Slope instability region target detection method based on remote sensing image deep learning

Abstract

The application discloses a kind of based on remote sensing image deep learning's side slope instability area target detection method, its method includes: S1, constructs remote sensing image sample set;S2, YOLOv8-FMC network model includes feature extraction core network, feature fusion enhancement network and target detection head, feature extraction core network is extracted multi-scale feature map F1~F3 layer by layer;Mask generation module MGB is used respectively to feature map F1~F3 Generation CAM heat map and with feature map F1~F3 corresponding scale multiplication, up-sampling and step-by-step fusion processing feature map N1~N3;S3, target detection head passes through convolution block attention module CBAM, and the output side slope instability class and boundary frame are output by convolution module;S4, obtains the side slope remote sensing image of research area input YOLOv8-FMC network model identification and divides out side slope instability boundary frame and exports.The application realizes comprehensive feature multi-scale extraction, global heat map auxiliary rapid and accurate judgment, can respond to complex mine environment and carry out the high-precision processing of side slope instability identification and segmentation.

Description

Technical Field

[0001] This invention relates to the field of slope instability detection, and more particularly to a method for detecting targets in slope instability areas based on deep learning from remote sensing images. Background Technology

[0002] With the intensification of construction activities and the impact of abnormal climate, the structure of slope soil or rock is prone to change, losing its original stability and causing slope instability phenomena, including collapse, cracking, and landslides. These phenomena threaten human safety, cause property damage, and damage the ecological environment. Against this backdrop, using remote sensing technology combined with target detection methods for automated identification of unstable slope areas can promptly detect potential risks and provide a scientific basis for disaster prevention and mitigation and ecological environmental protection.

[0003] Early extraction of information on slope instability areas relied on manual visual interpretation and delineation, which was inefficient and costly. With the development of deep learning, research has gradually shifted towards automated detection methods. For example, two-stage target detection models (such as the improved Mask R-CNN) are employed to address the shortcomings of the Mask branch in edge blurring and false detection, solving the problem of rough edges on slope cracks and thus improving the accuracy of slope crack identification. In contrast, YOLO has advantages such as lightweight structure, high detection speed, and high accuracy. To further improve detection performance, some studies have constructed an integrated model based on YOLOX-CNN and Mask R-CNN, building an integrated model suitable for small- to medium-scale target detection and large-scale target edge extraction, further enhancing the landslide target detection effect.

[0004] However, deep learning-based methods still face challenges in detecting slope instability areas: 1) Two-stage target detection models are computationally intensive, require high hardware resources, and have high training and deployment costs, limiting their widespread application. 2) Most studies only target single types of slope instability phenomena, without considering other types of instability. However, various instability phenomena (including collapse, cracks, and landslides) differ significantly in shape, size, and boundary features, significantly increasing the difficulty of target detection. 3) When targets are sparse in slope instability areas, traditional YOLO models tend to overlook scattered small targets or weak feature regions, requiring more macroscopic and global feature judgment capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a target detection method for slope instability areas based on deep learning from remote sensing images. The improved YOLOv8-FMC network model includes a feature extraction core network, a feature fusion enhancement network, and a target detection head. The feature extraction core network includes first to third layers of feature extraction networks containing a feature extraction module SDASPP. The feature extraction module SDASPP realizes comprehensive feature extraction of multi-scale information, edge features, and morphological perception. The mask generation module MGB extracts and processes the feature map to obtain a CAM heatmap. The feature extraction module SDASPP and the mask generation module MGB are combined to perform corresponding scale multiplication, upsampling, and stepwise fusion processing to obtain feature maps of different scales, realizing fast and accurate judgment of macroscopic and global features. The target detection head realizes feature enhancement through the convolutional block attention module CBAM and meets the practical application requirements of target detection in slope instability areas.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A target detection method for slope instability areas based on deep learning from remote sensing images, the method comprising:

[0008] S1. Construct a remote sensing image sample set, and label the slope instability areas in the remote sensing images in the remote sensing image sample set.

[0009] S2. Construct an improved YOLOv8-FMC network model and input it into the remote sensing image sample set for model learning and training. The YOLOv8-FMC network model includes a feature extraction core network, a feature fusion enhancement network, and a target detection head. The feature extraction core network extracts multi-scale feature maps F1 to F3 layer by layer. The feature fusion enhancement network uses the mask generation module MGB to generate CAM heatmaps of different scales from the feature maps F1 to F3 respectively, and performs corresponding scale multiplication, upsampling, and step-by-step fusion processing on the feature maps F1 to F3 to obtain feature maps N1 to N3 of different scales.

[0010] S3. The target detection head includes a convolutional block attention module (CBAM). The CBAM includes a channel attention module and a spatial attention module. The channel attention module extracts channel attention weights from feature maps N1 to N3 respectively, weights them, and then inputs them into the spatial attention module to obtain spatial attention weights. The convolutional module outputs the slope instability category and the slope instability bounding box.

[0011] S4. Obtain remote sensing images of slopes in the study area, input them into the trained YOLOv8-FMC network model, identify whether there is slope instability in the remote sensing images, identify and segment the slope instability boundary boxes in the remote sensing images where slope instability exists, and output them.

