Landslide detection method based on lightweight model

Abstract

The invention provides a landslide detection method based on a lightweight model, and belongs to the technical field of identification and monitoring of landslide disasters. According to the method, a YOLOv7-tiny model is adopted, and in a network structure of the YOLOv7-tiny model, a Mish activation function is used in a backbone network, and a Hard Sash activation function is used in a feature enhancement network and a detection network; in the backbone network, replacing an ELAN (Enhanced Local Area Network) module in the Stage 4 with a BoT-ELAN module; in the feature enhanced network, the standard convolution SC is replaced by the GSConv, and the GSConv is replaced by the ELAN to obtain an improved GSELAN module; and selecting a WIOU Loss function as a loss function of the bounding box. According to the invention, the lightweight landslide detection network is constructed based on YOLOv7-tiny, the landslide detection performance is fully improved through methods of improving the network structure, using a new activation function and the like, and remarkable advantages are shown.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide disaster identification and monitoring, and specifically to a landslide detection method based on a lightweight model. Background Art

[0002] Landslides are the most common and typical geological disasters, and at the same time, they are also geological disasters with the widest influence range and the greatest losses. They not only have a huge impact on casualties and the economy, but also cause a certain degree of damage to the ecology, resources, and environment. Therefore, timely collection of landslide information is crucial for taking appropriate remedial measures and evacuating people.

[0003] Currently, the identification of landslide disasters mainly relies on on-site inspections and visual interpretation. Although this method can provide accurate results, the investigation process is time-consuming and vulnerable to various factors, and there is a certain degree of danger. With the development of drone technology and satellite applications, some progress has been made in drone aerial photography and satellite image technology, but it is still limited by the complexity of spatial information, resulting in low classification accuracy.

[0004] Traditional computer algorithms have limitations in landslide image classification and are difficult to effectively eliminate interference data, thus affecting the classification accuracy. Although the development of computer vision and deep learning technologies has brought new opportunities for landslide identification, there is currently no efficient and accurate method that can overcome the limitations of the existing technology and achieve automatic and accurate identification of landslide areas.

[0005] In the field of object detection, especially in landslide detection, there are already some implementation schemes and existing technologies similar to this research. Some traditional object detection algorithms such as Faster R-CNN, SSD (Single Shot MultiBoxDetector), and the YOLO series (You Only Look Once) are widely used in object detection tasks, including landslide detection:

[0006] 1. Faster R-CNN: Faster R-CNN is a classic object detection algorithm that generates candidate boxes by introducing a Region Proposal Network, and then uses a classifier and a regressor for object detection. Faster R-CNN may face the problem of slow speed in landslide detection tasks because it requires two-stage training and inference and is relatively complex.

[0007] 2. SSD (Single Shot MultiBox Detector): SSD is a single-stage object detection algorithm that achieves object detection by predicting the location and category of objects at different scales. SSD is faster than Faster R-CNN, but there may be certain challenges in small object detection and localization accuracy.

[0008] 3. YOLO series: The YOLO series is a class of object detection algorithms known for their speed. Its design idea is to transform the object detection task into a regression problem of a single neural network. YOLOv7-tiny, as the basic framework of this study, has advantages in terms of speed and lightweight network design compared to traditional YOLO versions.

[0009] Although these traditional object detection algorithms have achieved certain results in landslide detection tasks, there is still room for improvement in terms of the accuracy, speed, and network complexity of landslide detection. Specifically, their defects mainly include technical aspects, cost aspects, and efficiency aspects.

[0010] 1. Disadvantages in technical aspects:

[0011] Relying on manual field inspections and visual interpretation: Traditional methods mainly rely on manual field inspections, which are subjective and limited, resulting in low recognition accuracy.

[0012] Insufficient classification accuracy: Existing technologies have difficulties in dealing with complex geological environments and eliminating interference data, affecting the accurate identification of landslide areas.

[0013] High complexity: Traditional algorithms such as Faster R-CNN and SSD have relatively complex network structures in landslide detection, with many model parameters, making training and optimization difficult.

