Identification of Pasture Degradation Using Remote Sensing Data and Machine Learning: A Case Study of Obichnik

Abstract

The degradation of pastures and meadows is a global problem with a wide range of impacts. It affects farmers in different ways, such as decreases in cattle production, milk yield, and forage quality. Still, it also has other side effects, such as loss of biodiversity, loss of resources, etc. In this study, the degradation of a semi-natural pasture near the village of Obichnik, Bulgaria, was evaluated using machine learning algorithms, and an unmanned aerial vehicle (UAV) obtained visual spectrum images. A high-quality (HQ) orthomosaic of the area was created and numerous regions of interest were manually marked for training and validation purposes. Three machine learning algorithms were used—Maximum likelihood, Random trees (RT), and Support Vector Machine (SVM). Furthermore, object-based and pixel-based approaches were utilized. The obtained results indicate that the object-based RT and SVM models provide significantly better accuracy, with their Cohen’s Kappa reaching 0.86 and 0.82, respectively. The performed classification showed that approximately 61% of the investigated pasture area is covered with grass, which indicates light-to-medium degradation.

1. Introduction

Pasture and meadow degradation is a global problem affecting not only cattle breeders who depend on healthy pastures for their livelihoods but also other people who suffer from the resulting hydrological disturbances, sandstorms, resource loss, and social consequences, including population displacement. Their health has a direct and indirect impact on biodiversity, as local flora and fauna are adapted to long-term evolutionary processes that have shaped these ecosystems [1].

Rangeland degradation has become an emerging topic in pasture management and environmental protection. Although many authors and international organizations call for the protection of natural grasslands, there is no consensus definition of what constitutes natural grassland degradation. Conceptual definitions for degradation are very broad, including declines in rangeland quality, decreases in vegetation cover, productivity, economic potential, service function, recovery capacity, and diversity [2,3,4]. Others focus only on reductions in forage production or grassland productivity, invasions by non-native plant species, dominance by unpalatable species, or functions related to biochemical and water cycles [5,6,7,8]. To sum up, the term ‘degradation’, referring to the condition of natural grasslands and pastures, is widely used in different circumstances by different stakeholders, from researchers and rural extension agents to policy makers, without a commonly agreed definition.

The causes of rangeland degradation are generally attributed to a combination of overgrazing, unscientific livestock management, historical-cultural barriers to adopting modern livestock management concepts, global climate change, and soil damage by small mammals [9]. From the perspective of farmers, pasture and meadow degradation reduces the economic performance and sustainability of livestock production, which may result in a reduction in the body weight and wool yield of sheep [10], forage quality [11], cattle production [12], milk production [13], etc.

Another important aspect is the criterion used for assessing the degradation of pastures and grasslands. The United Nations has proposed the changes in the land cover, productivity, and carbon reserves as sub-indicators for monitoring land degradation processes [14]. Other monitoring indicators that are often used are vegetation cover, vegetation height, aboveground biomass, grass yield, and net primary productivity, especially in remote-sensing scenarios [15]. Nevertheless, scientists classify the degradation of pasture/meadow/grassland into different categories, using different assessment criteria. In [16], the pastures in the state of Minas Gerais, Brazil, were categorized into only two classes, based on their chemical and physical characteristics: healthy pasture and degraded pasture. In another Brazilian study [17], three degradation levels of pastures are used, and the chosen criterion is the share of exposed soil. Level 1, no or low degradation (≤20% exposed soil); Level 2, medium degradation (20–45% exposed soil); and Level 3, high degradation (>45% soil exposed). Another study for Brazil assessed the degradation of humid tropic pastures in four classes, based on the erosion of the soil [18]: no visible degradation; light degradation (pastures with sheet erosion or shallow cattle tracks); moderate degradation (some vertical erosion is visible, such as gullies); and strong degradation (pastures that are partly interrupted and vertically broken with cattle tracks as well as large gullies). In a study dealing with Kazakhstani pastures [19], the vegetation cover was the main criterion for assessing their degradation. Furthermore, it applied the Drude scale [20], where the following classes are defined:

• Soc.: plants grow close to each other, joining with their aboveground parts (90–100%);
• Cop. 3: plants occur in very large numbers (75–90%);
• Cop. 2: plants occur in large numbers (50–75%);
• Cop. 1: plants occur in considerable numbers (25–50%);
• Cop. 1-Sp.: the species are relatively abundant (15–25%);
• Sp.: the species are abundant, but do not form a continuous cover (5–15%);
• Sol.-Sp.: low abundance (1–5%);
• Sol.: the species grow sparsely (<1%);
• Un.: the species occur in single instances.

