Monitoring the Area Change in the Three Gorges Reservoir Riparian Zone Based on Genetic Algorithm Optimized Machine Learning Algorithms and Sentinel-1 Data

Abstract

Riparian zones play a critical role in ecosystems. Accurately extracting the area of a riparian zone in open water is challenging due to human activities and climate change. This study used Sentinel-1 satellite data to investigate the capabilities of the support vector machine, extreme gradient boosting, and random forest methods, which were optimized by genetic algorithms for the detection of area changes in the riparian zone in the heart region of the Three Gorges Reservoir area of China. A total of 29 images were collected in 2020, and three models were created for each image, which were then transferred to other phases. The models’ performance metrics were validated using all of the images. The results indicated that the SVM method achieved the best performance with an accuracy of 0.945, an F1_Score of 0.950, and a kappa coefficient of 0.889. The optimal model was then used to monitor the area changes in the riparian zone over the study area in 2020. It was calculated that the area of the riparian zones was the smallest on 26 December and the largest on 17 June, with a maximum riparian zone of 31.97 km${}^2$. Overall, this study demonstrates that an SVM is the most stable method for detecting area changes in a riparian zone when using Sentinel-1 data compared to the RF and XGB methods. The findings are anticipated to provide a feasible plan for detecting the area dynamics in open-water riparian zones and to provide valuable information for the rational utilization of land resources and the ecological safety of the riparian zone in the Three Gorges Reservoir.

1. Introduction

The riparian zone is the area where the water level of reservoirs, lakes, or rivers changes as a result of periodic water storage and discharge [1]. As a transition zone between terrestrial and aquatic ecosystems, it is essential to accurately quantify the area dynamics of a riparian zone to better understand ecosystem processes such as soil sediment production and shift [2], plant community change [3], greenhouse gas flux [4], and fish habitat migration and change [5]. Specifically, monitoring the area of riparian zones contributes to assessing the health of ecosystems [6]. A long-term observation of changes in riparian zone areas helps to determine the extent of the damaged riparian zones, thus guiding ecological restoration projects and other initiatives [7,8]. Furthermore, changes in riparian zone areas provide valuable information for effective flood risk management and early warnings for natural disasters [9]. The accurate monitoring of riparian zone area changes aids local governments in formulating rational land use planning [10]. The traditional research on the area dynamic changes in land use, such as water, is usually based on expensive surface detection systems with artificial methods [11,12], which cannot cover most areas of the world. Remote sensing satellites have recently been an effective tool for studying large-scale land use [13,14]. Although they are limited by the temporal requirements for monitoring the area changes in riparian zones, some open-source optical sensor products, such as Landsat and Sentinel-2, have rich data content, high resolution, and are easy to obtain. However, these products cannot penetrate clouds. The impact of weather and day–night changes [15,16] make them unable to monitor the area changes in riparian zones for a long period of time, especially in cloudy areas. Conversely, active sensors, such as synthetic aperture radar (SAR), provide data that are unaffected by weather conditions and solar illumination [17] and that are sensitive to water [18,19], so they are often used to explore the area changes in open water [20]. Among them, Sentinel-1 is widely used due to its open-source data, wide coverage, and short revisit cycle [21,22].

SAR signals have a low backscatter coefficient in open water [22]. Therefore, the traditional use of the Sentinel-1 satellite for surface and water mapping and monitoring is conducted by setting the threshold value of the backscattering coefficient. For example, Souza et al. [23] used SAR backscatter thresholds to discuss the viability and constraints of Sentinel-1 data in monitoring the Poço da Cruz reservoir levels. They proved that Sentinel-1 data were effective in monitoring the changes in the reservoir’s water level. Similarly, Conded et al. [24] combined the SAR backscatter thresholds with Lee’s filtering [25] to obtain the most accurate map of anomalous floods in the Ebro River. Moghimi et al. [26,27] effectively monitored the changes in multitemporal SAR images by integrating the threshold and level-set methods. Despite the fact that such methods do not require supervised data and offer a high degree of automation, their results may be influenced by preprocessing errors and the inherent characteristics of these methods. Furthermore, as they are affected by human activities and climate change, changes in open water make it difficult to distinguish between water and land boundaries, which may result in errors in fixed threshold methods [28]. In addition, the SAR scattering coefficients of rivers and the surrounding land in different regions vary greatly, thus resulting in a poor generalization ability for such methods.

