Remote sensing image landslide detection method and system

Abstract

The application discloses a remote sensing image landslide detection method and system, and belongs to the technical field of landslide detection. The method adopts a data enhancement method to expand a landslide data set, and solves the problem of a small number of landslide samples in a training set. DarkNet53 is used to replace a commonly used backbone feature extraction network of FasterR-CNN, so that the problem of gradient disappearance caused by a too deep network is avoided, and the detection precision is improved.

Description

Technical Field

[0001] This invention relates to the field of landslide detection, and in particular to a method and system for landslide detection using remote sensing images. Background Technology

[0002] Landslides are a common natural disaster, damaging not only the natural environment but also villages, town buildings, and infrastructure, and in severe cases, causing numerous casualties. Landslide identification, early warning, and prevention are crucial for reducing and avoiding geological disaster losses, and are of great significance for landslide prevention and the establishment of landslide databases.

[0003] Traditional landslide identification relies on geological workers who create geological hazard catalogs through on-site exploration and visual interpretation of remote sensing images. However, due to limitations of traditional technologies and conditions, the completeness and timeliness of existing landslide catalogs are insufficient to meet the needs of current economic development and disaster prevention and mitigation. Landslide identification methods utilizing the spectral and textural features of landslide remote sensing images are relatively mature and can be mainly divided into manual visual interpretation methods, change detection methods, machine learning methods, and deep learning methods.

[0004] Visual interpretation is a commonly used method in geological hazard research. Researchers visually interpret remote sensing images and combine this with non-remote sensing data on geological hazards for analysis and reasoning to identify potential hazard areas. For example, Liu Chunling et al. successfully interpreted major engineering disturbance landslides in a transboundary region of Southeast Asia using Gaofen-1 satellite remote sensing data. Although this method demonstrates high accuracy in landslide identification, it heavily relies on the researchers' prior knowledge and experience.

[0005] Change detection is another commonly used landslide identification method. It uses optical images of the same area at multiple time points, combined with pixel-based multi-threshold analysis and object-oriented analysis methods. This method avoids the "salt and pepper" effect caused by noise in pixel analysis methods, and makes full use of various image information, thereby improving the accuracy and efficiency of landslide identification. However, this method requires optical image data from multiple time phases. For example, Zhang Shuaijuan et al. successfully extracted landslide information in the Jiama mining area of ​​Tibet Autonomous Region using change detection.

[0006] Traditional landslide identification mainly relies on visual interpretation by experts. With the development of computers, methods for detecting changes in multiple thresholds designed manually have gradually emerged. In recent years, with the development of artificial intelligence technology, landslide identification methods based on machine learning and deep learning have appeared.

[0007] Machine learning methods utilize various statistical models to extract relevant features between landslide bodies and optical images, and then use classifiers to identify the relationships between these features for landslide identification. While this method is time-consuming and requires less manpower, feature selection and parameter tuning are cumbersome. Currently, various machine learning methods, such as support vector machines and random forests, have been combined with pixel-based and object-oriented analysis methods to become new approaches in landslide identification. For example, Xi Wenfei et al. used an improved Retinex algorithm to remove image gross errors in the mountainous areas of northeastern Yunnan, combined with an SVM model to identify landslides.

[0008] Deep learning-based landslide recognition methods primarily utilize convolutional neural networks (CNNs) to construct deep learning networks that learn high-order semantic features of images to identify landslides. Currently, there are two main approaches to landslide recognition using deep learning: one is based on object detection algorithms, and the other is based on semantic segmentation algorithms. Object detection algorithms identify objects of the required type in an image, determining their category and location. Current object detection algorithms are mainly divided into single-stage and two-stage algorithms. Among single-stage algorithms, the YOLO series is currently the most widely used. For example, Heyi and Hou have used an improved YOLOX object detection model to detect complex landslides, improving the detection accuracy for complex, small landslides. In two-stage algorithms, Faster R-CNN (Faster Region Convolutional Neural Networks) is widely used. For instance, Niu and C have improved the Faster R-CNN model by incorporating an attention module, improving the detection accuracy for non-landslide and debris flow types. Single-stage detection algorithms outperform two-stage algorithms in terms of detection speed, but their accuracy is less precise.

