1. Introduction
Water is vital for the life of humans, animals, plants, and ecosystems. Human health, food security, economic growth, energy production, and ecosystems are all water-dependent. Growing population and urbanization, intensive industrial development, agriculture, increasing demand, and misuse of water have increased water stress, making water a scarce and expensive resource, especially in undeveloped countries.
This growing issue has been recognized and several policies have been adopted in order to provide sustainable management and prevent further decreases in water quality and quantity. The 2030 Agenda for Sustainable Development [
1], adopted by United Nations Member states, within Sustainable Development Goals (SDG) 6 [
2] emphasizes the water-related issue. SDG 6 has eight targets including water quality. In Europe, the Water Framework Directive (WFD) [
3] defines a framework for the protection of the aquatic environment (rivers, lakes, transitional waters, groundwaters, and coastal waters.). The primary aim of WFD is to achieve at least a good status in all water bodies. To assess the status of the water bodies, monitoring of biological, hydromorphological, and physicochemical water quality parameters (WQP) as defined within Annex V and Annex X [
4] needs to be conducted.
The WFD implies that rivers with catchment areas greater than 10 km
2 and lakes greater than 0.5 km
2 in surface area and all of the water bodies into which priority substances are discharged need to be included within the water status assessment and monitoring. WQP is traditionally determined by the collection of in-situ samples and then analyzing them in the laboratory [
3]. Although this method provides high accuracy, it is labor, time, and cost-intensive. Therefore, monitoring all water bodies as defined by WFD would require major financial investments. Moreover, the conventional methodology determines the WQP concentration at the sampling point. The water quality within water bodies is rarely constant due to unpredictable events such as storms, accidental spillages, or leakages. and it is highly influenced by hydrodynamic characteristics such as flow direction and discharge. Due to that the monitoring of spatial and temporal variations and trends in large water bodies by conventional methods is challenging.
To overcome those limits, remote sensing technologies, which have the advantage of large spatial coverage and high temporal resolution, have been used to identify and monitor water bodies more effectively and efficiently [
5,
6,
7]. The remote sensing monitoring of WQP is based on establishing the correlation between in-situ monitoring data and corresponding surface reflection. The spectral characteristics of water are functions of the hydrological, biological, and chemical characteristics of water [
8]. Therefore, the amount of radiation at various wavelengths reflected from the water surface can be used directly or indirectly to detect WQP.
The clear water reflects light with wavelengths < 600 nm, resulting in high reflectance in the blue-green while absorbing radiation at the Near-Infra Red (NIR) portion of the spectrum and beyond. The increase of chlorophyll-a (chl-a) concentration increases absorption in Red (R) and strongly absorbs Blue (B) light while the reflection peak is located at the green (G) part of the spectrum [
9]. Water clarity is the function of Total Suspended Solids (TSS) concentration. TSS is the measure of the weight of inorganic particulates suspended in the water column and it is responsible for most of the scattering [
10]. By influencing the scattering of light, TSS directly controls the transparency and oxygen content of the water body [
11]. The increased concentration of TSS causes the peak to shift from G toward the R region and increases water reflectance in the NIR region.
Thus, many studies have used band combinations and spectral indices to develop empirical algorithms for the estimation of optical active WQP and achieved good results [
12,
13]. Various spectral bands have been used to quantify the chl-a and TSS (
Table 1).
However, inland waters are seriously affected by human activities, due to optical properties being complex and highly variable. Therefore, each band is not only sensitive to one but also to other WQP which can lead to significant uncertainty in the results produced.
In addition, WQPs such as Total Nitrogen (TN), Total Phosphorus (TP), and Dissolved Oxygen (DO) are important information for understanding water body dynamics. Increased levels of nutrients can lead to algal blooms and oxygen depletion.