Machine learning (ML) algorithms have been popular in large-scale water surface and environmental monitoring in recent years due to their simplicity, low cost, and strong feasibility [29,30]. However, these algorithms have varied performances in distinguishing different land use problems. For example, Slagter et al. [31] used the random forest (RF) algorithm to capture the area dynamics of wetland and other land uses in St. Lucia, South Africa, with an overall accuracy of 87.1%. Acharya et al. [32] compared the capabilities of artificial neural networks, RF, support vector machines (SVMs), K-means clustering, and decision tree classifiers (DTCs) for surface and water monitoring based on Landsat 8 data. They found that DTCs and RF were more accurate compared to other algorithms. Luo et al. [33] put forward a lake extraction method using a combination of shape factors and ML, and they successfully extracted the lake and river areas of the Yellow River Delta in the last decade. They found that the SVM method performed the worst, while the backpropagation neural network (BP) and RF methods performed highly (and were close in performance to each other). Furthermore, numerous studies have demonstrated that optimizing ML hyperparameters through grid search [34], genetic algorithm (GA) [35], and random search [36] methods could improve model performance in most cases [37,38]. For example, Huang et al. [39] used a GA to optimize the DTC model, thus resulting in a 16.5% improvement in the accuracy for mapping the watershed of the Loess Plateau in Hongshimao. Sorkhabi et al. [40] demonstrated the capabilities of a GA, differential evolution (DE), and cat swarm optimization (CSO) in monitoring the dam reservoir storage in Iran through an unsupervised approach. In fact, GA searches for optimal solutions by mimicking the evolution of biological processes, and it has excellent global search capabilities that can rapidly seek all solutions in the solution space and not easily fall into local optima. Thus, it has been widely used in ML [41], signal processing [42], and combinatorial optimization [43].

However, few studies have paid attention to the change in riparian zone areas in the Three Gorges Reservoir region. To fill in the gap for monitoring the changes in the area of open-water riparian zones in the study area, this work used the easy access, short revisit cycle, and stable data acquisition characteristics of Sentinel-1 data to explore the ability of three ML algorithms for monitoring area changes in the riparian zone of the Three Gorges Reservoir (TGR) area in China. Unlike the common natural riparian zone, the riparian zone in this area is closely related to the Three Gorges Dam (TGD) and the flood season in the upper reaches of the Yangtze River. The research results are expected to provide a simple and low-cost method for monitoring changes in the area of open-water riparian zones and useful information for the rational use of land resources in the TGR area and ecological security in the surrounding areas. Specifically, preprocessed Sentinel-1 SAR images were used to evaluate three commonly used ML algorithms, namely RF, SVM, and XGB, optimized by genetic algorithms for monitoring the riparian zone in open water, and the optimal model was determined through test indicators and actual performance. Subsequently, the best model was used to determine the area changes of the riparian zone in the core area of the TGR in 2020.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area

As shown in Figure 1, Wanzhou District (30°23${}^{\prime }$50${}^{″}$–31°0${}^{\prime }$18${}^{″}$N, 107°52${}^{\prime }$22${}^{″}$–108°53${}^{\prime }$52${}^{″}$E) is located in the heart of the Three Gorges of Yangtze River, China. The Yangtze River runs from southwest to northeast through Wanzhou District, with a total transit length of 80.4 km, an average yearly flow of 13,200 m${}^3$/s, and an average annual transit flow of 416.3 billion m${}^3$. Water resources are abundant in Wanzhou District. There are 21 rivers including the Maojian River, the Wuqiao River, the Xintian River, the Shiqiao River, the Longbao River, and the Ramixi River that flows into the Yangtze River. The annual average temperature is 19.3 °C, and the yearly precipitation is 1531 mm. The flood season is usually from May to September, with 69.8% of the annual precipitation (1071 mm). The waters of the Three Gorges Reservoir (TGR) area are affected by the dynamic adjustment of the Three Gorges Dam and rainfall, with water surface varying between 145 and 175 m.