[0009] Faster R-CNN, as the most widely used two-stage object detection algorithm, demonstrates better detection accuracy compared to single-stage algorithms. Compared to its predecessor, Fast R-CNN, the proposed RPN can quickly provide a large number of region candidate boxes, improving detection speed while maintaining accuracy. However, the best-performing feature extraction networks in Faster R-CNN, VGG16 (Visual Geometry Group) and ResNet50 (Residual Networks), are prone to gradient vanishing during backpropagation due to their deep network structures. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for detecting landslides in remote sensing images, which can avoid the gradient vanishing problem caused by excessively deep networks and improve detection accuracy.

[0011] To achieve the above objectives, the present invention provides the following solution:

[0012] A method for detecting landslides in remote sensing images, comprising:

[0013] The improved Faster R-CNN model is obtained by replacing the backbone feature extraction network in the Faster R-CNN model with DarkNet53.

[0014] Obtain a landslide dataset containing multiple landslide remote sensing image samples;

[0015] The landslide dataset was augmented using data augmentation techniques to obtain the augmented landslide dataset.

[0016] The improved Faster R-CNN model was trained using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model;

[0017] The remote sensing image of the landslide to be detected is input into the trained improved Faster R-CNN model to identify the landslide in the image.

[0018] A remote sensing image landslide detection system includes:

[0019] The network improvement module is used to replace the backbone feature extraction network in the Faster R-CNN model with DarkNet53 to obtain an improved Faster R-CNN model.

[0020] The dataset acquisition module is used to acquire a landslide dataset containing multiple landslide remote sensing image samples;

[0021] The augmentation module is used to augment the landslide dataset using data augmentation methods to obtain an enhanced landslide dataset.

[0022] The training module is used to train the improved Faster R-CNN model using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model.

[0023] The identification module is used to input the remote sensing image of the landslide to be detected into the trained improved Faster R-CNN model to identify the landslide in the image.

[0024] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0025] This invention discloses a method and system for detecting landslides in remote sensing images. It uses data augmentation to expand the landslide dataset, solving the problem of insufficient landslide samples in the training set. It replaces the commonly used backbone feature extraction network of Faster R-CNN with DarkNet53, avoiding the gradient vanishing problem caused by excessive network depth and improving detection accuracy. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart of a landslide detection method for remote sensing images provided in an embodiment of the present invention;

[0028] Figure 2 A network structure diagram of the improved Faster R-CNN model provided in this embodiment of the invention;

[0029] Figure 3 This is a partial sample of the landslide dataset provided in the embodiments of the present invention;

[0030] Figure 4 This is a diagram showing the results of data augmentation processing of the landslide dataset provided in an embodiment of the present invention.

[0031] Figure 5 Sample images of the batch detection effect of the method of the present invention provided in the embodiments of the present invention;

[0032] Figure 6 This is a sample image of the batch detection effect of ResNet50 provided in an embodiment of the present invention;

[0033] Figure 7 The image shows a sample of the batch detection effect of VGG16 provided in an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] like Figure 1 As shown, this embodiment of the invention provides a method for detecting landslides in remote sensing images, including:

[0037] Step 1: Replace the backbone feature extraction network in the Faster R-CNN model with DarkNet53 to obtain the improved Faster R-CNN model.

[0038] Faster R-CNN commonly uses VGG16 and ResNet50 for feature extraction, employing multiple 3×3 convolutional kernels to extract features at greater depths. However, as the network depth increases, accuracy decreases instead of increasing. This is because excessive network depth leads to vanishing gradients during backpropagation, making it impossible to adjust the weights of earlier layers.

[0039] In contrast, YOLOv3's feature extraction network, DarkNet53, avoids the vanishing gradient problem caused by excessive network depth due to its skip connections in the residual structure, and boasts advantages such as small parameter count and high accuracy. However, in YOLOv3, the feature layers extracted by a single DarkNet53 do not undergo RPN (Region Proposal Network) generation to generate candidate regions, resulting in YOLOv3's recognition accuracy being lower than Faster R-CNN. Therefore, this invention attempts to combine the advantages of DarkNet53 and Faster R-CNN, replacing the VGG16 and ResNet50 feature extraction networks in Faster R-CNN with DarkNet53, and using RPN to generate candidate regions for the extracted feature layers, thereby improving the accuracy of the object detection model.

[0040] During the experiment, because the residual edges of ResiduaBlock directly map the input layer features of the Feature Map to the output layer after the skip connections, it causes uneven distribution of the output layer, affecting the network's fitting ability. Therefore, this invention adds a BN layer and a Leaky ReLU layer between DarkNet53 and RPN to ensure that the output feature layers have the same distribution, avoid overfitting, and increase the network's nonlinearity.