[0014] 2. Disadvantages in cost aspects:

[0015] High cost: Traditional manual field inspections require a large amount of time and human resources, resulting in high costs.

[0016] Equipment and technology costs: Using drone aerial photography and satellite image technology requires investment in expensive equipment and technical support, increasing the cost burden. Additionally, some existing technologies may require high-performance GPUs or TPUs for training and inference, which increases the investment cost of hardware devices.

[0017] 3. Disadvantages in efficiency aspects:

[0018] Large time consumption and low efficiency: Traditional methods take a long time and cannot quickly respond to the monitoring needs of landslide disasters. Existing technologies are not efficient and cannot achieve rapid and accurate identification of large-scale areas, affecting the effectiveness of disaster warning and emergency response.

[0019] Slow training and inference speed: Traditional algorithms have a slow training speed when dealing with large-scale datasets. At the same time, the inference speed may not be fast enough for real-time landslide monitoring, affecting the effect of real-time monitoring. Moreover, the complex network structure and a large number of parameters lead to a high computational complexity of the algorithm, reducing the overall efficiency.

[0020] Low resource utilization: The existing technologies have low utilization of computing resources, resulting in a large amount of resource waste and low efficiency during the training and inference processes. Summary of the Invention

[0021] The technical problem to be solved by the present invention is: how to develop a landslide detection method with a more streamlined algorithm and higher accuracy and precision.

[0022] Another technical problem to be solved by the present invention is how to further reduce the cost of landslide detection and improve the efficiency.

[0023] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0024] A landslide detection method based on a lightweight model, which uses the YOLOv7-tiny model. In the network structure of the YOLOv7-tiny model: the Mish activation function is used in the backbone network, and the Hard Swish activation function is used in the feature enhancement network and the detection network; in the backbone network, the ELAN module in Stage 4 is replaced with the BoT-ELAN module; in the feature enhancement network, GSConv is used to replace the standard convolution SC, and GSConv is replaced into ELAN to obtain the improved GSELAN module; the WIOU Loss function is selected as the loss function of the bounding box.

[0025] Preferably, the Mish activation function is as shown in the following formula 1:

[0026]

[0027] The Hard Swish activation function is as shown in the following formula 2:

[0028]

[0029] Preferably, the BoT-ELAN module is an improved BoT-ELAN module obtained by using Bottleneck Transformer to replace the ordinary convolution of the original ELAN.

[0030] Preferably, in the process of replacing the standard convolution SC with GSConv, GSConv first transforms a feature map M1 with an input channel number C1 into a feature map M2 with an output channel number of C2 / 2 using SC; then, applies DSC to M2 to obtain another feature map M3 with an output channel number of C2 / 2; then, concatenates M2 and M3 through the Concat operation and uses the Shuffle operation to obtain a feature map with a channel number of C2.

[0031] Preferably, in the process of replacing GSConv into ELAN, replace the two SCs with a 3×3 convolution kernel in the original ELAN module with GSConv, and keep the SC with a 1×1 convolution kernel unchanged.

[0032] Preferably, the calculation formula of the WIOU Loss function is shown in the following formulas 3 to 8:

[0033] L WIOUv3 = rL WIOUv1 Formula 3;

[0034]

[0035] L IOU = 1 - IOU Formula 6;

[0036] L WIOUv1 = R WIOU L IOU Formula 7;

[0037]

[0038] Among them, x and y represent the position of the predicted bounding box, x gt , y gt are the positions of the corresponding target bounding box, w c , h c represent the width and height of the predicted bounding box respectively, IOU is the intersection over union of the predicted bounding box and the ground truth bounding box, β is the constructed non-monotonic focusing coefficient, and α and δ are hyperparameters.

[0039] The present invention provides a landslide detection method based on a lightweight model. This technical solution constructs a lightweight landslide detection network based on YOLOv7-tiny, and through methods such as improving the network structure and using a new activation function, fully improves the performance of landslide detection, showing significant advantages.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. Higher degree of automation: The present invention adopts a deep learning algorithm, reduces the dependence on manual on-site inspections, realizes a higher degree of automated landslide detection, and improves the objectivity and accuracy of detection.