In [21], the grassland degradation was assessed in Northern Tibet. The suggested classification was based on the grassland degradation index (GDI):

$$GDI={\displaystyle \frac{\left({\sum }_{i=1}^5{D}_i\times A_i\right)}{A}},$$

(1)

where $D_i$ is the grading score of grassland degradation, $A_i$ is the distribution area of grade $i$, and $A$ is the total area of the investigated grassland. The degradation was then categorized into five classes: undegraded (GDI ≤ 1); light degradation (GDI ≤ 2); moderate degradation (GDI ≤ 3); serious degradation (GDI ≤ 4); and extremely serious degradation (GDI > 4).

Another approach is the classification of degradation, based on the coverage of a certain plant or the canopy coverage. In [22], the grassland degradation in China was assessed based on the coverage of Stellera chamaejasme (SC). Five degradation classes were defined as follows: degradation stage 1 (SC ≤ 19%); degradation stage 2 (SC ≤ 39%); degradation stage 3 (SC ≤ 59%); degradation stage 4 (SC ≤ 79%); and degradation stage 5 (SC ≤ 95%). In another study aimed at assessing the grassland degradation in Inner Mongolia, China, the canopy coverage (CC) was selected as a criterion [23], and three degradation classes were defined based on it: dense grassland (CC > 50%); moderately dense grassland (CC > 20%); sparse grassland (CC > 5%). Similarly, in [24] the vegetation cover (VC) was used as the criterion for assessing degradation: none (VC > 80%); light to moderate (VC > 60%); moderate to severe (VC > 30%); and extreme (VC < 30%).

From the performed analysis, it can be seen that numerous criteria and classifications exist, which could be used to assess the degradation of pastures, meadows and grasslands. In each individual study, the criterion and the scale/categories of degradation were chosen individually, depending on the local specifics of the investigated terrain: the observed problems, the species composition, the biodiversity, their impact on the farmers’ production, etc.

With the advancement of unmanned aerial vehicles (UAV) and satellite imaging technologies, the interest in management of grassland and pastures, their degradation, desertification, etc., has increased [25,26,27]. High-spatial-resolution drone and satellite-obtained data are becoming increasingly cheap and globally available, meaning that the use of these datasets for rangeland mapping, monitoring, and tracking is affordable [28,29]. The results shown in [30] confirm the possibility of UAV imagery being applied as a tool for precision grassland surveys. The authors reached R² = 0.62 for their best-performing biomass estimation model and concluded that such an approach can be an alternative to conventional field methods. Different approaches are used when assessing the quality and efficiency of pastures with remote sensing, which are based mostly on vegetation indices and machine learning (ML). According to [31], the first Scopus publication on the application of ML in pasture evaluation is from 2013, and by 2020, the number of articles on the topic increased to 456, which shows that the application of ML algorithms with other metrics gives promising results. The study also stated that the most commonly used ML algorithms for this purpose are random forest (RF) and partial least squares regression (PLSR).

Different studies have used ML regression models for the evaluation of aboveground biomass and dry matter of grasslands and pastures. In [32], machine learning-based predictive models were trained for the estimation of aboveground biomass using satellite and UAV-obtained data. For large-scale evaluation, the Huber model returned the best results with R² = 0.60, while for short-scale the best-performing models were K-nearest neighbours with R² = 0.76 and Extra trees with R² = 0.75. The grassland biomass was also estimated in [33], using a combination of Sentinel-2 imagery with machine learning algorithms. Even though the authors did not test their approach on large-scale grasslands, their best-performing model achieved R² = 0.6, which they consider enough to support farmers in the optimization of the biomass management of their pastures. In [34], a UAV-mounted multispectral camera was used, in order to estimate the grass stand of pastures. Using a combination of statistical models and machine learning allowed predicting the pre-grazing pasture biomass production with an error of less than 20%. In [35], computer vision algorithms were used to extract the biophysical parameters from images, which were then used to train machine learning algorithms that predict grassland height and biomass. The study achieved up to R² = 0.656 for pasture height estimation and R² = 0.496 for biomass estimation.

