipes) Detection Using Machine Learning Algorithms in Lake Tana, Ethiopia” compares several machine learning methods in the context of water hyacinth detection. The algorithm comparison study may be interesting enough for some readers, but some clarifications and major revisions are needed before publication. Please see the details below:

1. My major concern is about the significance and novelty of the research. Lines 15 – 18 and lines 75 – 80 are about the contributions of the paper. However, I’m not sure if the comparison of several traditional machine learning algorithms is significant enough for the community. The authors claim the significance of the research in lines 85 – 87, but “invaluable information” is a very strong statement and is hardly shown in this manuscript. I recommend the authors acknowledge the existence of other more modern and advanced machine learning or deep learning algorithms, with proper citations, and clarify that one of the research’s limitations is the selection of algorithms. I also recommend tuning down the statement a bit rather than claiming it’s “invaluable”.

2. Line 162: “Advanced machine learning methods..” These are not considered “advanced” methods compared to modern ones.

3. Line 167 – 170: Function names such as “ee.Classifier.smileRandomForest” should be introduced and briefly explained.

4. Table 2: Some equations are missing the right-side parentheses.

5. Figures 3-4: Black circles should be explained in the figure description or on the plots themselves.

6. Tables 3-4: SVM, CART, and RF’s performances are very similar in this case. I don’t think the difference is significant enough to draw a conclusion. If possible, I recommend comparing more methods, including both more advanced ones (e.g., CNN) and basic ones (e.g., logistic regression), so that more insightful conclusions can be drawn.

7. Table 4: In the table, “Machine learning algorism” -> “Machine learning algorithm”

8. I noticed quite a few grammatical errors. The manuscript needs an extensive English language check.

Author Response

Dear Editor and Reviewer

Reviewer #4:

My major concern is about the significance and novelty of the research. Lines 15 – 18 and lines 75 – 80 are about the contributions of the paper. However, I’m not sure if the comparison of several traditional machine learning algorithms is significant enough for the community. The authors claim the significance of the research in lines 85 – 87, but “invaluable information” is a very strong statement and is hardly shown in this manuscript. I recommend the authors acknowledge the existence of other more modern and advanced machine learning or deep learning algorithms, with proper citations, and clarify that one of the research’s limitations is the selection of algorithms. I also recommend tuning down the statement a bit rather than claiming it’s “invaluable”.

Response: The significance and novelty of the study has been revised as per the comments given by the reviewer (Page 2&3 in the manuscript)

. I also recommend tuning down the statement a bit rather than claiming it’s “invaluable”.

Response: This statement also have been changed as shown in the last paragraph of the Introduction

Abstract

-----------------------------------------------------------------------------------------------------------------------------------------------Using appropriate remote sensing detection methods and determining the seasonal distribution of the weed is important for decision-making, water resource management, and environmental protection. As the demand for reliable estimation of E. crassipes mapping from satellite data grows, comparing the performance of different machine learning algorithms could help in identifying the most effective method for E. crassipes detection in the Lake. Therefore, this study aimed to examine the ability of random forest (RF), support vector machine (SVM), and classification and regression tree (CART) machine learning algorithms for detecting E. crassipes and estimating seasonal spatial coverage of the weed on the Google Earth Engine (GEE) platform using Landsat 8 and Sentinel 2 images. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Introduction

---------------------------------------------------------------------------------------------------------------------------------------------------------------------- Among machine learning algorithms, Random Forest (RF), Classification and Regression Tree (CART), and Support Vector Machine (SVM) are widely used for remote sensing image analysis due to their accuracy in classification [36]. RF and SVM are known for robustness and accuracy in image classification tasks, and the ability to handle high-dimensional data [37,38]; CART is known for interpretability and simplicity [39]. Additionally, deep learning techniques, such as Convolutional Neural Networks (CNNs), are also commonly used for image classification tasks and have shown promising results. However, they are computationally more expensive than common machine-learning methods [40]. There is no single machine-learning algorithm that can handle all pattern-recognition tasks due to differences in texture and spectral reflectance of the ground features, and different algorithms often produce variable results [31]. Moreover, the relative performance of various classification methods will vary by season due to changes in the physiological characteristics of plants and life-cycle influence on spectral reflectance.