[0041] Simultaneously, data augmentation techniques were combined to process the landslide dataset, and the improved algorithm was applied to the landslide dataset. Comparative experiments show that the algorithm performs well on the landslide dataset.

[0042] Figure 2In this code, CBR represents Conv layer, BN layer, and LR layer; LR layer represents Leaky ReLU activation function layer; Bottleneck represents residual network block; ROI Pool represents ROI pooling layer; Classifier represents classifier; bbox_pred represents regression prediction; and cls_pred represents classification prediction.

[0043] Step 2: Obtain a landslide dataset containing multiple landslide remote sensing image samples.

[0044] To train and test the improved Faster R-CNN landslide detection algorithm proposed in this invention, this invention selects the latest publicly available dataset from Wuhan University, China, the Bijie Landslide Dataset, which contains 770 labeled landslide sample locations.

[0045] Figure 3 This dataset showcases selected landslide examples from this publicly available dataset. The dataset comprises 770 landslide samples, sourced from TripleSat satellite images taken between May and August 2018, with a ground resolution of 0.8m. Figure 3 Parts (a), (b), (c), (d), (e), (f), (g), and (h) in the figure show partial samples of the landslide dataset.

[0046] Step 3: Use data augmentation methods to expand the landslide dataset to obtain the augmented landslide dataset.

[0047] In the field of object detection, the number of training samples is a crucial factor determining the performance of deep learning models. However, the current number of training samples is often insufficient for training deep learning models, leading to overfitting issues during training. Therefore, this invention addresses this problem through offline data augmentation techniques. These techniques not only expand the number of training samples but also ensure the diversity of sample features, effectively solving the overfitting problem with small datasets and improving the model's training performance.

[0048] Because the Bijie landslide dataset from Wuhan University has a limited sample size, this invention uses the ImgAug image enhancement library to augment the original dataset. The image enhancement methods selected in this invention include eight types: HSV (color dithering), HSV + luminance, HSV + luminance + Gaussian noise, Gaussian noise, luminance, flip, scaling + translation, and rotation.

[0049] This invention uses the Bijie landslide dataset from Wuhan University. The original dataset contained 770 landslide images, while the augmented dataset contains 6930 landslide images. The augmented result is shown below. Figure 4 As shown. Figure 4Part (a) in the image is the original image. Figure 4 Part (b) in the image is the flipped version. Figure 4 Part (c) in the image is the scaled and translated image. Figure 4 Part (d) in the image is the image after adding Gaussian noise. Figure 4 Part (e) in the image shows the image after the brightness has been changed. Figure 4 Part (f) in the image is the rotated image. Figure 4 The (g) part in the image is the image after color dithering. Figure 4 The (h) part in the image represents the image after color dithering and brightness adjustment. Figure 4 Part (i) in the image is the result of color jittering, brightness adjustment, and Gaussian noise addition.

[0050] Step 4: Train the improved Faster R-CNN model using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model.

[0051] The deep learning models of this invention all use the PyTorch 1.10.0 deep learning framework. The computer processor used is an Intel Xeon Silver 4110, the operating system environment is Windows 10, the CPU memory is 64GB, the GPU model is NVIDIA GeForce GTX 1080Ti, and the video memory is 11GB. During training, CUDA 11.3 and CUDNN 8.2.0 graphics card drivers are used for GPU-accelerated computation. The initial learning rate of the algorithm model of this invention is set to 0.0001, and the number of training epochs is set to 120 epochs.

[0052] To avoid memory performance issues, this invention divides the training process into two phases based on different batch sizes: a trunk freezing phase and a trunk unfreezing phase. The first 50 rounds are the freezing phase with a batch size of 8; the last 70 rounds are the unfreezing phase with a batch size of 4.

[0053] To ensure the accuracy of the experimental results and the reproducibility of the model, this experiment uses a fixed random number seed to generate a consistent set of random numbers. This seed is used in each model training iteration, ensuring that the initial weights are identical and thus maintaining stable training results. Furthermore, the experimental results can be reproduced. It is important to note that the predictive performance of the neural network model is primarily related to the network structure and the number of iterations, and is independent of the random number seed setting.

[0054] Step 5: Input the remote sensing image of the landslide to be detected into the trained improved Faster R-CNN model to identify the landslide in the image.