Other studies have used ML classification algorithms for categorization of the types of grasslands/meadows based on different factors. In [36], the authors used support vector machine (SVM), RF, neural network, 3D fully convolutional network, 3D convolutional neural network, and the systematic classification method (SCM) with satellite-obtained multispectral images of grassland, in order to identify the types of vegetation. A large amount of ground-based reference data were used for training and validation, which allowed reaching accuracies between 43.42% to 99% for the different species with the SCM method. In [37], the authors tried to identify pastures in Brazil using Landsat 8 data and the RF algorithm. The study created a new pasture map for Brazil and reached an overall accuracy of 87%. In [38], an object-based image analysis was applied to Sentinel-2 data with seven vegetation indices in order to classify different types of pastures (herbaceous and shrubby). The authors used the simple linear iterative clustering (SLIC) algorithm for image segmentation and the random forest for image classification, which allowed them to achieve 87% accuracy. In another study, a pasture was classified into grazed and ungrazed categories, using ground-based camera images and the Modified Green Red Vegetation Index [39].

Different studies dealing specifically with the assessment of grasslands and pasture degradation exist, which are based on appropriate vegetation indices. In [40], shrub degradation of grasslands in Kyrgyzstan was identified based on Google Earth data. The study results showed that the shrubs’ normalized difference vegetation index (NDVI) is higher, by approximately 0.20, than the grass NDVI during the summer months. In another study, the grassland degradation index (GDI) and satellite-obtained images were used to classify the degradation of grasslands [41]. The vegetable pasture cover was derived from NDVI, and with its help the GDI was estimated. NDVI and satellite-obtained data were also used in [42], in order to assess the degradation of pastures in Brazil over a five-level scale. The obtained degradation map was validated using ground samples, and 84.2% overall accuracy was achieved. In [16], pixel-based analysis was also applied, in order to estimate the degradation of pastures. LANDSAT-8 and Google Earth data were analyzed over a three-year period in order to obtain four vegetation index signatures of healthy and degraded pastures (NDVI and three RGB indices: total brightness quotient of blue, green and blue). Similarly, in [19], a degraded pasture land in Kazakhstan was investigated with the use of satellite imaging from Google Earth, Landsat 5 and 8 as well as other reference data. Thereafter, isolinear maps were created by dividing the area of herbaceous or woody vegetation by the total area. In [24], NDVI and the modified soil-adjusted vegetation index 2 (MSAVI2) were used on WorldView-2 satellite data to assess grassland degradation in the Georgian Caucasus. Thereafter, the random forest regression model was used to fit the results.

Other studies have used machine learning to evaluate grassland degradation. In [22], the authors tried to assess it based on the percent of Stellera chamaejasme coverage. Data from ground-based cameras and deep neural networks allowed a success rate between 84% and 96% to be achieved for the different classes. In another study, multispectral and multitemporal data from two satellites (RapidEye Ortho 3A and Landsat 5 TM L1T) were used to identify the degradation of a pasture [18]. The images were analyzed using the random forest classification algorithm, where a total of 210 regions of interest from the four degradation classes were identified. Furthermore, a total of 460 polygons were used for training and validation purposes. The obtained overall accuracy for the pasture degradation data reached 61% using the Landsat data, 75% using the RapidEye data and 77.5% using the combined data. In [43], the decision tree method was used to analyze NDVI data, based on Landsat images, aimed at predicting land degradation in Brazil. The validation of the decision tree achieved Kappa 0.82.

As can be seen from the performed analysis, there are many studies dealing with management of pastures and meadows. They are aimed at assessing the pasture biomass and dry matter, identifying the vegetation species, monitoring the regenerative grazing, large-scale categorization of pasture lands, etc. There are also many studies specifically aimed at assessing and monitoring degradation, though most of them rely on vegetation indices and multispectral data. So far, very few authors have used the machine learning approach in a combination with visual spectrum images for remote assessment of pasture degradation. Considering also the broad understanding of the term degradation by the different stakeholders, a visible research gap can be observed.