Previous studies have utilized different classifiers, such as the maximum likelihood classifier (MLC) [1,14,19], decision tree [19], and random forest (RF) [41] to determine the spatial and temporal distribution of E. crassipes in Lake Tana. However, most studies on Lake Tana focused only northeastern part and used Level 1 Top of Atmosphere (TOA) satellite imagery without any atmospheric correction and have only used spectral bands. For example, Dersseh et al. [41] used the RF method to classify and estimate the areal cover of E. crassipes from Sentinel-2 MSI Level-1C TOA data but focused within the lake shoreline boundary without atmospheric corrections. Similarly, Worqlul et al. [1] used the MLC method on PlanetScope Level 1 Top of Atmosphere (TOA) products without considering atmospheric correlation in their analysis. However, atmospheric corrections are crucial for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Besides, Asmare et al. [19] employed a decision tree algorithm for E. crassipes detection in Landsat 8 images but did not evaluate the classifier's accuracy, but it is crucial for evaluating the performance of a classifier. To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help to identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

Line 162: “Advanced machine learning methods..” These are not considered “advanced” methods compared to modern ones.

Response: Advanced machine learning methods have been checked everywhere in the manuscript and changed to Non-parametric/common machine learning methods as shown below.

For example, Three non-parametric machine learning methods available in the GEE platform were used and chosen for their ability to handle outliers in the training data. A detailed explanation of each classifier and process is given as follows:------------------------------------------------------------------------------.

Line 167 – 170: Function names such as “ee.Classifier.smileRandomForest” should be introduced and briefly explained.

Response: Description of the classification detection algorithm and process have been included and improved as per the comments given by the reviewer (Pages 5& 6).

--------------------------------------------------------------Three non-parametric machine learning methods available in the GEE platform were used and chosen for their ability to handle outliers in the training data. A detailed explanation of each classifier and process is given as follows:

Classification and Regression Tree (CART): CART is a binary decision classification tree [39]. The algorithm enables simple decision-making in logical if-then scenarios. It recursively divides the feature space into smaller regions, or nodes, based on the values of one or more input features by selecting the feature and the threshold value that will produce the most "pure" child nodes. CART is simple to interpret and understand, but is heavily influenced by the sample size employed in each class. Its effectiveness is also hindered by high-dimensionality data, which results in complex tree architectures [36]. CART can be prone to overfitting, especially when the tree is deep and complex. To deal with overfitting, the data tree can be pruned by setting a minimum number of observations per leaf. In this study, a cross-validation technique was used to determine the optimal number of observations per leaf, and a value of 10 produced high accuracy, resulting in the minimum number of observations per leaf being set to 10. The ee.Classifier.smileCart () function is part of Google Earth Engine (GEE) JavaScript API and creates a CART classifier. The ee.Classifier.smileCart () function was used to train the CART classifier with a given set of training data and feature bands. The classify () function was used to apply the trained classifier to the target imagery to classify each pixel of the image into one of the predefined classes.

Random Forest (RF): RF method was first introduced by Breiman [43]. RF is an ensemble machine learning method that uses multiple CART trees to make predictions. In RF, multiple trees are created by selecting a random subset of the data and combined to give a more accurate and stable prediction [43]. Because of the randomness of sample and feature selection, RF has the advantages of strong noise resistance, generalization ability, and strong processing ability for high-dimensional data without the need to manually select features[37]. The number of trees in the 50 – 500 range performed better with higher accuracy for RF classification. In the present study, 300 trees produced good results, thus, the number of trees was set to 300 for all datasets. The ee.Classifier.smileRandomForest () function is part of the GEE JavaScript Application Programming Interface (API) and creates a RF classifier. The function ee.Classifier.smileRandomForest () was used to train the RF classifier, and then classify () function was used to apply the trained classifier to the target imagery.

Table 2: Some equations are missing the right-side parentheses.

Response: Parentheses have been checked and corrected in Table 2.

Table 2: Equations and sources of spectral indices

Index	Formula/equations	Source
NDVI	)	[50]
NDWI	)	[51]
SRI		[52]
MNDWI		[53]

Figures 3-4: Black circles should be explained in the figure description or on the plots themselves.

Response: Black circles (ovals) shape in the figure represents the area of misclassification and these explanations have been included in the figure caption (from Figure 3-Figure 6) and the result descriptions too (Page 9 and 10)

Tables 3-4: SVM, CART, and RF’s performances are very similar in this case. I don’t think the difference is significant enough to draw a conclusion. If possible, I recommend comparing more methods, including both more advanced ones (e.g., CNN) and basic ones (e.g., logistic regression), so that more insightful conclusions can be drawn.