2.1.2. Data

The interferometric wide-swath (IW) ground-range detection (GRD) images covering the study area were obtained from the Copernicus Open Centre (https://scihub.copernicus.eu/dhus/#/home (accessed on 7 December 2021)). According to publicly available government data (www.gov.cn/xinwen/2021-12/03/content_5655676.htm (accessed on 13 January 2022)), in 2020, the Yangtze River witnessed the most severe basinwide flood since 1998, with the largest inflow peak flow rate into the Three Gorges Reservoir since its construction reaching 75,000 cubic meters per second. Based on this information, a total of 29 images from 2020 were used in this study (

Table 1. The images used in the current study.

Date	Code *	Date	Code *	Date	Code *	Date	Code *	Date	Code *
1 January	0101	13 March	0313	24 May	0524	4 August	0804	27 October	1027
13 January	0113	25 March	0325	5 June	0605	16 August	0816	8 November	1108
25 January	0125	6 April	0406	17 June	0617	28 August	0828	20 November	1120
6 February	0206	18 April	0418	29 June	0629	21 September	0921	2 December	1202
18 February	0218	30 April	0430	11 July	0711	3 October	1003	26 December	1226
1 March	0301	12 May	0512	23 July	0723	15 October	1015

Note: * the images were encoded by month and day in numbers.

). The acquired Sentinel-1 data were preprocessed by Sentinels Application Platform software (SNAP 8.0) and sequentially processed by Apply Orbit File, Thermal Noise Removal, Calibration, Linear To From dB, Multilook, Speckle Filtering, and Terrain-Correction. A digital elevation model (DEM) with a 30 m resolution (https://www.webmap.cn (accessed on 17 January 2022)) was used for terrain correction. All imageries were used to analyze the area changes of the riparian zone of the study area. For each image, the VV polarization, VH polarization, projected incidence angle, and ALPHA band were obtained and used for machine learning modeling.

For model calibration and validation, sample sites were selected with the help of Google Earth Pro (7.3.4.8248) and the land use map (1:10,000). Although optical satellites cannot depict the changes in the riparian zone over time, their higher resolution and clarity can help us with sampling. First, 5177 water samples and 538,724 nonwater samples were obtained over the entire region using the land use map. In order to ensure that the sample points were consistent in all images, 29 Sentinel-1 data images and Google Earth Pro were used to screen them. Considering the resolution (20 m) of the data from Sentinel-1, water surfaces that could not be identified at this resolution were removed. Water samples that were not always in water bodies in all images were deleted. In this way, 4057 water samples were obtained. To satisfy the requirement of data balance, a downsampling technique was applied to remove the majority class samples. This process eliminated most nonwater examples. Finally, a total of 8499 samples were obtained, comprising 4057 water samples and 4442 nonwater samples. About 60% of these were used for model calibration and 40% for model validation (Figure 2).

In addition, three sites along the Yangtze River were picked to test the properties of the models and visualize the changes in the riparian zone (Figure 3). At site A, there is a bridge over the river and its boundary with the water area is not clear. Site B contains only water and riverbanks and is one of the common scenarios in the study area. However, the shaded features of the riverbanks are somewhat similar to those of the waters. Sites A and B were used to assess the model’s resistance to disturbance and robustness. Additionally, site C was located in the southwest part with a flat terrain and used to demonstrate riparian zone changes in the study area.

As rainfall somewhat influences the variation in the Three Gorges reservoir riparian zone, Figure 4 shows the monthly rainfall in 2020 for Chongqing, where the study area was located. In 2020, rainfall in Chongqing was mainly concentrated between May and September, with heavy rainfall occurring in June and July.

2.2. Methodology

2.2.1. Support Vector Machine

Cortes and Vapnik first proposed support vector machines (SVMs) in the 1990s [44], which are broad linear classifiers that classify data by supervised learning. The decision boundary is determined by learning the largest margin hyperplane of the sample solution. In addition, the classification decision of SVMs is determined by the support vector independent of the sample dimensionality, avoiding to some extent the dimensionality disaster [44]. This makes support vector machines very suitable for remote sensing image analysis with a large dimensionality [45].

2.2.2. Extreme Gradient Boosting

Extreme gradient boosting [46] (XGB) is an improvement on the gradient boosting decision tree (GBDT) algorithm. It was proposed by Chen et al. [46] and uses the Taylor expansion to replace the objective function in GBDT. New tree models are generated by continuously learning from previously generated tree models. In addition, XGB also adds a regularization term to the objective function to adjust the complexity of the model during the learning process, resulting in stronger generalization capability. The characteristics of XGB make it widely used in fields such as vegetation mapping applications [47] based on remote sensing data.