The main objective of this work is to investigate the efficacy of different supervised pixel-based and object-based machine learning algorithms for remote assessment of the degradation of a pasture near the village of Obichnik, Bulgaria. To achieve this, appropriate degradation classes should be defined, corresponding to the specifics and problems of the investigated location. This study relies on UAV-obtained data from visual spectrum sensors, though satellite-obtained data are also applicable. The obtained results should provide new knowledge about the application of remote sensing and machine learning for evaluation of pasture and meadow degradation.

The rest of the paper is organized as follows. Section 2 describes the object and means of the investigation, as well as the methodology for data collection and data analysis. Section 3 presents the obtained row data and their processing, application, and evaluation by three machine learning algorithms—maximum likelihood (MxL), random trees (RT), and SVM. Furthermore, the optimal models are compared in terms of Kappa, precision, and recall in order to gain an in-depth understanding of their performance. The obtained accuracies are also compared with those obtained in other studies, in order to evaluate the effectiveness of the proposed approach. Finally, Section 4 presents the conclusions from the study.

2. Materials and Methods

2.1. The Study Area

The researched area (Figure 1) is a pasture near the village of Obichnik, Momchilgrad municipality, Kardzhali region, Bulgaria (coordinates 41.494062, 25.478968). The pasture is located in the mountainous region of the Eastern Rhodopes and is characterized as semi-natural dry pastures. The altitude is about 546 m and also falls within the Natura 2000 site under Directive 92/43/EEC (Habitats Directive).

The observed pasture territory is freely used by cattle, and no grazing system is applied. Therefore, in the long term, the species composition of the pasture is greatly disturbed, which also leads to a change in the ratio between cereal grasses, leguminous grasses, and other species, as well as lower-quality grasses. As a result, exposed areas are also observed, due to a decrease in species that are involved in the formation of healthy turf. Previously conducted field studies show that the observed territory is characterized by a rich species composition and diversity of the biological and economic qualities of the species. In the composition of the vegetation, 73 species of higher plants from 22 botanical families are found.

Three types of degradation could be identified on the investigated pasture:

• Shrubs and trees (Figure 2a)—since no grazing system is applied on the pasture, shrubs and trees are steadily increasing their share;
• Stones and rocks (Figure 2b)—the pasture is surrounded by abandoned old houses and other constructions, which are the source of this type of degradation;
• Exposed soil and cattle tracks (Figure 2c)—they are created mostly by the grazing animals themselves. One of the reasons for this is the slope of the terrain, which is up to 15%.

2.2. Data Acquisition and Data Analysis

Data collection was carried out using a DJI Mavic 2 Enterprise Advanced UAV by SZ DJI Technology Co., Ltd. (Shenzhen, China). The characteristics regarding the applicability of this type of drone are related to the range of visibility that is provided, the possible speeds of movement, the capabilities of the camera, and the possibilities of use in windy weather. DJI Mavic 2 Enterprise Advanced is characterized by a speed of up to 50 kph in positioning mode, and the vertical and horizontal accuracy ranges are ±0.5 m and ±1.5 m, respectively. Its max flight time is between 24 and 31 min and it has resistance to wind speeds of up to 10 m/s. The gimbal has 3-axis (tilt, roll, and pan) stabilization, and the max control speed on a tilt is 120°/s with an angular vibration range of ±0.005°. The most important element for this study is the camera of the UAV. It uses a ½” CMOS 48 Mpx visual spectrum sensor with a resolution of 8000 × 6000 px. For the needs of this study, the UAV has been set up to take photos with ISO 100, an 80% overlap, and the relative flying altitude and flying speed were 40 m and 30 km/h, respectively.

The image processing and analysis have been implemented using the software ArcGIS Pro v. 2.8.6, developed by Esri Inc. (California, CA 92373, USA). This tool provides a wide range of options, such as combining overlapping images into a single orthomosaic map, generation of classification maps based on vegetation indices, machine learning, deep learning, etc. Furthermore, ArcGIS has numerous integrated tools for raster- and object-based analysis.

2.3. Methodology of the Study

Considering the available types of degradation on the investigated pasture, the following classes are defined as follows:

• Grass—this class corresponds to areas with no or insignificant degradation, which are covered by grass vegetation;
• Shrubs—this class corresponds to areas which are covered with bushes and in the long term with trees;
• Soil—this class represents areas with exposed soil and cattle tracks;
• Stones—this class represents areas which are covered with stones and rocks.