Response: All methods have an overall accuracy and F1 score of greater than 90% in both data sets. Although the accuracy differences among methods are small, Random Forest (RF) was found to be the most accurate with an overall accuracy ranging from 97.4-99.13% in all seasons, while CART and SVM had nearly the same accuracy ranging from 96.93-98% and from 96.92-98.54%, respectively. Furthermore, there were significant differences in the spatial distribution of water hyacinth in all methods, as shown in Figures 3- 6. Additionally, there were meaningful differences among the three methods in terms of the coverage areas of different land use/cover types and water hyacinth, as shown in Table 3. Despite the small differences in accuracy among the methods, the coverage areas of water hyacinth varied depending on the algorithm employed, with RF, CART, and SVM. Therefore, it is crucial to evaluate the performance of different algorithms and select the most accurate method for accurately estimating water hyacinth coverage. In this study, RF was found to be the most accurate in terms of overall accuracy, kappa coefficient, and F1 score, and also we have confirmed the accuracy spatially on the field. But, due to time constraints, we are unable to compare RF, SVM, and CART with CNN and logistic regression. (Pages 8&11)

Table 4: In the table, “Machine learning algorism” -> “Machine learning algorithm”

Response: Machine learning algorism has been changed to Machine learning algorithm in Table 4, now have been moved to supplementary material because of another reviewer's comment.

I noticed quite a few grammatical errors. The manuscript needs an extensive English language check.

Response: The grammatical errors and clarity have been checked and corrected.

Author Response File: Author Response.pdf

Reviewer 5 Report

This study compares the performance of three classification methods in assessing total area coverage of an invasive aquatic plant in Lake Tana, Ethiopia and additionally reports on the seasonal pattern of coverage. The study appears to have followed sound methodology, although I would like greater clarity/detail about the methods. The results are presented in several different styles and some important information is missing. My overall feeling is that the central focus on comparing the accuracy of the three classifiers is not highly impactful. The importance of this paper could be improved by emphasizing the seasonal dynamics of water hyacinth coverage or by going even further to compare those dynamics across different regions of the lake's shoreline. As the discussion mentions, the performance characteristics of the three methods employed are already fairly well understood and the outcome observed in this paper is not surprising.

My detailed comments are included in the attached PDF, but I will try to give highlights here.

Methods - Please make it clearer that you used the raw spectral bands as well as the nine derived indices. The manuscript already says that, but not as clearly as it could. I have concerns that using multiple vegetation and water indices that are likely highly correlated will lead to problems. I would encourage the authors to look at the correlation of these indices and select a subset to include in their analysis. DVI can surely be eliminated, since it is a linear combination of two spectra that will already be captured by inclusion of the raw spectra. I encourage the authors to report permutation importance of the inputs. This would give them the opportunity to talk about how best to measure WH coverage and whether these recommendations differ by method or season.

Results: Reflectance curves: I don't think it is necessary to use figure space to make the point that Landsat has fewer bands than Sentinel 2. However, it would be great to see these curves (for Sentinel 2 only) in all four seasons.

Performance: The opening paragraph and Figure 3-4 could contribute more to the manuscript. I would like to see quantitative statements about the rate of misclassification by cover type, season, and method. Seems like this would be a table. Please also add to the Figs. 3-4 captions to say what the ovals indicate. I assume it is areas of particular misclassification trouble.

The following paragraph on page 8 could make much more general statements to help the reader understand outcomes and let Table 3 communicate the precise numbers. Could even put tables 3 and 4 in the supplemental info and focus instead (here and in the paper in general) on the spatiotemporal patterns of WH coverage.

Discussion: I think the authors could make a strong assertion that they've shown the choice of method and sensor is not a major factor in the study of WH coverage.

In terms of producing information to support WH management/control or understand the population dynamics of the weed, is there a meaningful difference between the methods?

Does this study demonstrate anything else that the previous studies the authors cite have not already established? Does it conflict with previous findings or resolve important questions posed for WH detection?

Were the methods employed here more accurate than the studies mentioned in the introduction using MLA or discriminant analysis?