[0055] In the implementation of the algorithm, this invention selects a deep learning framework based on PyTorch and replaces the commonly used VGG16 and ResNet50 networks in Faster R-CNN with DarkNet53. Before RPN, BN layer and Leaky ReLU layer are added to make the extracted feature layers satisfy the same distribution, increase the nonlinearity of the network, and thus improve the detection accuracy of the model.

[0056] Furthermore, addressing the issue of insufficient landslide samples in the landslide dataset, this invention employs offline data augmentation to expand the landslide dataset, enabling more thorough model training and further improving detection performance. Experiments demonstrate that the algorithm proposed in this invention, compared to the original algorithm, improves detection accuracy while reducing the number of model parameters, enabling more accurate landslide detection. This has significant practical value in the field of landslide detection and provides strong technical support for further enhancing landslide identification and early warning prevention.

[0057] The detection results of the method of the present invention will be evaluated through experimental means below.

[0058] (1) Model evaluation indicators

[0059] To verify the performance of the algorithm of this invention, the Wuhan University landslide dataset was augmented to 6930 samples using data augmentation methods. The proposed algorithm was then compared with Faster R-CNN (VGG16) and Faster R-CNN (ResNet50) algorithms. This was done to determine the optimal training effect of the network model. 30 The dataset is divided into two parts: a test set (10%) and a validation set (90%). The training and validation sets are split in an 8:2 ratio, and different network models are trained on each set separately. This invention uses four metrics—AP, Recall, Precision, and F1 Score—to evaluate the performance of the algorithm.

[0060] This invention trained three algorithms—Faster R-CNN (VGG16), Faster R-CNN (ResNet50), and the proposed algorithm—for 120 rounds on the training set. The optimal training weights for each algorithm were selected, and the model accuracy was evaluated on the test set. The evaluation results are shown in Table 1. Table 1 shows that the proposed algorithm achieved AP, Precision, Recall, and F1 Score of 92.05%, 69.59%, 93.80%, and 0.80, respectively. Compared to Faster R-CNN (VGG16), AP improved by 10.71%, Precision by 10.44%, Recall by 7.51%, and F1 Score by 0.10. Compared to Faster R-CNN (ResNet50), AP improved by 0.42%, Precision by 1.52%, Recall decreased by 1.87%, and the F1 Score remained the same.

[0061] The algorithm proposed in this invention outperforms Faster R-CNN (VGG16) in all evaluation metrics, and outperforms Faster R-CNN (ResNet50) in AP and Precision, achieving parity in F1 Score, but slightly lower in Recall. This indicates that the use of DarkNet53 in this invention enhances feature propagation and improves the gradient vanishing problem, thereby improving the model's AP and Precision. The addition of BN and Leaky ReLU layers before RPN ensures that the feature layers satisfy the same distribution, enhancing network nonlinearity and maintaining model performance, resulting in a good F1 Score. Tables 2 and 3 show that, compared to ResNet50, the algorithm proposed in this invention has fewer parameters and a shorter detection time, leading to a slight decrease in the number of sample detections when Precision is high, resulting in a slightly lower Recall than ResNet50.

[0062] Table 1 Comparison of model evaluation metrics for the three algorithms

[0063]

[0064] Note: Bold numbers indicate the optimal value for each evaluation indicator.

[0065] This invention compares the number of parameters of the three algorithms, and the comparison results are shown in Table 2. VGG16 has the fewest parameters, while ResNet50 has the most. The algorithm proposed in this invention increases the number of parameters by 27.581 Mb compared to VGG16, and reduces it by 80.833 Mb compared to ResNet50.

[0066] Table 2 Comparison of total model parameters for the three algorithms

[0067] algorithm Parameter quantity (Mb) FasterR-CNN (VGG16) 28.275 Faster R-CNN (ResNet50) 136.689 Algorithm of this invention 55.856

[0068] Note: Bold numbers indicate the optimal values ​​for the parameters.

[0069] Although VGG16 has the smallest number of parameters, as shown in Table 1, it is the worst in all evaluation metrics. Although ResNet50 is better than VGG16 in all evaluation metrics, as shown in Table 2, its large number of parameters leads to a decrease in model detection speed. Although the algorithm proposed in this invention has fewer parameters than VGG16, it is better than VGG16 in all evaluation metrics, and its number of parameters is much lower than that of ResNet50 while its AP and Precision are higher.

[0070] (2) Batch image detection effect

[0071] This invention uses three algorithms to perform batch detection on 770 landslide samples in the training set. The detection time reflects the detection speed. A comparison of the batch detection times of the three algorithms is shown in Table 3. As shown in Table 3, VGG16 has the shortest detection time of 47s; the algorithm of this invention is the second shortest at 57s; and ResNet50 has the longest detection time of 1min18s.