2.2.3. Random Forest

Random forest (RF) is an ensemble learning method originally developed by Breiman [48]. It uses multiple decision tree models to solve the same problem and calculates the vote of each decision tree to obtain the final prediction. This process can counteract overfitting and outlier instances that can occur with single decision trees, making the RF insensitive to outliers and noise. In addition, RF also uses bagging to automatically split data into training and test sets, improving the stability and accuracy of the model. Random forest is widely used as a user-friendly tool in fields such as remote sensing and data mining [49].

2.2.4. Hyperparameters Optimization-Seeking Algorithms

Hyperparameters of machine learning models that are not learned from the data are often predesigned or set to default values by humans, such as the depth of the tree in a decision tree (DT) model or the number of DTs in a random forest [50]. The performance of the model is largely affected by these hyperparameters. For greater performance, the model is always optimized hyperparametrically by means of a grid search (GS) [51] or a random optimization search [34]. A GS is essentially an exhaustive search algorithm for specifying parameter values, which can lead to a high time complexity when the quantity of parameters or the search range is large [51]. Currently, there are many studies indicating that stochastic optimization search algorithms such as particle swarm optimization (PSO) [52] and genetic algorithms (GAs) [53] perform well in finding the optimal hyperparameters of a model [54,55]. These types of algorithms obtain a better combination of hyperparameters by simulating biological migration or genetics in nature [56]. Genetic algorithms are very active in applications such as remote sensing image classification and feature detection [54,55]. For instance, Wang et al. used a GA to optimize the combination of SVM hyperparameters, which significantly enhanced the overall classification accuracy of fine Lccom [57].

The specific process of a genetic algorithm is as follows:

• Construct an initial population (where each individual in the population represents a solution).
• Calculate the fitness of individuals in the population using an evaluation function.
• Use fitness values to calculate the probability of selecting individuals from the population.
• Generate the next generation of the population by combining different individuals through crossover.
• Introduce a certain mutation probability to randomly change some elements of individuals’ sequences to prevent becoming trapped in local optima.
• Iterate until termination conditions are met (e.g., reaching a maximum number of generations or satisfying convergence criteria).

In order to improve the model performance, the genetic algorithm was applied to optimize two critical hyperparameters for each model (

Table 2. Hyperparameters and search ranges of different machine learning models.

Classifier	Hyperparameter	Candidate Value	Combination Number	Optimization Result
SVM	c	0.01–1000	2.56 × ${10}^{11}$	15.886
	gamma	0.0001–256		0.0039
RF	max_depth	1–600	6 × ${10}^5$	256
	n_estimators	1–1000		997
XGB	learning_rate	0.01–1.0	${10}^5$	0.104
	sub_sample	0.01–1.0		0.772

). For SVM, the kernel function was the widely used radial basis function (RBF) in this work. Real integer coding was employed to establish the individuals of the genetic algorithm [35]. The count of individuals in each population was set to 50, the populations size was 100, and the upper limit of reproduction was 100 generations. The parameters of the optimized model after optimization through the genetic algorithm can be seen in Table 2. In fact, compared to before using GA optimization, the model accuracy improved to a certain extent (1–2%). See Appendix A for more optimization details.

2.2.5. Statistical Indicators

The confusion matrix is a common tool for evaluating machine learning classification accuracy. For classification problems with n class elements, the predictions of the model are recorded by an n × n table.

Table 3. Confusion matrix.

		Prediction
		Positive (Pos)	Negative (Neg)
Ground truth	Positive	TP (True Pos)	FN (False Neg)
	Negative	FP (False Pos)	TN (True Neg)

displays the confusion matrix for a binary classification problem. In this work, TP and TN represent correctly classified waters and nonwaters, respectively, while FP and FN indicate incorrectly classified waters and nonwaters. According to the confusion matrix, the accuracy, kappa, and F1_Score metrics can be computed [58].