How does this study contribute to the broader effort to improve environmental assessment and monitoring via remotely-sensed data?

There was a statement in the introduction about understanding spatiotemporal dynamics. Building that up more and then coming back to it here would be good.

Comments for author File: Comments.pdf

Author Response

Dear Editor and Reviewer

Reviewer #5:

Methods - Please make it clearer that you used the raw spectral bands as well as the nine derived indices. The manuscript already says that, but not as clearly as it could. I have concerns that using multiple vegetation and water indices that are likely highly correlated will lead to problems.

I would encourage the authors to look at the correlation of these indices and select a subset to include in their analysis. DVI can surely be eliminated, since it is a linear combination of two spectra that will already be captured by inclusion of the raw spectra.

Response: In addition to the raw spectra bands presented in Table 1, classifiers were provided with four derived environmental indices of water and vegetation. Nine spectral indices, such as the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), chlorophyll index green (Clg), normalized difference water index (NDWI), simple ratio index (SRI), modified normalized difference water index (MNDWI), green normalized difference vegetation index (GNDVI), optimized soil adjusted vegetation index (OSAVI), and difference vegetation index (DVI) were generated. To identify less correlated indices, a pair-wise correlation between the bands and indices was calculated using the ee.Reducer.pearsonsCorrelation() function in GEE. Five indices (GNDVI, OSAVI, DVI, EVI, and Clg) were identified as being correlated to each other and were removed. The remaining four indices (NDVI, SRI, NDWI, and MNDWI) (Table 2) were less correlated and used as input for the classifier, along with the spectral bands to improve E. crassipes detection accuracy.

Table 2: Equations and sources of spectral indices

Index	Formula/equations	Source
NDVI	)	[50]
NDWI	)	[51]
SRI		[52]
MNDWI		[53]

SRI is sensitive to the chlorophyll content of the plant, while NDVI is sensitive to the greenness of the plant, both of which are important when mapping water hyacinths. The use of NDWI and MNDWI in combination is important for detecting water hyacinths in lakes with spatially variable turbidity. In a lake with spatially variable turbidity (like Lake Tana), different areas of the lake may have different levels of turbidity, which can affect the NDWI and MNDWI values. By using both NDWI and MNDWI, it is possible to detect water hyacinths in different areas of the lake, regardless of the level of turbidity. When NDWI values are decreased in areas with higher turbidity, MNDWI values can still provide reliable information about the presence of water hyacinths and vice versa. This can provide a more robust and accurate detection of water hyacinths in a lake with spatially variable turbidity.

Figure S: Pair-wise correlation between bands and indices in Sentinel 2 image

I encourage the authors to report permutation importance of the inputs. This would allow them to talk about how best to measure WH coverage and whether these recommendations differ by method or season.

Response: We attempted to determine the permutation importance but encountered coding issues and need more time to incorporate it. To our knowledge, GEE does not have built-in support for permutation importance in RF, SVM, and CART methods, it needs Payton code for it and encountered coding problems. Instead, we have included the feature importance of the RF classifier during different seasons using the explain() function.

--------------------------------------------------------------------------------------------------------------------------------------------------------Feature importance of the RF classifier was assessed using the explain () method in GEE for different seasons in Sentinel 2 image to evaluate the influence of spectral bands and indices on the classifier's accuracy. Feature importance values indicate how much each input feature, such as spectral bands or indices, contributed to the predictions made by the RF classifier (Page 7).

3.4 Feature Importance (Page 13).

The importance of spectral bands and indices in the RF Classifier varied with the season (Figure 8). In the Sentinel 2 datasets, spectral indices and spectral bands were found to be the most important variables for classification using the RF classifier. Band 12 (B12) contributed the most to autumn and spring image classification when compared to other spectral bands and indices. In contrast, the Simple Ratio Index (SRI) and band 2 (B2) contributed the most in spring and summer, respectively.

Figure 8: Relative feature importance of RF classified image in autumn (a), winter (b), spring (c), and (d) summer from Sentinel 2 image.

Results: Reflectance curves: I don't think it is necessary to use figure space to make the point that Landsat has fewer bands than Sentinel 2. However, it would be great to see these curves (for Sentinel 2 only) in all four seasons.

Response: The Reflectance curve section has been revised as per the comments given by the reviewer (Page 8).