[0072] Because VGG16 has the fewest parameters, its detection time is the shortest; ResNet50 has the most parameters, so its detection time is the longest. Although VGG16 has the shortest detection time, due to... Figure 7 It is evident that its detection results for landslide images are not ideal, failing to accurately identify landslide boundaries and occasionally misidentifying roads as landslides. Although the algorithm proposed in this invention has a shorter detection time than VGG16, it is still superior in many aspects. Figure 5 It can be seen that its detection results are the most accurate, the determination of landslide boundaries is the most precise, and there is no problem of misdetecting non-landslides as landslides.

[0073] Table 3 Comparison of batch detection time for landslide samples on the training set by the three algorithms

[0074] algorithm Detection time (Time) FasterR-CNN (VGG16) 47s Faster R-CNN (ResNet50) 1 minute 18 seconds Algorithm of this invention 57s

[0075] Note: Bold numbers indicate the optimal detection time.

[0076] The proposed algorithm and its batch detection performance on 770 landslide samples compared with ResNet50 and VGG16 are shown in the figure below. Figures 5 to 7 ,in Figure 5 This is the detection result of the algorithm proposed in this invention. Figure 6 These are the detection results from ResNet50. Figure 7 This is the test result for VGG16. Figure 5 Parts (a), (b), (c), and (d) of this paper demonstrate the detection results of the proposed algorithm on four different landslide samples. Figure 6 Parts (a), (b), (c), and (d) in the figure show the detection results of ResNet50 on four different landslide samples. Figure 7 Parts (a), (b), (c), and (d) show the detection results of VGG16 on four different landslide samples. Figure 6 The detection results shown have false detection issues. ResNet50 is prone to identifying non-landslides as landslides, and it is not accurate enough for landslide boundary detection. Figure 7 The test results, compared to Figure 6 While it exhibits fewer false detections, VGG16 remains insufficiently accurate for landslide boundary detection; conversely... Figure 5 The detection results showed almost no false detections, and the algorithm proposed in this invention is quite accurate for landslide boundary detection.

[0077] The ResNet50 PR curve exhibits excessive precision fluctuations when Recall is low, easily leading to misidentification of non-landslides as landslides. While the VGG16 PR curve is relatively accurate with low Recall and fewer false detections, its precision drops too rapidly with high Recall, resulting in inaccurate landslide boundary detection. The algorithm proposed in this invention exhibits a smoother precision fluctuation in the PR curve when Recall is low, thus accurately identifying landslides and minimizing false detections; and when Recall is high, the precision does not drop too rapidly, thus ensuring accurate landslide boundary detection.

[0078] Currently, the Faster R-CNN-based target detection algorithm is highly effective for landslide disaster detection. However, due to the excessive depth of the feature extraction network, Faster R-CNN suffers from gradient vanishing during training, making it impossible to adjust network layer weights. Therefore, this invention improves upon the Faster R-CNN algorithm, proposing a new remote sensing image target detection algorithm. This invention replaces the backbone feature extraction networks VGG16 and ResNet50 in the Faster R-CNN algorithm with DarkNet53. DarkNet53 employs skip connections to avoid the gradient vanishing problem caused by excessive network depth. Furthermore, this invention uses data augmentation to expand the landslide dataset to address the issue of insufficient landslide samples in the training set. Finally, experiments are conducted to compare the proposed algorithm with Faster R-CNN (VGG16) and Faster R-CNN (ResNet50) algorithms to evaluate the effectiveness of the proposed algorithm. Experimental results show that, on the same landslide dataset, the algorithm of this invention can achieve an AP of 92.05%, an F1 score of 0.80, a recall of 93.80%, and a precision of 69.59%. Compared with the Faster R-CNN (VGG16) algorithm, the AP is improved by 10.71%, and compared with the Faster R-CNN (ResNet50) algorithm, the AP is improved by 0.42%.

[0079] In order to perform the methods corresponding to the above embodiments and achieve the corresponding functions and technical effects, a remote sensing image landslide detection system is provided below, including:

[0080] The network improvement module is used to replace the backbone feature extraction network in the Faster R-CNN model with DarkNet53 to obtain an improved Faster R-CNN model.

[0081] The dataset acquisition module is used to acquire a landslide dataset containing multiple landslide remote sensing image samples.