Accuracy (ACC) is the simplest and most intuitive metric for classification problems, reflecting the proportion of samples correctly forecasted by the model to the total number of samples. The greater the accuracy, the higher the model performance.

$$ACC=\frac{TN+TP}{TP+FN+FP+TN},$$

(1)

The kappa coefficient is utilized to test the consistency of the model and is within the range [−1, 1], but it usually falls between 0 and 1. In terms of kappa coefficient, the model gives low [0.0–0.2], fair [0.21–0.40], moderate [0.41–0.60], high [0.61–0.80], and perfect [0.81–1.0] consistency. The kappa coefficient is calculated as follows:

$$Kappa=\frac{ACC-Pe}{1-Pe},$$

(2)

where

$$Pe=\frac{TP\times \left(FN+TP\right)+\left(TN+FP\right)\times TN}{{\left(TP+FN+FP+TN\right)}^2},$$

(3)

Recall (R) indicates how numerous positive cases are properly forecasted in the sample.

$$R=\frac{TP}{FN+TP},$$

(4)

Precision (P) indicates the number of samples predicted to be positive that are actually positive samples.

$$P=\frac{TP}{FP+TP},$$

(5)

R and P indicators contradict each other in some cases, so F1_Score was introduced by taking a weighted average of them. The higher the F1_Score, the more effective the model.

$$F{1}_{Score}=\frac{2\times R\times P}{R\times P},$$

(6)

2.3. Overall Process

The overall flow of this study is shown in Figure 5.

3. Results

3.1. Optimal Phase

To obtain a robust model with superior classification accuracy during the study period, the developed models using one image were also transferred to other phases. Model performance was evaluated by the validation set of each image.

Table 4. Accuracy (ACC), F1_score, and kappa coefficient of the RF, XGB, and SVM methods for different time-phase images. Higher indicators are marked in bold.

Image	ACC			F1_score			Kappa
	RF	XGB	SVM	RF	XGB	SVM	RF	XGB	SVM
0101	0.9406	0.9411	0.9395	0.9465	0.9457	0.9456	0.8803	0.8812	0.8779
0113	0.9416	0.9403	0.9400	0.9471	0.9458	0.9460	0.8824	0.8796	0.8790
0125	0.9400	0.9400	0.9386	0.9455	0.9454	0.9447	0.8791	0.8792	0.8762
0206	0.9410	0.9370	0.9451	0.9449	0.9399	0.9501	0.8815	0.8736	0.8894
0218	0.9433	0.9432	0.9402	0.9476	0.9446	0.9460	0.8859	0.8840	0.8794
0301	0.9435	0.9431	0.9437	0.9465	0.9463	0.9490	0.8845	0.8849	0.8866
0313	0.9410	0.9344	0.9424	0.9445	0.9365	0.9479	0.8815	0.8686	0.8839
0325	0.9438	0.9424	0.9434	0.9479	0.9456	0.9489	0.8851	0.8844	0.8860
0406	0.9424	0.9423	0.9436	0.9452	0.9454	0.9490	0.8831	0.8841	0.8862
0418	0.9406	0.9383	0.9445	0.9443	0.9412	0.9491	0.8807	0.8762	0.8882
0430	0.9433	0.9388	0.9451	0.9473	0.9419	0.9503	0.8860	0.8772	0.8894
0512	0.9430	0.9431	0.9412	0.9462	0.9450	0.9470	0.8853	0.8850	0.8814
0524	0.9381	0.9320	0.9415	0.9422	0.9355	0.9462	0.8756	0.8636	0.8823
0605	0.9250	0.9240	0.9327	0.9309	0.9289	0.9392	0.8492	0.8474	0.8642
0617	0.9234	0.9172	0.9338	0.9260	0.9189	0.9376	0.8465	0.8343	0.8671
0629	0.9422	0.9383	0.9427	0.9465	0.9426	0.9477	0.8837	0.8761	0.8847
0711	0.9372	0.9345	0.9414	0.9410	0.9382	0.9460	0.8738	0.8686	0.8822
0723	0.9368	0.9339	0.9406	0.9414	0.9384	0.9457	0.8729	0.8672	0.8803
0804	0.9431	0.9407	0.9421	0.9479	0.9454	0.9475	0.8855	0.8807	0.8834
0816	0.9433	0.9391	0.9445	0.9472	0.9420	0.9494	0.8861	0.8779	0.8882
0828	0.9400	0.9361	0.9438	0.9435	0.9383	0.9488	0.8795	0.8720	0.8869
0921	0.9435	0.9437	0.9454	0.9475	0.9459	0.9504	0.8864	0.8850	0.8899
1003	0.9370	0.9329	0.9399	0.9410	0.9363	0.9444	0.8735	0.8654	0.8791
1015	0.9281	0.9191	0.9411	0.9308	0.9201	0.9454	0.8560	0.8382	0.8816
1027	0.9413	0.9411	0.9439	0.9454	0.9441	0.9492	0.8821	0.8819	0.8869
1108	0.9402	0.9382	0.9405	0.9452	0.9411	0.9464	0.8824	0.8762	0.8800
1120	0.9377	0.9386	0.9387	0.9435	0.9413	0.9445	0.8789	0.8770	0.8758
1202	0.9415	0.9422	0.9432	0.9465	0.9437	0.9486	0.8851	0.8806	0.8856
1226	0.9380	0.9395	0.9403	0.9464	0.9434	0.9464	0.8841	0.8803	0.8796