[0042] 2. Higher classification accuracy: By optimizing the algorithm and feature extraction technology, the present invention performs better in dealing with complex geological environments and eliminating interference data, improving the accurate recognition rate of landslide areas and further enhancing the classification accuracy.

[0043] 3. Simplified complexity: The algorithm design of the present invention simplifies the network structure, reduces the model complexity and the number of parameters, making training and optimization easier and improving the efficiency and performance of the landslide detection algorithm.

[0044] 4. Lower overall cost: By reducing the human resource requirements, simplifying the equipment requirements and technical costs, the present invention reduces the overall cost of the landslide detection process, making the landslide monitoring technology more cost-effective and operable.

[0045] 5. Higher efficiency and speed: The present invention optimizes the algorithm and calculation process, improving the efficiency and speed of landslide detection, achieving faster response and real-time monitoring capabilities, and meeting the requirements for efficiency and speed in some application scenarios.

[0046] In addition, the beneficial effects of the present invention are also reflected in the following aspects:

[0047] 1. Technical aspects:

[0048] Reduced dependence on manual on-site inspections: The present invention introduces an automated landslide detection algorithm based on deep learning, reducing the need for manual on-site inspections and improving the objectivity and accuracy of detection.

[0049] Improved classification accuracy: By introducing advanced image processing and feature extraction technologies, the present invention can better handle complex geological environments and eliminate interference data, thereby improving the accurate recognition rate of landslide areas.

[0050] Simplified complexity: The present invention adopts a streamlined network structure and parameter optimization method, reducing the complexity of the algorithm and improving the efficiency and performance of landslide detection.

[0051] 2. Cost aspects:

[0052] Cost reduction: By reducing the human resource requirements and simplifying the equipment requirements, the present invention significantly reduces the overall cost of the landslide detection process, making the landslide monitoring technology more cost-effective and easy to apply.

[0053] Equipment and technology costs: The present invention adopts a more intelligent landslide detection method, reducing the dependence on expensive equipment and technical support and lowering the hardware equipment investment cost.

[0054] 3. Efficiency aspects:

[0055] Improve efficiency: Through the algorithm optimization and technological innovation of the present invention, the efficiency of the landslide detection process is significantly improved, enabling a rapid response to the monitoring requirements of landslide disasters and enhancing the disaster warning and emergency response effects.

[0056] Accelerate training and inference speed: The present invention optimizes the algorithm structure and calculation process, improves the training and inference speed, realizes real-time landslide monitoring, and meets the application scenarios with high speed requirements.

[0057] Improve resource utilization rate: By optimizing resource management and computational resource utilization, the present invention reduces resource waste, improves the efficiency and performance in the landslide detection process, and solves the problem of low resource utilization rate in the existing technology. Description of the Drawings

[0058] Figure 1 It is the overall structure of the lightweight YOLOv7 network.

[0059] Figure 2 It is the structural diagram of the BoT-ELAN module.

[0060] Figure 3 It is the structural diagram of the GS-ELAN module.

[0061] Figure 4 It is the visualization comparison diagram of feature maps.

[0062] Figure 5 It is the visualization comparison diagram of landslide detection between the improved algorithm and the original algorithm. Detailed Embodiments

[0063] The following will describe the detailed embodiments of the present invention in detail. To avoid excessive unnecessary details, the well-known structures or functions will not be described in detail in the following embodiments. The approximate language used in the following embodiments can be used for quantitative expressions, indicating that certain changes in quantity are allowed without changing the basic functions. Unless otherwise defined, the technical and scientific terms used in the following embodiments have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0064] I. Research Method