[0012] Preferably, the slope instability areas of the remote sensing images in the remote sensing image sample set are labeled with slope instability boundary boxes using LabelMe software to generate JSON files corresponding to the remote sensing images. The JSON files are then converted into TXT tags, and the TXT tags are used to create label data corresponding to the remote sensing images and associated with the remote sensing images to construct the remote sensing image sample set.

[0013] To better implement the present invention, the feature extraction core network includes a first-layer feature extraction network, a second-layer feature extraction network, and a third-layer feature extraction network. The first to third-layer feature extraction networks all include an improved feature extraction module SDASPP and a max pooling module. The first-layer feature extraction network extracts features to obtain feature map F1. Feature map F1 is input into the second-layer feature extraction network to obtain feature map F2. Feature map F2 is input into the third-layer feature extraction network to obtain feature map F3.

[0014] To better implement this invention, the feature extraction module SDASPP is an improved network structure. SDASPP includes five branches: the first branch includes a 1×1 convolution module and a Sobel-DCNv1 module; the second branch includes a 3×3 convolution module with a dilation rate of 6 and a Sobel-DCNv1 module; the third branch includes a 3×3 convolution module with a dilation rate of 12 and a Sobel-DCNv1 module; the fourth branch includes a 3×3 convolution module with a dilation rate of 18 and a Sobel-DCNv1 module; and the fifth branch includes a global average pooling module and a Sobel-DCNv1 module. The Sobel-DCNv1 module includes parallel Sobel computation modules and deformable convolution DCNv1. The Sobel computation module calculates the gradient magnitude G of the input feature map using the Sobel operator, outputting a feature map T2, which is then convolved and fused with the feature map T1 processed by the deformable convolution DCNv1. The five branches of the feature extraction module SDASPP perform feature extraction at different scales, and the feature outputs from the five branches are convolved, convolved, and fused to finally output a feature map.

[0015] Preferably, the feature fusion enhancement network multiplies feature map F3 with the CAM heatmap corresponding to feature map F3 and concatenates them to obtain feature map M1; the feature fusion enhancement network multiplies feature map F2 with the CAM heatmap corresponding to feature map F2 and upsamples feature map M1 for concatenation, and then performs feature extraction processing by the feature extraction module SDASPP to obtain feature map M2; the feature fusion enhancement network multiplies feature map F1 with the CAM heatmap corresponding to feature map F1 and upsamples feature map M2 for concatenation, and then performs feature extraction processing by the feature extraction module SDASPP to obtain feature map M3.

[0016] Preferably, the feature fusion enhancement network processes feature map M3 through the feature extraction module SDASPP to obtain feature map N1; feature map N1 is processed by dimension downsampling and channel adjustment (CBS) and concatenated with feature map M2, and then processed by the feature extraction module SDASPP to obtain feature map N2; feature map N2 is processed by dimension downsampling and channel adjustment (CBS) and concatenated with feature map M1, and then processed by the feature extraction module SDASPP to obtain feature map N3.

[0017] Preferably, the channel attention module obtains two feature maps through two parallel adaptive global max pooling layers and an adaptive global average pooling layer. The two feature maps are then input into a multilayer perceptron, and the output feature maps are summed and a sigmoid activation function is used to generate channel attention weights for each channel. The spatial attention module generates feature maps through two parallel adaptive max pooling layers and an average pooling layer, and performs channel concatenation and convolution operations. Finally, a sigmoid activation function is used to obtain the final spatial attention weights.

[0018] Preferably, the feature maps N1 to N3 are processed by parallel dimension downsampling and channel adjustment (CBS) and then stitched together before being input into the convolutional block attention module (CBAM) of the target detection head.

[0019] Preferably, the target detection head uses a 1×1 convolution module to predict, identify, and output the slope instability category; the target detection head uses another 1×1 convolution module to identify, segment, and output the slope instability bounding box of the remote sensing image containing slope instability.

[0020] Preferably, the overall loss function expression of the YOLOv8-FMC network model is as follows:

[0021] ,in The overall loss of the model, Classification loss for the target For bounding box regression loss, For distributed regression loss, The loss is handled by the mask generation module MGB.

[0022] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0023] (1) The improved YOLOv8-FMC network model of the present invention includes a feature extraction core network, a feature fusion enhancement network and a target detection head. The feature extraction core network includes a first to third layer feature extraction network containing a feature extraction module SDASPP. The feature extraction module SDASPP realizes the comprehensive feature extraction of multi-scale information, edge features and morphological perception. The mask generation module MGB extracts and processes the feature map to obtain the CAM heat map. The feature extraction module SDASPP and the mask generation module MGB are combined to perform corresponding scale multiplication, upsampling and step-by-step fusion processing to obtain feature maps of different scales, realizing the fast and accurate judgment of macro and global features. The target detection head realizes feature enhancement through the convolutional block attention module CBAM and meets the actual application requirements of target detection in slope instability areas.