presented the averaged ACC, F1_score, and kappa coefficient of the SVM, RF, and XGB for each image. The performance of the models generally varied from image to image, and even different models produced from the same image varied, but in general, the accuracy of the models produced by the SVM was better than that of the RF and XGB. Clearly, the models built with the image acquired on 21 September (0921) gave the highest values of ACC, F1_score, and kappa for SVM, XGB, and RF. According to the kappa coefficient, the three models gave perfect consistency (0.81–1.0). Compared with the RF and XGB, the SVM performed the best with an average ACC of 0.945, F1_Score of 0.950, and kappa coefficient of 0.889.

3.2. Identification of Surface with Optimal Models

To illustrate the details of the classification accuracy, the three optimal models, namely, the RF, XGB, and SVM developed from image 0921, were applied to extract the surface at site A and site B. The identification of the surface in winter, spring, summer, and autumn at the two sites is shown in Figure 6 and Figure 7. At site A, there was a bridge over the river, and the models built by the RF and XGB models did not recognize the water surface affected by the interfering bridge well; a part of the bridge was also classified as water surface, while the SVM performed the best and the predicted results were very similar to the original map, distinguishing more clearly the boundary between the water surface and the bridge, showing some robustness in terms of resistance to interference.

At site B, the SVM was barely affected by the shadows of the riverbank and clearly separated the water surface from the riverbank shadows, while both the RF and XGB models classified some riverbank shadows as water surfaces. In the right part of site B, there is a narrow, long valley. The SVM could identify it from the water, while the RF and XGB models could not distinguish well between the shadow of the valley and the water.

3.3. Surface Change in the Riparian Zone

According to the public information of the government, the release dates of the Three Gorges Dam (TGD) are commonly from April to June, mainly related to the flooding of the Yangtze River basin in that year and the fact that the experimental storage starts from September to November. Therefore, six Sentinel-1 images of 0406, 0524, 0617, 0723, 1027, and 1226 time phases at site C were selected to explore the riparian zone changes in Wanzhou (Three Gorges Reservoir area) according to the optimum model. The predicted result and the elevations of 145 and 175 m are presented in Figure 8, and in order to better display the riparian zone changes, the difference between the six images and the maximum riparian area of the place is also marked.

At site C, Figure 8a shows the level of water before the release of water from the Three Gorges Reservoir (TGR) in April, which was around 175 m. Although the water was high at this time, the riparian zone was not completely submerged by water. As rainfall increased from April in Chongqing (the upstream of the reservoir), it was likely that the Three Gorges Reservoir would be gradually released from that moment to cope with the Yangtze flood. After a period of water discharge from the reservoir, the level of water gradually decreased, with the level of water at the end of May being higher than 145 m (Figure 8b). At that time, the area of the riparian zone gradually expanded. By mid-June, the water level had dropped to a minimum and was basically at the same level as the 145 m. The area of the riparian zone reached its maximum. Due to the surge in rainfall upstream of the TGD in June and July, the water level rose again at the end of July, presumably to regulate the downstream flood pressure and to avoid or mitigate the possibility of natural disasters such as flooding downstream. The TGD reduced its discharge compared to the release period, causing the level of water to rise again in the middle of the flood (Figure 8d) and the exposed riparian zone was mostly submerged by water again. When the reservoir flood period ended and the TGD restarted storage, the water level was maintained close to but below 175 m due to the rainfall in June and July and the regulation of the TGD, so the storing of water at the TGD only caused the level of water to rise to 175 m (Figure 8e). The TGD then maintained storage to meet tasks such as power generation in the reservoir and water regulation downstream. Site C reached its highest level in December 29 of the 29 images recorded by Sentinel-1 in 2020 (Figure 8f). Similar to what is represented in Figure 8e, at that time the exposed riparian zone almost disappeared.