[0065] To meet the effectiveness of the landslide detection network and simultaneously realize the lightweight of the detection network, the present technology proposes a lightweight landslide detection algorithm based on YOLOv7-tiny (hereinafter referred to as YOLOv7), and the overall framework of this algorithm is as Figure 1As shown in the figure. The present technology has made the following four improvements to the YOLOv7 network: (1) The Mish function and the Hard Swish function can enhance the non-linear expression ability and robustness of the network. Therefore, the present technology uses the Mish activation function in the backbone network and the Hard Swish activation function in the feature enhancement network and the detection network; (2) In the backbone network, the present technology replaces the ELAN module in Stage 4 with the BoT-ELAN module, reducing the network model parameters while improving the network's ability to associate context information; (3) In the feature enhancement network, the present technology uses GSConv to replace the standard convolution (SC), and replaces GSConv into ELAN to obtain the improved GSELAN module, reducing the number of parameters and the computational complexity of the network model and strengthening the fusion of features extracted from the backbone network; (4) The present technology selects the WIOU Loss function as the loss function for the bounding box, avoiding the instability of network training caused by low-quality samples and improving the generalization ability of the network model.

[0066] 1. Setting of the activation function

[0067] There are some problems with the original Leaky ReLU activation function. For example, in the two intervals where the domain is greater than zero and less than zero, the expression of the Leaky ReLU function is linear, which limits the model's ability to process complex data. The function is not smooth, which may lead to instability during training. Therefore, in view of the above problems, the present invention selects the Mish function and the Hard Swish function as the activation functions of the network model.

[0068] The Mish function has the characteristics of being smooth, non-linearly monotonic, unbounded above and bounded below. The characteristics of the smooth function endow it with better generalization ability. The non-linearly monotonic characteristic stabilizes the gradient flow and improves the fitting ability of the network. The unbounded above characteristic can avoid the gradient saturation problem and prevent the training speed from dropping sharply. The bounded below characteristic can impose a strong regularization effect and effectively avoid overfitting of the network. The expression of the Mish function is shown in Equation (1):

[0069]

[0070] The Hard Swish function is an approximately linear activation function. This function takes into account both non-linear expression and computational efficiency, and has a lower computational complexity. At the same time, the Hard Swish function has good gradient propagation characteristics. It can be backpropagated through simple derivative calculations and the chain rule, thereby effectively updating the network parameters. This helps to alleviate optimization problems such as gradient disappearance and gradient explosion during training. The expression of the Hard Swish function is shown in Equation (2):

[0071]

[0072] After multiple experiments, it is proved that the Mish activation function has stronger feature extraction ability in the backbone network. The approximate linear characteristic of the Hard Swish function makes it easier to train and optimize in deeper neural network structures and can better adapt to complex data representation requirements. Therefore, the present invention finally selects the Mish function as the activation function of the backbone network and the Hard Swish function as the activation function of the feature enhancement network and the detection network.

[0073] 2. BoT-ELAN Module Based on Bottleneck Transformer

[0074] Bottleneck Transformer is a simple but effective backbone architecture, which is widely used in various computer vision tasks. Compared with convolutional neural networks, the neural network based on Bottleneck Transformer not only has fewer parameters, but also can extract image context information more effectively, model the correlation between global features of the graph, thereby improving the efficiency of object detection. For the landslide detection task, the features of the old landslide area are often similar to those of the background area and difficult to distinguish, while the global context information can provide more comprehensive scene semantic information, which helps the model better understand the entire scene. Therefore, the present invention uses Bottleneck Transformer to replace the ordinary convolution of the original ELAN to obtain the improved BoT-ELAN module, and its structure is as Figure 2 shown.

[0075] Although the BoT-ELAN module has fewer parameters, using this module in the shallow network will greatly increase the computational amount of the model. Therefore, the present invention chooses to replace the ELAN in Stage 4 with BoT-ELAN only, achieving a balance between model performance and model computational amount.

[0076] 3. GSConv and GS-ELAN Modules

[0077] In order to improve the extraction of feature information in the image while reducing the number of parameters, the present invention selects GSConv, which can reduce the model parameters and ensure the detection accuracy, to replace SC, and obtains the improved GS-ELAN module based on GSConv.