Results

3.1 Water Hyacinth Spectral Reflectance Curve

The seasonal spectral response curves of water, water hyacinth, cultivated land, grazing land, and forest (church forest) in and around Lake Tana were assessed from Sentinel 2 (Figure 2). The spectral reflectance values of each land use/cover varied with the season. Water hyacinth showed the highest reflectance value in the red edge 2 (B6), red edge 3 (B7), near-infrared (B8), and red edge 4 (B8A) bands during autumn, winter, and spring. In the summer season, water hyacinth displayed a relatively lower spectral reflectance value compared to both forest and grazing land in the red edge 2 (B6), red edge 3 (B7), near-infrared (B8), and red edge 4 (B8A) band regions. In the autumn season, the spectral reflectance value of forest was higher than cultivated land, grazing land, and water in the red edge 2 (B6), red edge 3 (B7), near-infrared (B8), and red edge 4 (B8A) bands. In winter and spring, the reflectance value of forest was less than cultivated land and grazing land, respectively. In general, the spectral reflectance value of water hyacinth was highest in autumn and lowest in summer relative to other seasons.

Figure 2: Average spectral reflectance curves of different land use/cover from Sentinel 2 image in (a) autumn (Oct 2021), (b) winter (Jan 2022), (c) spring (Mar 2022) and (d) summer (Jun 2022).

Performance: The opening paragraph and Figures 3-4 could contribute more to the manuscript. I would like to see quantitative statements about the rate of misclassification by cover type, season, and method. Seems like this would be a table.

Response: The opening paragraph and Figures 3-6 have been revised (Pages 8 -10). Quantitative information about the rate of misclassification by cover type, season, and method has also been included in the manuscripts as per the comments given by the reviewer (Page 11).

3.2 Performance of Machine Learning Algorithms: SVM, CART and RF

All three machine learning classifiers were able to detect the spatial distribution of water hyacinths (E. crassipes) in Lake Tana (Figure 3 - 6). However, when using SVM and CART from Sentinel 2 images in the autumn, spring and summer, some other vegetation types were misclassified as E. crassipes as indicated in the black oval shape in the figures (Figures 3b,c, 5b,c, 6b,c and Table 3). Misclassification of other vegetation was more prevalent in the SVM during the autumn and summer (Table 3, Figures 3c and 6c). Misclassification of other vegetation was also more prevalent in CART classifiers during spring (Table 3 and Figure 5b). Some water was also misclassified as other vegetation in the CART and RF classifiers during the autumn and winter, and this was more prevalent in the CART classifier (Figures 3a,b and 4a,b). Similarly, some E. crassipes were misclassified as other vegetation in the CART and SVM classifiers during the winter in the Sentinel 2 data (Figure 4b,c).

Figure 3: Machine learning classifier comparison in autumn (October 2021): Sentinel 2 (a) RF, (b) CART, (c) SVM; Landsat 8 (d) RF, (e) CART, (f) SVM land use/cover map. Black oval shape in the figure represents the area of misclassification.

During the winter, spring, and summer seasons, some other vegetation was classified as E. crassipes in Landsat 8 CART and SVM classified images as shown in the black oval shape in the figures (Figure 4e, f-6e, f). Misclassification was highest in CART during winter and summer (Figure 4e and 6e and Table 3) and in SVM during spring season (Table 3 and Figure 5f). The RF machine learning algorithm detected E. crassipes relatively better than the SVM and CART algorithms in all seasons as confirmed by field observations (Figure 3a,d-6a,d). Some water was also misclassified as other vegetation in the CART classifiers across all seasons in Landsat 8 data (Figure 3e-6e).

Seasonal spatial coverage of E. crassipes and other land uses/covers in RF, CART, and SVM methods are shown in Table 3. The area coverage of E. crassipes and other land uses/covers differed in the three methods in all seasons. In the autumn Sentinel 2 image, there was a 2.7 km² and 6.6 km² differences in the E. crassipes spatial coverage between the RF and CART methods, and between the RF and SVM method, respectively. This indicates that other land uses/covers were misclassified as E. crassipes more often using the SVM method compared to other methods during the autumn season. The largest misclassification of other land use/cover occurred in the summer Landsat 8 CART and SVM classified images at 10.13 km² and 8.59 km² area differences from RF, respectively (Table 3). In both Sentinel 2 and Landsat 8 data sets the area coverage of E. crassipes and other land uses/covers varies among methods (RF, CART and SVM) in all season. The spatial coverage of E. crassipes varies depending on the algorithm employed and yielded different results in both data sets (Table 3). As a result, evaluating the performance of different algorithms and selecting the most accurate method is critical for accurately estimating E. crassipes coverage.