The methodology of this study is summarized in Figure 3. It includes the following steps:

• Step 1. A UAV is used to make overlapping images of the investigated pasture;
• Step 2. The overlapping images are combined into one big high-quality (HQ) map, representing the investigated pasture area;
• Step 3. The HQ map is used to select numerous polygons (regions of interest) from each class, to be used for training and validation purposes;
• Step 4. The HQ map is used to generate a segmentation map, by manipulating three parameters of the pixels: spectral detail, spatial detail, and minimum segment size in pixels;
• Step 5. Using the combined HQ map, the training data, and the segmentation map, a classification map is created, using one of the available algorithms;
• Step 6. The accuracy of the classification map is assessed by using randomly selected pixels from the chosen polygons in Step 3 and comparing them with the reference class. Steps 4, 5, and 6 are repeated numerous times until the best-performing models are obtained;
• Step 7. Once the optimal classification models are obtained, they are used to evaluate and analyze the degradation of the pasture.

Figure 3. Overview of the study methodology.

Figure 3. Overview of the study methodology.

In this study, two types of classification approaches are used and compared: pixel-based and object-based. Furthermore, for each one of them the following algorithms are applied as follows:

-

Maximum likelihood—the idea behind the maximum likelihood is to estimate the parameters of a probabilistic model by maximizing the likelihood function, which measures how well the model explains the observed data. By iteratively adjusting the model parameters, MxL finds the parameter values that make the observed data most probable under the model [44];

-

Random trees—a very straightforward and easy-to-implement algorithm, based on the idea of using a random subset of features and samples from the training data. The randomness improves the models’ accuracy and the overfitting, leading to a robust and more generalized model [45];

-

Support Vector Machine—it is especially useful for models that work well with a small dataset or to separate the data into clearly distinguished classes. The goal of SVM is to find a hyperplane in multidimensional space that separates different classes in classification tasks. A hyperplane is a geometric representation of the dividing line in two-dimensional space or the plane in three-dimensional space, for example. The margin is the distance between the nearest points of the two classes and the hyperplane. The goal of SVM is to maximize this margin, because a wider margin usually leads to better generalization and fewer errors on new data. The support vectors are the data points that are closest to the hyperplane. They are key to defining the hyperplane and pattern, as the margin is calculated relative to them [46].

3. Results and Discussion

An experimental study at the investigated pasture was conducted on 21 April 2023. The UAV made a total of 39 overlapping images of parts of the pasture according to the trajectory shown in Figure 4a. Next, using the ArgGIS Pro v. 2.8.6 software, a combined HQ map of the pasture was generated (Figure 4b).

According to Step 3 of the developed methodology, the HQ map was used to select a total of 280 regions of interest (ROI) for training and validation purposes (Figure 4c), covering a total of 4,004,502 pixels. Their distribution among the classes is summarized in

Table 1. ROI and pixels for each class, that are used for training and validation purposes.

Class	Number of ROI	Number of Pixels	Relative Area
Grass	94	1,220,995	30.5%
Shrubs	92	2,626,642	65.6%
Soil	32	24,452	0.6%
Stones	62	132,413	3.3%
Total	280	4,004,502	100.0%

. It can be seen that a high imbalance exists between the different classes. The used ROI from the “Soil” and “Stones” classes are 0.6% and 3.3% of the total number of pixels, respectively, and this mostly corresponds to the actual situation on the investigated pasture. Therefore, this characteristic of the training data should be considered when analyzing the obtained results.

Next, steps 4, 5, and 6 of the methodology were repeated numerous times, in order to estimate the optimal segmentation parameters, which returned the highest accuracy results for the object-based classification. The highest results were obtained with the following segmentation characteristics: spectral detail—14, spatial detail—2, and minimum segment size in pixels—50. The generated optimal classification maps using the three object-based algorithms are presented in Figure 5. Similarly, the generated maps using the pixel-based algorithms are shown in Figure 6.

The accuracy of the six classification maps has been assessed using 10,000 stratified random points, which are randomly distributed within each class, proportionally to their relative area. The overall accuracy of each classification method is presented in

Table 2. Comparison of the different algorithms’ overall accuracy.