[0078] SC can preserve the hidden correlations between each channel. Depth-wise separable convolution (DSC) can reduce the number of network parameters and computational volume, but it loses the correlation information between channels. Therefore, when the network depth increases, the feature information extracted by DSC will be much lower than that of SC. GSConv combines the structures of SC and DSC, achieving the reduction of parameters while preserving the correlation information between channels. GSConv first transforms a feature map M1 with an input channel number C1 into a feature map M2 with an output channel number of C2 / 2 using SC; secondly, uses DSC on M2 to obtain another feature map M3 with an output channel number of C2 / 2; finally, connects M2 and M3 through the Concat operation and uses the Shuffle operation to obtain a feature map with a channel number of C2. Since GSConv first performs SC and then DSC, using GSConv will inevitably deepen the network structure, and each spatial (width and height) compression and channel expansion of the feature map will result in the loss of some semantic information. If GSConv is used throughout the model, the network layer of the model will be deeper, which will exacerbate the resistance of the data flow and significantly increase the network training time. In the feature enhancement network, the channel dimension of the feature map reaches the maximum, and the width and height dimensions reach the minimum, and no further dimension transformation is required. Therefore, the present invention chooses to introduce GSConv into the feature enhancement network, and based on GSConv, replaces the SC in the original ELAN module with GSConv to obtain an improved GS-ELAN module. The structure of the GS-ELAN module is as Figure 3 shown.

[0079] In order to avoid the network structure from being too deep and ensure the minimum loss of semantic information, the present invention chooses to replace the SC with two 3×3 convolutional kernels in the original ELAN module with GSConv, and keep the SC with a 1×1 convolutional kernel unchanged. This improvement further reduces the number of parameters and avoids the network becoming deeper due to the use of too many GSConv, thus reducing the network performance.

[0080] 4. WIOU Loss Function

[0081] The bounding box loss function is a function used to measure the difference between the predicted bounding box and the true bounding box in the object detection task. Its reasonable definition will significantly improve the performance of the object detection model.

[0082] The original network used the CIOU loss function. CIOU is insensitive to the scale of the predicted bounding box and has a small penalty for larger bounding boxes. In addition, there may be low-quality examples in the dataset. If the bounding box regression of low-quality samples is blindly emphasized, it will harm the detection performance of the model. To address these issues, the present invention chooses to use WIOU as the bounding box loss function.

[0083] WIOU evaluates the quality of anchor boxes by adopting a dynamic non - monotonic focusing mechanism and combines the method of gradient gain. It can not only ensure the effect of high - quality anchor boxes but also reduce the influence of low - quality samples, thereby improving the overall performance of the algorithm. There are three versions of WIOU. Through experiments, it is finally proved that WIOU v3 with the best effect is selected as the loss function of the network of the present invention, and its calculation formula is as follows:

[0084] L WIOUv3 =rL WIOUv1 Equation 3;

[0085]

[0086] L IOU =1 - IOU Equation 6;

[0087] I WIOUv1 =R WIOU L IOU Equation 7;

[0088]

[0089] Among them, x and y represent the position of the predicted box, x gt , y gt are the positions of the corresponding target boxes, w c , h c respectively represent the width and height of the predicted box, IOU is the intersection - over - union of the predicted box and the ground - truth box, β is the constructed non - monotonic focusing coefficient, and α and δ are hyperparameters.

[0090] II. Simulation Experiments and Analysis

[0091] 1. Dataset Preparation

[0092] The dataset of the present invention contains 340 landslide images, which mainly come from the landslide area of the Loess Plateau, but the landslide areas of these images are not labeled. Therefore, the present invention uses the labelimg image annotation tool to label the landslide areas in the images. Deep neural networks usually require a large amount of data. To solve this problem, the present invention uses data augmentation to increase the diversity of the landslide dataset to better train the neural network model. The data augmentation methods of the present invention include rotation transformation, scaling transformation, Mixup, and Mosaic. Finally, a dataset with a total of 640 landslide images is obtained, and the dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0093] 2. WIOU Loss Function

[0094] In the experiments of the present invention, the model performance evaluation metrics include Precision, Recall, Average Precision (AP), the number of parameters (Params), and the model computational amount (GFLOPs). These metrics are used to evaluate the performance of the model in the landslide detection task. The formulas for Precision and Recall are shown in Equations (9) and (10).