Figure 9 shows that the change in the riparian zone mainly occurred from April to October in 2020, with the level of water remaining high in the other months. However, this does not mean that the riparian zone disappeared during that period, even though its area was very small. In order to cope with the rainy season, the water level in the reservoir began to drop around April and reached its lowest on 17 June, causing the riparian zone to be completely exposed. With the arrival of the rainy season in the reservoir and the regulation of the TGD, the riparian zone was gradually covered by water flow, and its area gradually shrank. Through calculation, the maximum area of the riparian zone was found to be 31.97 km${}^2$. This is consistent with the reality. See Appendix B for the area changes of the other two stations.

4. Discussion

4.1. The Model Performance

For land type classification [31,32], the sample points of land types are generally acquired through field exploration or determined by the human eye through clean, high-resolution satellite imagery. However, manual methods and dynamic changes in the land-use type may cause noise in the dataset required by the experiment, which leads to differences in model performance [31]. In contrast, the final prediction process of the SVM is controlled by only a small number of support vectors [44], and the small number of support vectors allows the SVM to capture key samples while eliminating redundant or noisy data. This may be the reason why SVM performed well in handling the noise of Sentinel-1 data (site A) and distinguishing complex land classes (site B). At the same time, this also means that when detecting the area changes in the riparian zone, the SVM is a more robust method than the RF and XGB. When considering more complex optical satellite data, the SVM’s characteristics in low-dimensional learning and high-dimensional classification [59] become even more advantageous. Furthermore, compared to threshold-based methods [23,24,26], which do not require training data, the acquisition of thresholds depends on complex procedures. Their results can be easily affected by preprocessing, specific-method hyperparameters, and other factors, which increases the difficulty in monitoring changes in the riparian zone area. In contrast, the simple modeling process of machine learning holds an advantage in terms of reproducibility and generalizability.

Table 4 indicates that the same algorithms exhibited various performances with different temporal images. The model accuracy was relatively low in June when the water level was low, which may be due to the fact that the exposed zone increased the complexity of land types in that period, and the coarser resolution of Sentinel-1 could not capture this information, especially in the areas with complicated terrain. It is necessary to consider using higher-resolution satellite data or more advanced methods, such as deep-learning-based image segmentation [60].

4.2. The Area Change in the Riparian Zone

The water level in the upstream riparian zone of the TGR area was high from October to April of the following year, and the exposed riparian zone area was relatively stable. However, the riparian zone area changed significantly from May to September. Although there was a large amount of rainfall during that period, the flood discharge mechanism of the TGD made the riparian zone reach its maximum in mid-June. This is consistent with the reality and also means that the TGR riparian zone changes more frequently during that period. In fact, frequent changes are more likely to affect the riparian zone, so during that period, it is more necessary to monitor the riparian zone to reduce the impact of possible pollutants or other factors on the riparian zone.

According to public information, the TGD generally regulates the water level to 145 m by water release before June to cope with the Yangtze River flood season. This exposes a large area of the riparian zone. In the flood season, the heavier rainfall gradually takes up the capacity set aside by the TGD, causing the water level to rise. When the flood season is over, the TGD operates at a peak water level of 175 m to ensure its own performance, its power generation plan, and downstream water demand. This is also the reason why the area of the riparian zone remains relatively stable during that period. Even during the full-basin flood in 2020, the water level in the reservoir area remained relatively stable. Therefore, compared to natural basins, the changes in the TGR riparian zone are closely related to the storage and discharge cycle of the TGD.

5. Conclusions

This study compared the capabilities of different machine learning algorithms in monitoring area changes in riparian zone areas. For the first time, it determined the trends and extent of riparian zone area changes in the core area of the Three Gorges Reservoir. These results provide an accurate dataset for designing suitable land use and ecological security plans in that region.

Some factors such as topography, which was not considered in the current work, may influence the model’s performance. In the areas with complicated terrain, the limited spatial resolution of the satellite may also result in poor model accuracy.

In the future, it is recommended to explore the use of satellite data with higher spatial and temporal resolutions, as well as advanced image segmentation algorithms. These efforts will contribute to obtaining more precise dynamics of riparian zone areas at the target locations.