Table 3: Area (km²) of seasonal land use/cover from RF, CART, and SVM classifiers during autumn, winter, spring, and summer

	Area of land use/cover (km²)
	Sentinel 2				Landsat 8
land use/cover type	RF	CART		SVM	RF	CART	SVM
Autumn
Water	3029.3	3031.1		3042.9	3047.12	3014.59	3059.82
Water hyacinth	22.4	25.1		29	19.87	17.77	21.06
Other vegetation	553.6	548.3		554.5	440.14	478.42	445.22
Bareland	122.1	122.8		100.8	219.47	215.80	200.49
Winter
Water	3006.5	2970.4		3009	3044.16	3019.87	3034.68
Water hyacinth	11.2	9.5		6	14.12	17.55	16.72
Other vegetation	318.1	441.4		382.7	264.60	339.09	330.89
Bareland	391.4	306		329.7	403.80	350.16	344.41
Spring
Water	3014.3	3008.9		3028.8	3023.18	3016.33	3038.96
Water hyacinth	4.2	7		4.6	5.35	5.45	6.81
Other vegetation	294.2	306.6		264.5	258.10	282.93	256.99
Bareland	414.6	404.8		429.4	440.07	421.98	423.89
Summer
Water	2983.1		2970.8	2999.5	3005.10	2999.32	3016.59
Water hyacinth	2.2		2.5	4.4	4.09	14.22	12.68
Other vegetation	310.4		340.4	272.2	463.99	499.27	448.31
Bareland	431.6		413.6	451.3	254.27	214.64	249.87

Please also add to the Figs. 3-4 captions to say what the ovals indicate. I assume it is areas of particular misclassification trouble.

Response: Black ovals shape in the figure represents the area of misclassification and these explanations have been included in the figure caption (from Figure 3-6) and the result descriptions too (Pages 9 and 10)

The following paragraph on page 8 could make much more general statements to help the reader understand outcomes and let Table 3 communicate the precise numbers. Could even put tables 3 and 4 in the supplemental info and focus instead (here and in the paper in general) on the spatiotemporal patterns of WH coverage.

Response: The paragraph and Tables 3 & 4 have been rewritten as per the comments given by the reviewer and Tables 3 and 4 have been moved to supplementary materials (Pages 11&12).

--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The overall accuracy of the Sentinel 2 classified image was greater than 95% using all three methods. RF had the highest accuracy ranging from 97 - 99% across all seasons, followed by CART (97 - 98%) and SVM (97 - 98%). Similarly, the Kappa coefficient for RF, SVM, and CART was greater than 0.96 across the seasons ranging from 0.97 - 0.99 in RF; 0.96 - 0.98 in SVM; 0.96 - 0.97 in CART. Although the differences in performance among the methods were small, RF had relatively higher accuracy while the CART and SVM methods were similar in Sentinel 2 classified images. The overall accuracy of the Landsat 8 classified images was also greater than 93% in all seasons ranging from 95 - 98% in RF; 94 - 96% in SVM; 94 - 96% in CART. The Kappa coefficient of the Landsat 8 classified images using all methods was greater than 0.90 in all seasons ranging from 0.94 - 0.97 in RF and 0.92 - 0.95 in both SVM and CART. The differences in accuracy among the methods were also small in Landsat 8, but RF had the highest accuracy relative to SVM and CART. Overall accuracy and Kappa coefficients for E. crassipes classified image detection are slightly lower in Landsat 8 compared to Sentinel 2 amongst all three classifiers and all four seasons. In terms of overall accuracy and Kappa coefficient, the RF algorithm slightly outperformed the SVM and CART algorithms across all seasons (Table S1).----------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Discussion: I think the authors could make a strong assertion that they've shown the choice of method and sensor is not a major factor in the study of WH coverage.