[0095]

[0096] Among them, TP (True Positive) represents the number of correctly predicted positive samples; FN (False Negative) represents the number of incorrectly predicted negative samples; FP (False Positive) represents the number of incorrectly predicted positive samples. Average Precision can be used to measure the quality of the network model, which is represented by the area under the Precision-Recall (Pr-Re) curve, and its calculation process is shown in Equation (11).

[0097]

[0098] Among them, Pr(Re) is the functional expression of the curve composed of Precision and Recall.

[0099] The number of parameters (Params) refers to the total number of all learnable parameters in the neural network model. The number of parameters is usually used to measure the size and complexity of the model. The model computational amount (GFLOPs) is the number of floating-point operations executed by the model in one second, which is used to evaluate the computational efficiency of the model. Its calculation formula is as follows:

[0100] Params = C in ×C out ×K 2 Equation 12;

[0101] GFLOPs = W × H × C in ×C out ×K 2 Equation 13;

[0102] Among them, C in and C out respectively represent the number of input and output feature channels, K represents the size of the convolutional kernel, and W and H respectively represent the width and height of the input feature map.

[0103] 3. Experiments and Analysis

[0104] The experiments of the present invention were carried out under the Windows 10 operating system; a single GPU with the model of NVIDIA TESLA T4 (12GB RAM) was used; the deep learning frameworks used were Pytorch 1.11.0 and Torchvision 0.12.0; the compiler used was Python 3.7. The input image size was set to 640×640 in the experiment, and the Adam optimizer with an initial learning rate of 1e-2 and a minimum learning rate of 1e-4 was used to guide the optimization. In the network training, the cosine annealing learning rate strategy (COS) was adopted to adjust the learning rate. The batch size (Batch_size) was set to 16, and the number of training epochs was set to 500.

[0105] To verify the effectiveness of the activation function selected in the present invention, the present invention conducted experimental comparisons, comparing the lightweight network that uses only the Mish activation function and the HardSwish function with the network proposed in the present invention. The experimental results are shown in Table 1. It can be observed from the table that the activation function selected in the present invention has been improved in various indicators, proving the effectiveness of the method of the present invention.

[0106] Table 1 Comparison of experimental results of different activation functions

[0107] Activation function Precision Recall AP Mish / % 95.45 72.41 82.12 HardSwish / % 95.56 74.14 82.23 Mish + HardSwish / % 96.70 75.86 85.26

[0108] Secondly, the present invention conducted a comparative experiment, comparing the WIOU loss function adopted with the CIOU, DIOU, GIOU, EIOU, and SIOU loss functions to verify the effectiveness of the loss function selected in the present invention in the landslide detection task. The experimental results are shown in Table 2.

[0109] Table 2 Comparison of experimental results of different loss functions

[0110] Loss function Precision Recall AP CIOU / % 95.65 75.86 84.14 DIOU / % 88.04 69.83 77.02 GIOU / % 91.95 68.97 81.12 EIOU / % 92.22 71.55 81.58 SIOU / % 96.70 74.14 85.26 WIOU / % 96.70 75.86 85.26

[0111] It can be observed from the table that the WIOU loss function has reached the optimal level in various indicators, indicating that the WIOU loss function is more suitable for the landslide detection task. Therefore, the present invention selects WIOU as the loss function of the network model.

[0112] To verify the effectiveness of the improved modules in the backbone network and the feature enhancement network, ablation experiments were conducted on the improved modules for comparison. Among them, the BoT-ELAN module (Module A) introduced in the backbone network and the GSConv and GSELAN modules (Module B) introduced in the feature enhancement network. The experimental results are shown in Table 3. It can be seen from the table that when Module A and Module B are added separately, the number of network parameters and the model calculation amount are both reduced, meeting the requirements of lightweight, and at the same time, the detection performance of the network is improved. Finally, when Module A and Module B are added to the network at the same time, the results show that the number of network parameters is reduced by 1.139M, the model calculation amount is reduced by 1.388G, and the detection performance of the network model is further improved.