No.	Classification Algorithm	Cohen’s Kappa
1	Optimal object-based MxL	0.53
2	Optimal object-based RT	0.86
3	Optimal object-based SVM	0.82
4	Pixel-based MxL	0.54
5	Pixel-based RT	0.41
6	Pixel-based SVM	0.43

and is expressed in terms of Cohen’s Kappa. The Kappa coefficient is a statistical measure, used for evaluating the performance of the trained models. It could be interpreted as follows [47]: k ≤ 0—no agreement; 0.01 ≤ k ≤ 0.20—none to slight agreement; 0.21 ≤ k ≤ 0.40—fair agreement; 0.41 ≤ k ≤ 0.60—moderate agreement; 0.61 ≤ k ≤ 0.80—substantial agreement; 0.81 ≤ k ≤ 1.00—almost perfect agreement. The obtained results show that the object-based RT classification is the best-performing model with Kappa 0.86, followed by the object-based SVM classification with Kappa 0.82. These values indicate that there is almost perfect agreement between the validation data and the predicted data. All other algorithms (pixel-based and object-based) returned worse results, which is indicated by their lower Kappa, ranging between 0.41 and 0.54, i.e., between “moderate agreement” and “substantial agreement”.

The obtained results can be compared with those obtained in similar studies, in order to evaluate the efficacy of the proposed approach. In [43], a decision tree model with validation Kappa 0.82 was obtained. These results are practically identical to the optimal SVM model in this study and slightly lower than the optimal RT model. Similar results were also received in [18], where for a RF model the obtained Cohen’s Kappa varies between 0.85 and 0.88, depending on the satellite map used. Similarly, in [24], a coefficient of determination between the predicted and observed values of R² = 0.79 was obtained for the pixel-based analysis of grassland degradation, which is slightly lower compared to our results. Nevertheless, when performing such analyses, one should consider that the other studies used different classification categories, so no direct comparison could be made.

Considering the high imbalance in the training data between the different classes, it is worth going deeper into the confusion matrices of the object-based RT and SVM classifications, as the best-performing algorithms (

Table 3. Confusion matrix for object-based RT classification.

		Reference				Precision	Recall	F-Score
		Grass	Stones	Soil	Shrubs
Results	Grass	3052	0	10	343	0.896	0.993	0.942
	Stones	0	251	0	191	0.568	0.787	0.660
	Soil	0	0	47	89	0.346	0.825	0.487
	Shrubs	21	68	0	5922	0.985	0.905	0.943

and

Table 4. Confusion matrix for object-based SVM classification.

		Reference				Precision	Recall	F-Score
		Grass	Stones	Soil	Shrubs
Results	Grass	3020	8	10	413	0.875	0.983	0.926
	Stones	0	259	0	387	0.401	0.812	0.537
	Soil	0	3	39	0	0.929	0.684	0.788
	Shrubs	53	49	8	5744	0.981	0.878	0.927

, respectively). If compared, it can be seen that their precision and recall when identifying grasses and shrubs are almost identical. There is some difference when it comes to the identification of stones. It can be seen that the RT classification has 0.568 precision, compared to 0.401 of the SVM one, which indicates that the latter misclassified more than half of the stones as shrubs. On the other hand, the SVM algorithm returns fewer false positives, as its recall is 0.812, compared to 0.787 of the RT one. Once again, the false positives come mostly from misclassifying stones as shrubs.

Significantly different is the situation when it comes to the identification of soil. The SVM algorithm performs significantly better, and its precision reaches 0.929, compared to 0.346 for RT. The reason for this is that the RT algorithm identifies many shrubs as soil. On the other hand, SVM returns more false positives and the recall reaches 0.684, compared to 0.825 for the RT algorithm.

Examples of the misclassification of stones and shrubs are presented in Figure 7. In the upper part of the image, a large area of different shrubs (Shrubs 1) is selected, some of which were falsely classified as stones by the SVM algorithm. False-positive stones are also classified with the RT algorithm, though the areas are smaller. If we zoom in, it can be seen that the main reason for this is the color of the shrubs, as some of them are peeling. This allows us to speculate that if the investigation of the pasture degradation is made one month later when all the bushes and trees have completely turned green, the accuracy of both algorithms will increase.