[0082] The augmentation module is used to augment the landslide dataset using data augmentation methods to obtain an enhanced landslide dataset.

[0083] The training module is used to train the improved Faster R-CNN model using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model.

[0084] The identification module is used to input the remote sensing image of the landslide to be detected into the trained improved Faster R-CNN model to identify the landslide in the image.

[0085] A BR layer is added between the DarkNet53 and RPN layers in the improved Faster R-CNN model; the BR layer includes a BN layer and a Leaky ReLU layer.

[0086] Data augmentation methods include: color dithering, color dithering + brightness adjustment, color dithering + brightness adjustment + Gaussian noise addition, Gaussian noise addition, brightness adjustment, flipping, scaling + translation, and rotation.

[0087] When training the improved Faster R-CNN model: the initial learning rate is set to 0.0001, and the number of training rounds is 120; the training is divided into two phases: the backbone freezing phase and the backbone unfreezing phase; the first 50 rounds are the backbone freezing phase with a batch size of 8; the last 70 rounds are the backbone unfreezing phase with a batch size of 4; the same random number seed is used in each iteration of training the improved Faster R-CNN model to ensure that the initial weights of the model training are the same.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0089] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for detecting landslides in remote sensing images, characterized in that, include: The backbone feature extraction network in the Faster R-CNN model is replaced with DarkNet53 to obtain an improved Faster R-CNN model. A BR layer is added between the DarkNet53 and RPN layers in the improved Faster R-CNN model. The BR layer includes a BN layer and a Leaky ReLU layer to ensure that the output feature layers meet the same distribution and avoid overfitting. Obtain a landslide dataset containing multiple landslide remote sensing image samples; The landslide dataset was augmented using data augmentation techniques to obtain the augmented landslide dataset. The improved Faster R-CNN model was trained using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model. The training of the improved Faster R-CNN model was divided into two phases: a backbone freezing phase and a backbone unfreezing phase. The first 50 rounds were the backbone freezing phase with a batch size of 8. The last 70 rounds were the backbone unfreezing phase with a batch size of 4. The same random number seed was used in each iteration of training the improved Faster R-CNN model to ensure that the initial weights of the model training were the same. The remote sensing image of the landslide to be detected is input into the trained improved Faster R-CNN model to identify the landslide in the image.

2. The landslide detection method for remote sensing images according to claim 1, characterized in that, The data augmentation methods include: color dithering, color dithering + brightness adjustment, color dithering + brightness adjustment + Gaussian noise addition, Gaussian noise addition, brightness adjustment, flipping, scaling + translation, and rotation.

3. The landslide detection method for remote sensing images according to claim 1, characterized in that, When training the improved Faster R-CNN model: The initial learning rate was set to 0.0001, and the number of training rounds was 120.

4. A landslide detection system based on remote sensing images, characterized in that, include: The network improvement module is used to replace the backbone feature extraction network in the Faster R-CNN model with DarkNet53 to obtain an improved Faster R-CNN model; a BR layer is added between the DarkNet53 and RPN layers in the improved Faster R-CNN model; the BR layer includes a BN layer and a Leaky ReLU layer, so that the output feature layers meet the same distribution and avoid overfitting problems. The dataset acquisition module is used to acquire a landslide dataset containing multiple landslide remote sensing image samples; The augmentation module is used to augment the landslide dataset using data augmentation methods to obtain an enhanced landslide dataset. The training module is used to train the improved Faster R-CNN model using the enhanced landslide dataset to obtain the trained improved Faster R-CNN model. The training of the improved Faster R-CNN model is divided into two training phases: a backbone freezing phase and a backbone unfreezing phase. The first 50 rounds are the backbone freezing phase with a batch size of 8. The last 70 rounds are the backbone unfreezing phase with a batch size of 4. The same random number seed is used in each iteration of training the improved Faster R-CNN model to ensure that the initial weights of the model training are the same. The identification module is used to input the remote sensing image of the landslide to be detected into the trained improved Faster R-CNN model to identify the landslide in the image.

5. The remote sensing image landslide detection system according to claim 4, characterized in that, The data augmentation methods include: color dithering, color dithering + brightness adjustment, color dithering + brightness adjustment + Gaussian noise addition, Gaussian noise addition, brightness adjustment, flipping, scaling + translation, and rotation.

6. The landslide detection system based on remote sensing images according to claim 4, characterized in that, When training the improved Faster R-CNN model: The initial learning rate was set to 0.0001, and the number of training rounds was 120.

Images