[0113] Table 3 Comparison of ablation experiments for each module

[0114]

[0115]

[0116] To more intuitively verify the effectiveness of the algorithm of the present invention, the feature maps of the original algorithm and the improved algorithm were visualized, and the results are as Figure 4 shown. Since the information finally extracted by the network is the coordinate information of the anchor box, the visualization of the feature map mainly represents the visualization of the center point of the detection box. The closer the heat map is to the center of the landslide area, the more accurate the detection is, and the darker the color of the heat map indicates the higher the confidence of the detection box. From Figure 4 a, it can be seen that the center of the detection box of the improved algorithm is closer to the center of the landslide area, indicating that the improved algorithm has higher accuracy. From Figure 4 c, it can be seen that the improved algorithm can detect the landslide areas missed by the original algorithm. Therefore, the improved algorithm is more effective in the task of detecting landslides.

[0117] To better verify the performance of the improved algorithm, the present invention visually compared the detection results of the original algorithm and the improved algorithm, and the results are as Figure 5 shown. In Figure 5 a, there are misdetection problems in the detection of landslides by the original algorithm. Only some landslide areas can be detected with a high confidence, indicating that the misdetection problem of the original algorithm is relatively serious. Although the confidence of the improved algorithm is not high, it can detect the entire landslide area completely. In Figure 5 b, there are multiple small-scale landslides. The original algorithm has a missed detection problem, while the improved algorithm successfully detects all the landslide areas in the figure. In addition, Figure 5 c shows a picture of an old landslide. The original algorithm performs poorly in the detection of old landslides, while the improved algorithm successfully identifies the old landslide.

[0118] To verify the superiority of the lightweight algorithm of the present invention compared to other algorithms, the present invention conducts a comparative experiment on the lightweight algorithm with classic algorithms such as SSD, RetinaNet, Fcos, YOLOv4-tiny, YOLOv5s, and YOLOv5s-v6.0. The experimental results are shown in Table 4.

[0119] Table 4 Comparison of Experimental Results between the Algorithm of the Present Invention and Other Classic Algorithms

[0120] Precision Recall AP Params GFLOPs SSD 79.31％ 39.66％ 58.17％ 11.671M 13.100G Retinanet 80.00％ 58.62％ 69.54％ 19.771M 109.031G Fcos 75.76％ 64.66％ 70.42％ 32.111M 161.135G YOLOv4 - tiny 87.78％ 68.10％ 78.98％ 5.874M 16.175G YOLOv5s 84.38％ 43.68％ 52.17％ 7.064M 16.477G YOLOv5s - v6.0 91.11％ 70.69％ 79.36％ 7.235M 16.625G The present invention 96.70％ 75.86％ 85.26％ 4.875M 11.793G

[0121] By comparing the data in the table, it can be concluded that the algorithm proposed in the present invention has advantages in performance compared to various classic one-stage detection algorithms based on anchor boxes. At the same time, this algorithm not only improves the detection performance but also successfully reduces the number of model parameters and the computational complexity. Compared with YOLOv5s-v6.0 with the best performance among the classic algorithms, the algorithm proposed in the present invention has an average precision improvement of 5.1%, a recall improvement of 4.59%, an F1-score improvement of 5.83%, a reduction of 2.360M in the number of parameters, and a reduction of 4.832G in the computational complexity. Therefore, the improved algorithm model proposed in the present invention achieves a good balance between accuracy and lightweight, and successfully solves the problem that accuracy and lightweight cannot be optimized simultaneously.

[0122] III. Summary

[0123] The present invention uses landslide images on the Loess Plateau as data sources, and establishes a landslide detection dataset using object detection annotation tools and image enhancement techniques. The lightweight YOLOv7 landslide detection network proposed by the present invention is verified on this dataset. The network of the present invention uses YOLOv7-tiny as the basic structure of the model, adopts the Mish function as the activation function of the backbone network, and adopts the Hard Swish function as the activation function of the feature enhancement network and the detection network. In addition, the present invention designs a BoT-ELAN module based on Bottleneck Transformer and applies it to the backbone network, and applies the GSConv and GS-ELAN modules to the feature enhancement network. These improvements reduce the number of network parameters and the amount of computation, while improving the detection performance of the network. Finally, the present invention uses the WIOU loss function to supervise network training, effectively reducing the impact of low-quality samples in the dataset on the network performance, thereby improving the stability of network training. The proposed landslide detection network performs excellently in terms of accuracy, stability, and simplicity, and has excellent generalization ability and robustness. Experimental results on the landslide detection dataset show that compared with the algorithm before improvement, the lightweight algorithm proposed by the present invention reduces the number of network parameters by 18.94%, reduces the model computation amount by 10.53%, while the Precision increases by 2.59%, the Recall increases by 2.58%, and the AP increases by 2.03%. This indicates that the method proposed by the present invention has made remarkable improvements in the landslide detection task.