Response: All methods have an overall accuracy and F1 score of greater than 90% in both data sets. Although the accuracy differences among methods are small, there were notable differences in the spatial distribution of water hyacinth in all methods, as shown in Figures 3-6 and Table 3. Additionally, there were meaningful differences among the three methods in terms of the coverage areas of different land use/cover types and water hyacinth, as shown in Table 3. Despite the small differences in accuracy among the methods, the coverage areas of water hyacinth varied depending on the algorithm employed, such as RF, CART, and SVM. The average overall accuracy of classifiers for Sentinel 2 images was 97.8%, while the average overall accuracy of classifiers for Landsat 8 images was 95.5%. Though the difference in accuracy between the two is not significant, there is a meaningful difference in the estimation of land use/cover and water hyacinth area coverage (Table 3). Therefore, it is crucial to evaluate the performance of different algorithms and sensors and select the most accurate method and a sensor for accurately estimating water hyacinth coverage (Page 14).

The average overall accuracy of RF, CART, and SVM in the present study was greater than 95% for Sentinel 2 images and greater than 90% for Landsat 8 images. Although the difference in accuracy among the three methods is small, there is a significant difference in the final estimation of spatial coverage for E. crassipes and other land uses/covers (Table 3). The average overall accuracy of classifiers for Sentinel 2 images was 97.8%, while the average overall accuracy of classifiers for Landsat 8 images was 95.5%. Though the difference in accuracy between the two is not significant, there is a meaningful difference in the estimation of land use/cover and water hyacinth area coverage (Table 3). During the autumn season, the F1 score of the E. crassipes class was highest across all classifiers; however, the E. crassipes class had the lowest F1 score and Kappa coefficient during the winter in Sentinel 2 and during winter and spring in Landsat 8 images. The low F1 score and Kappa coefficient of the E. crassipes class during winter are attributable to the presence of different irrigated crops around Lake Tana. Some dense irrigated crops have similar reflectance values to the invasive weed, which will make it difficult to distinguish E. crassipes from crops. --------------------------------------------------------

In terms of producing information to support WH management/control or understand the population dynamics of the weed, is there a meaningful difference between the methods?

Response: Yes, The study found that the coverage area of E. crassipes varies among RF, CART, and SVM methods in both Sentinel 2 and Landsat 8 data sets, as shown in Table 3. Hence, evaluating the performance of different algorithms and selecting the most accurate method is critical for accurately estimating E. crassipes spatial coverage. Though the difference in accuracy among methods and between the two sensors is small, there is a meaningful difference in the estimation of land use/cover and water hyacinth area coverage (Table 3)(Pages 11&14).

Does this study demonstrate anything else that the previous studies the authors cite have not already established? Does it conflict with previous findings or resolve important questions posed for WH detection?

Response: The present study identified the most accurate and suitable method for water hyacinth detection in Lake Tana. Random Forest (RF) classifier was the most accurate and effective method for detecting water hyacinth in the Lake. Similarly, the appropriate (optimal) seasons were identified for water hyacinth mapping. Important vegetation and water indices (NDVI, MNDWI, NDWI, and SRI) were also identified in and around Lake Tana, these are very important for improving water hyacinth detection accuracy. Besides, the study provided new information on the seasonal distribution of water hyacinths, which is important for management and control efforts. These findings, which have not been reported in previous studies, are crucial for the management and control of water hyacinths.

------------- Previous studies have utilized different classifiers, such as the maximum likelihood classifier (MLC) [1,14,19], decision tree [19], and random forest (RF) [41] to determine the spatial and temporal distribution of E. crassipes in Lake Tana. However, most studies on Lake Tana focused only northeastern part and used Level 1 Top of Atmosphere (TOA) satellite imagery without any atmospheric correction and have only used spectral bands. For example, Dersseh et al. [41] used the RF method to classify and estimate areal cover of E. crassipes from Sentinel-2 MSI Level-1C TOA data, but focused within the lake shoreline boundary without atmospheric corrections. Similarly, Worqlul et al. [1] used the MLC method on PlanetScope Level 1 Top of Atmosphere (TOA) products without considering atmospheric correlation in their analysis. However, atmospheric corrections are crucial for accurately estimating coverage area and spatiotemporal patterns of aquatic plants in lake environments [26,42]. Besides, Asmare et al. [19] employed a decision tree algorithm for E. crassipes detection in Landsat 8 images, but did not evaluate the classifier's accuracy, but it is crucial for evaluating the performance of a classifier. To our knowledge, no comparative studies have evaluated the effectiveness of different machine-learning algorithms for E. crassipes detection during different seasons in Lake Tana. Given the unique characteristics of the lake and the species, certain algorithms may perform better under different conditions. Conducting a study that compares the performance of various algorithms for E. crassipes detection will help to identify the most effective approach for detecting the weed in the lake. Further, the conditions and distribution of E. crassipes in the lake may have evolved since previous studies, and there may be a need to update the information on the lake's current status.