However, since the relationship between surface reflectance and concentration of those parameters is indirect and non-linear, the estimation of their concentration represents a great challenge if they are based on traditional empirical algorithms. In recent years, with the increase in processing power and the development of artificial intelligence, machine learning (ML) algorithms have been increasingly used for WQP monitoring. The most common ML models for water quality parameters are Random Forest (RF), Supported Vector Machine (SVM) and Artificial Neural Network (ANN).
Guo et al. [
23] used the Landsat and MODIS reflection and SVM for monitoring of DO in Lake Huron. Results show good robustness with average R
2 = 0.91. Qian et al. [
24] tested Multiple Linear Regression (MLR), SVM, RF and ANN for monitoring of three non-optical (pH, DO, Electrical Conductivity (EC)) and one optical parameter (Turbidity) at Qingcaosha Reservoir based on Sentinel 2 images. The results indicated that ANN showed more robust performance for all WQP (RMSE: 0.33; 0.49; 0.38; 0.26 for pH, DO, EC, and Turbidity, respectively) compared to traditional ML algorithms. Guo et al. [
25] monitored the TP, TN, and Chemical Oxygen Demand (COD) by using Sentinel 2 imagery and NN, RF, and SVM algorithms. Their results showed that ML can significantly improve the estimation accuracy of non-optical parameters with Normalized Root Mean Square Error (NRMSE) of TP: 16.8%; TN: 29.64% and COD 18.75. Similarly, Ref. [
26] tested the performance of MLR, SVM, and ANN for monitoring of chl-a, DO, Turbidity, blue-green algae (BGA), and fluorescent dissolved organic matter (fDOM) from Sentinel 2 and Landsat 8 images. The DNN outperformed the ML algorithms resulting in Root Mean Square Error (RMSE) of 0.86, 7.56, 1.81, 14.50, and 5.19 for BGA, chl-a, DO, fDOM, and Turbidity, respectively. Hafeez et al. [
20] estimated the concentration of TSS, chl-a and Turbidity with several ML algorithms including ANN, RF, and SVM by using Landsat (5, 7, 8) imagery. ANN outperformed RMSE chl-a:1.4; TSS: 2; Turbidity: 3.10) followed by SVM. Leggesse et al. [
27] compared the six ML algorithms integrated with Landsat 8 imagery for the prediction of three optically active WQP (chl-a, Turbidity and Total Dissolved solids (TDS)). The results indicated that XGBoost regression performed best for chl-a (RMSE: 9.47) while RF performed best for the rest of the parameters (RMSE TDS: 12.3; Turbidity: 7.82) while ANN and SVM provided lower accuracy. Gomez et al. [
28] tested the performance of RF, SVM and ANN on a balanced dataset for the monitoring of chl-a based on Sentinel 2 images. The results showed that RF performed better compared to others (RMSE: RF 0.82; SVM 1.45; ANN 1.75).
It has been shown that ANN and SVM have provided excellent performance in monitoring both optically active and non-active WQP [
20,
26,
28,
29]. ANN, as a nonlinear approximation method, is more flexible for WQP monitoring. However, the resulting accuracy of ML is generally a function of the selected model and the quality and size of the training data. The development of an ANN model requires large training datasets and extensive experience in order to determine the optimal NN architecture. Using too many layers can result in overfitting, which involves the fitting of noise in training data and lower generalization to new data [
30]. On the other hand, a low number of layers can lead to underfitting when the model cannot represent the complexity of data adequately. Due to that, SVM and RF can have a higher generalization ability than ANN. Govedarica et al. [
7] tested the performance of ANN and SVM for monitoring Turbidity, TSS, TN, and TP. The results showed that SVM outperformed ANN for Landsat 8 data while ANN produced better results for Sentinel 2 data. The reason for the higher performance of SVM can be due to being less sensitive to small data samples and mixed pixels [
30,
31] and it avoids the occurrence of overtraining and optimization of fewer parameters [
32,
33]. However, an increase in the number of training data can make SVM difficult to implement.