[0124] The innovation of the present invention is concentrated in the following aspects:

[0125] 1. Optimization of deep learning algorithms: The key point of the present invention is to optimize deep learning algorithms for landslide detection, including network structure design, feature extraction, and model training, to achieve efficient and accurate landslide identification.

[0126] 2. Automatic landslide detection technology: The present invention adopts automatic landslide detection technology as a protection point, and realizes the rapid and accurate identification of landslide areas by introducing intelligent algorithms and image processing technology.

[0127] 3. Cost-benefit optimization: The present invention focuses on optimizing cost-benefit, and improves the economic benefits of landslide monitoring by simplifying processes, reducing hardware requirements, and reducing labor input, bringing protection points for the invention.

[0128] 4. Real-time monitoring technology: The key technical points include improving the training and inference speed to achieve real-time landslide monitoring to meet the real-time requirements of application scenarios, providing protection points for the invention.

[0129] The above has described the embodiments of the present invention in detail, but the above content is only the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the scope of the application of the present invention shall be included within the protection scope of the present invention.

Claims

1. Landslide detection method based on a lightweight model. This method uses the YOLOv7-tiny model, and is characterized in that, In the network structure of the YOLOv7-tiny model: the Mish activation function is used in the backbone network, and the Hard Swish activation function is used in the feature enhancement network and the detection network; in the backbone network, the ELAN module in Stage4 is replaced with the BoT-ELAN module; In the feature enhancement network, GSConv is used to replace the standard convolution SC, and GSConv is replaced into ELAN to obtain the improved GSELAN module; the WIOU Loss function is selected as the loss function of the bounding box.

2. The landslide detection method based on a lightweight model according to claim 1, wherein The Mish activation function is shown in Equation 1 below: The Hard Swish activation function is shown in Equation 2 below:

3. The landslide detection method based on a lightweight model according to claim 1, wherein The BoT-ELAN module is an improved BoT-ELAN module obtained by using Bottleneck Transformer to replace the ordinary convolution of the original ELAN.

4. The landslide detection method based on a lightweight model according to claim 1, wherein In the process of using GSConv to replace the standard convolution SC, GSConv first uses SC to transform a feature map M1 with an input channel number C1 into a feature map M2 with an output channel number of C2 / 2; then, M2 is processed with DSC to obtain another feature map M3 with an output channel number of C2 / 2; then, M2 and M3 are connected through the Concat operation and the Shuffle operation is used to obtain a feature map with a channel number of C2.

5. The landslide detection method based on a lightweight model according to claim 1, wherein, In the process of replacing GSConv into ELAN, the two SCs with a convolution kernel of 3×3 in the original ELAN module are replaced with GSConv, and the SC with a convolution kernel of 1×1 remains unchanged.

6. The landslide detection method based on a lightweight model according to claim 1, wherein, The calculation formula of the WIOU Loss function is shown in Equations 3 to 8 below: L WIOUv3 = rL WIOUv1 Equation 3; L IOU = 1 - IOU formula 6; L WIOUv1 = R WIOU L IOU Equation 7; Among them, x and y represent the positions of the predicted bounding boxes, and x gt , y gt are the corresponding positions of the target bounding boxes, w c , h c represent the width and height of the predicted bounding box respectively, IOU is the intersection over union of the predicted bounding box and the ground truth bounding box, β is the constructed non-monotonic focusing coefficient, and α and δ are hyperparameters.

Images