The spectral reflectance of E. crassipes varies with age and season, and the most suitable machine-learning algorithm for detection during different seasons must be clarified. Non-parametric machine learning methods, such as Random Forests (RF), Classification and Regression Trees (CART), and Support Vector Machines (SVM), can effectively handle outliers in training data [43] and often perform better on complex and non-linear datasets. The purpose of this study is to compare the performance of three non-parametric machine-learning algorithms (RF, SVM, and CART) for detecting E. crassipes, and quantify spatial coverage using atmospherically corrected Sentinel 2 and Landsat 8 images spectral bands and indices in combination through the Google Earth Engine (GEE) platform during different seasons. Comparing performance and identifying the most efficient machine-learning method and quantify spatial coverage of E. crassipes using the most accurate method will provide a more complete understanding of the seasonal variability and spatial coverage of the infestation, inform management and control strategies, and improve decision-making by providing accurate and up-to-date information on the E. crassipes infestation.

Were the methods employed here more accurate than the studies mentioned in the introduction using MLA or discriminant analysis?

Response: Previous studies in Lake Tana focused only northeastern part and have utilized MLC, decision tree, and RF methods. Previous studies have only employed spectral bands as inputs for their classifiers without any atmospheric correction process, resulting in a high rate of misclassification of church forests and irrigated crops as water hyacinths compared to the RF classifiers employed in the present study. When compared to SVM, CART, and other preview methods, an RF classifier using Sentinel 2 spectral bands and water and vegetation indices is an effective classifier for water hyacinth detection in Lake Tana.

Studies conducted in the past have utilized MLC, decision tree, and RF, but none have used discriminant analysis for the detection of water hyacinths in Lake Tana.

How does this study contribute to the broader effort to improve environmental assessment and monitoring via remotely-sensed data?

Response: The research can help to improve decision-making by providing accurate and up-to-date information on water hyacinth infestations, which can aid in the development of effective management and control strategies. Detection of water hyacinth on remote sensing data using machine learning methods during different seasons can provide a more comprehensive understanding of the seasonal variability of the infestation and will contribute to the broader effort of improving environmental assessment and monitoring through remote sensing data by providing a more accurate, cost-effective, and efficient method of monitoring water hyacinth infestations in large bodies of water such as Lake Tana (Page 15).---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Comparing performance and identifying the most efficient machine-learning method and quantify spatial coverage of E. crassipes using the most accurate method will provide a more complete understanding of the seasonal variability and spatial coverage of the infestation, inform management and control strategies, and improve decision-making by providing accurate and up-to-date information on the E. crassipes infestation. --------------------------------------------------------------------------While this study only investigated seasonal changes in total area coverage, accurate monitoring of E. crassipes using remote sensing will allow for the assessment of spatiotemporal changes in coverage among different regions of the Lake Tana shoreline, both within seasons and across years. Detecting E. crassipes infestation from remote sensing data using machine learning methods during different seasons can provide a more comprehensive understanding of the seasonal variability of the infestation and will contribute to the larger effort of improving environmental assessment and monitoring through remote sensing data, as it provides a more accurate, cost-effective, and efficient method of monitoring water hyacinth infestations in large water bodies such as Lake Tana.

There was a statement in the introduction about understanding spatiotemporal dynamics. Building that up more and then coming back to it here would be good.

Response: we have revised the manuscript as per the comments given by the reviewers.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I authors made required changes. I suggest to accept this manuscript.

Thank you

Reviewer 3 Report

All the comments have been addressed.

Reviewer 4 Report

The manuscript has addressed my comments and improved a lot, and I recommend publication.

Reviewer 5 Report

I thank the authors for their responses to my comments. Their edits were appropriate and I feel the manuscript is improved.

Article Menu

Detection of Water Hyacinth (Eichhornia crassipes) in Lake Tana, Ethiopia, Using Machine Learning Algorithms

Further Information

Guidelines

MDPI Initiatives

Follow MDPI