On the other hand [
27,
28] show that RF had better generalization ability and was less affected by overfitting compared with ANN and SVM. It was noticed that there was an increase in RF performance with an increase in the number of features used in the prediction [
28] while it can be decreased for small training datasets [
34,
35]. The RF algorithm is characterized by the considerable time expenditures for training the trees in the ensemble when the datasets are large [
36]. Compared to SVM, RF can take up to four times longer to train and optimize [
37].
In addition to ML, deep learning algorithms (DL) have been widely applied in remote sensing image classification. Convolution Neural Networks (CNN) are capable of extracting intrinsic features and have provided state-of-the-art accuracy. Pu et al. [
38] used CNN to classify the water quality of a lake based on Landsat 8 images. The results showed that CNN outperformed SVM and RF (OA: CNN 97.12%; SVM 96.89%; RF 86.15%). Cui et al. [
39] used CNN and a combination of Landsat 8 and Sentinel 2 images for monitoring water transparency reaching an R
2 of 0.85. Similarly, Ref. [
40] demonstrated chl-a retrieved from Sentinel-2 images using CNN regression resulting in an R
2 of 0.92. Although CNN has demonstrated increased accuracy and robustness, most of the research that is based on moderate-resolution satellite images deals with large water bodies such as lakes, and transitional or coastal waters. This is mostly due to the fact that CNN uses convolution filters of varying sizes (3 × 3, 5 × 5, or 7 × 7 pixels) to extract meaningful higher-level abstract features and increase accuracy. However, taking into account spatial resolution and the width of rivers these patches can represent heterogeneous classes limiting the accuracy of the model [
40].
The main aims of this paper are (a) to develop a comprehensive ANN-based model for monitoring water body status, and (b) to test the usability of the developed model in real-case scenarios.
5. Conclusions
The study successfully established a robust water quality monitoring program using a 38-year time series of Landsat and in-situ data, coupled with a back-propagated Artificial Neural Network (ANN) model. This model demonstrated high accuracy in monitoring various water quality parameters (WQP), showcasing its potential for sustainable water resource management.
The correlation analysis revealed strong associations between remote sensing data and specific WQPs, such as Total Suspended Solids (TSS), chlorophyll-a (chl-a), Dissolved Oxygen (DO), Total Nitrogen (TN), and Total Phosphorus (TP). Optimal band combinations for each parameter were identified, providing valuable insights into the spectral relationships aiding accurate WQP estimation.
The ANN-based model exhibited exceptional accuracy, particularly for optically active parameters like chl-a and TSS, surpassing results from previous studies that used different remote sensing techniques. This underscores the superiority of the developed model in achieving high precision in WQP estimations, surpassing various existing methods and algorithms.
The study highlighted the efficacy of Neural Networks as a nonlinear approximation method for WQP monitoring. It outperformed other techniques but emphasized the necessity for substantial training datasets to avoid overfitting or underfitting. Optimal input data selection and the use of extensive time series data contributed significantly to model accuracy enhancement.
The independent validation of the developed model revealed a strong ability to classify WQP concentrations accurately. Notably, while the models for TSS and TN provided consistent and accurate classifications, DO estimation faced challenges, especially during high variation periods. This underscores the complexity of DO dynamics in water bodies, particularly during seasonal shifts.
Regarding further research, exploring the integration of additional data sources, such as high-resolution imagery or meteorological data, could further refine the model’s accuracy. Incorporating these data could potentially improve predictions by capturing more intricate environmental parameters that contribute to water quality dynamics. This would contribute to advancing the understanding of the model’s robustness and applicability in different environmental contexts, potentially improving its performance and expanding its utility for broader water quality monitoring and management objectives. Aligning when the sampling data are obtained with satellite overpasses would also be recommended in order to increase the accuracy of the models.
Addressing these weaknesses could potentially strengthen the paper’s findings by providing a more comprehensive assessment of the model’s performance in real scenarios and expanding its applicability across various water body sizes and types.