In the bottom part of Figure 7 are shown samples for misclassifying stones as shrubs. The first marked stones area (Stones 1) is characterized by small white stones with a black background, and the second area (Stones 2) is characterized by small white stones surrounded by green background (grass and/or shrubs). It can be seen that both algorithms fail to correctly identify large parts of the area “Stones 1”. For area “Stones 2”, the SVM algorithm performs slightly better than the RT one. Nevertheless, this particular misclassification is not that crucial, as such mixed areas are hard to classify even with a human intellect.

The top part of Figure 8 presents a large marked area of bare soil. It can be seen that both algorithms fail to identify some parts of it, though the RT classification generally performs significantly better in this particular case. In all cases, the misclassified soil areas are falsely classified as shrubs. Furthermore, two areas of shrubs were marked. It can be seen that the RT algorithm has some misclassified grass areas, which are considered to be shrubs. On the other hand, the SVM algorithm is misclassifying some of the shrub areas as grass.

The bottom part of Figure 8 shows another example that represents a large grass area with numerous tiny defects, such as small stones and small soil areas. Both algorithms are able to identify them relatively well; however, RT classifies all small soil areas as shrubs, while SVM classifies them correctly in most cases.

Finally, the degradation of the pasture was assessed using the two best-performing algorithms and is summarized in

Table 5. Analysis of the pasture’s degradation using the selected optimal models.

Class	Area, m2	Relative Area, %
Optimal object-based RT classification
Grass	6065	61.50
Stones	102	1.03
Soil	290	2.94
Shrubs	3405	34.53
Optimal object-based SVM classification
Grass	6037	61.21
Stones	221	2.24
Soil	111	1.13
Shrubs	3493	35.42

. It can be seen that there is no significant difference between the classification models, based on RT and SVM. The relative area for classes “Grass” and “Shrubs” is almost identical, with less than 1% difference. The difference in the results of RT and SVM for the “Stones” class is slightly more than 1%, though this corresponds to an almost 100% relative difference. Similarly, for the “Soil” class, the difference between the two algorithms is less than 2%, though this corresponds to almost 200% relative difference. Nevertheless, from the point of view of a farmer, such differences can be considered insignificant, as the results for the “Grass” class are practically identical, and this is enough to evaluate the degree of degradation. The reason for this is that any further actions that are going to be taken to recover the degraded pasture will require an in-situ evaluation and micro-management.

If the degradation scale from [24] is used, the investigated pasture can be classified with “Light to moderate” degradation, which is characterized by vegetation coverage between 60% and 80%.

4. Conclusions

This study investigates the effectiveness of several machine learning methods for the identification of different types of pasture degradation using UAV-obtained visual spectrum images. The data were obtained from a pasture, located in the Rhodopa mountain near the village of Obichnik, Kardzhali region, Bulgaria. Unlike previous studies, where the degradation was estimated in terms of severity, in this paper it was assessed by classifying the pasture’s surface into four categories (grass, stones, soil, and shrubs) and by obtaining their shares.

Three pixel-based and three object-based models were trained, and their performance was assessed using Cohen’s Kappa. The obtained results indicate that the object-based RT and SVM models perform better, which is indicated by their Kappa coefficients—0.86 and 0.82, respectively. The comparison of their confusion matrices showed that the two models returned very accurate results for the “Grass” and “Shrubs” classes, with F1 scores ranging between 0.926 and 0.943. The RT model gave better results when identifying “Stones” (F1 = 0.660); however, SVM performed better with the identification of “Soil” (F1 = 0.788). In general, both algorithms showed that approximately 61% of the investigated pasture’s surface can be classified as “grass” and approximately 35% can be classified as “shrubs”, which indicates that, currently, it has light to medium degradation. It is important to continue the monitoring of the pasture’s degradation and identify appropriate measures for improving its state.

The obtained results also prove that the application of machine learning with visual spectrum data for the identification of pasture degradation can be as effective as other approaches, based on multispectral data and vegetation indices, with or without machine learning. This makes it applicable to small and medium-sized farmers using low-cost UAVs. Such an approach would allow them to optimize pastures and grasslands management and increase their financial results.

It is important to verify if the accuracies of the ML algorithms would be influenced by the annual time of UAV inspections and compare them with the results in this study. It is also of interest to investigate the application of satellite-obtained RGB images with machine learning for identification of pasture degradation, in order to compare the results with the UAV-obtained ones. Such comparisons were not performed in this study and are an object for future research.