[0024] (2) The improved YOLOv8-FMC network model of this invention enhances the feature perception capability. The feature extraction module SDASPP realizes edge enhancement, morphological perception and multi-scale perception, enabling the model to accurately detect targets in unstable slope areas of irregular shape and different size, thereby maintaining efficient and comprehensive detection capability in complex scenarios. The improved YOLOv8-FMC network model of this invention improves the sparse target detection accuracy. The convolutional block attention module CBAM realizes macroscopic and global feature judgment, enabling the model to more effectively capture sparse and scattered unstable slope area targets and improve detection accuracy.

[0025] (3) The improved YOLOv8-FMC network model of this invention enhances the detection performance under complex backgrounds. The convolutional block attention module CBAM is introduced into the target detection head to realize feature fusion and interaction between the predicted bounding box and the category, effectively suppressing background interference, thereby improving the model's ability to distinguish different types of slope instability phenomena in complex environments, and enhancing the recognition accuracy and robustness of various instability phenomena. Attached Figure Description

[0026] Figure 1 This is a flowchart of the target detection method for slope instability areas according to the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the structural principle of the improved YOLOv8-FMC network model in the embodiment;

[0028] Figure 3 This is a schematic diagram illustrating the improved structure and principle of the feature extraction module SDASPP in the embodiment.

[0029] Figure 4 This is a schematic diagram illustrating the structural principle of the deformable convolution DCNv1 in the embodiment;

[0030] Figure 5This is a schematic diagram illustrating the structural principle of the mask generation module MGB in the embodiment;

[0031] Figure 6 This is a schematic diagram illustrating the structural principle of the Convolutional Block Attention (CBAM) module in the embodiment.

[0032] Figure 7 for Figure 6 A schematic diagram illustrating the structural principle of the channel attention module inside the CBAM (Convolutional Block Attention Module).

[0033] Figure 8 for Figure 6 A schematic diagram illustrating the structural principle of the spatial attention module inside the convolutional block attention module (CBAM). Detailed Implementation

[0034] The present invention will be further described in detail below with reference to embodiments:

[0035] Example

[0036] like Figure 1 As shown, a target detection method for slope instability areas based on deep learning from remote sensing images is proposed, the method comprising:

[0037] S1. Construct a remote sensing image sample set, and label the slope instability areas of the remote sensing images in the remote sensing image sample set. In some embodiments, the slope instability areas of the remote sensing images in the remote sensing image sample set are labeled with slope instability boundary boxes using LabelMe software to generate corresponding JSON files for the remote sensing images. The JSON files are converted into TXT tags, and the TXT tags are used to create the corresponding label data for the remote sensing images (including whether there are slope instability areas and slope instability boundary boxes), and associated with the corresponding remote sensing images to construct the remote sensing image sample set.

[0038] S2. Construct an improved YOLOv8-FMC network model and input it into the remote sensing image sample set for model learning and training, such as... Figure 2 As shown, the YOLOv8-FMC network model includes a feature extraction core network Backbone (Backbone is an abbreviation for the module), a feature fusion enhancement network Neck (Neck is an abbreviation for the module), and an object detection head Detect (Detect is an abbreviation for the module). The feature extraction core network extracts multi-scale feature maps F1 to F3 layer by layer. Figure 2As shown, the core feature extraction network includes a first-layer feature extraction network, a second-layer feature extraction network, and a third-layer feature extraction network. Each of the first to third layers includes an improved feature extraction module SDASPP and a max-pooling module. The first-layer feature extraction network extracts features to obtain feature map F1. In this embodiment, taking a remote sensing image of B×C×H×W as an example, the feature extraction module SDASPP of the first-layer feature extraction network extracts multi-scale, shape, and size information. Then, it undergoes max-pooling (S=8) processing by the max-pooling module to obtain feature map F1 (also known as the first-layer feature map F1, B×C1×H / 8×W / 8), significantly expanding the receptive field while preserving detail. Feature map F1 is input into the second-layer feature extraction network to obtain feature map F2 (also known as the second-layer feature map F2, B×C2×H / 16×W / 16). Feature map F2 is input into the third-layer feature extraction network to obtain feature map F3 (also known as the third-layer feature map F2, B×C2×H / 32×W / 32). The core feature extraction network of this invention uses the first to third layers of the feature extraction network to perform layer-by-layer feature extraction, which can simultaneously take into account the detailed representation of small, medium and large targets and global contextual information at different levels, providing richer multi-scale feature representation for subsequent detection.

[0039] The experimental parameter settings for the improved YOLOv8-FMC network model are shown in Table 1, and the performance of the experimental server is shown in Table 2.

[0040] Table 1 Network Parameter Settings

[0041] method Parameter settings YOLOv8-FMC network model Model structure configuration file: yolov8.yaml, training and validation dataset configuration file: 1.yaml, number of training epochs: 100, batch size: 1, number of data loading threads (workers): 0, computing device: GPU or CPU available by default.

[0042] Table 2 Server Performance Configuration

[0043] category Configuration CPU Intel(R)Xeon(R)Gold5118CPU@2.30GHz graphics card NVIDIA GeForce RTX 2080 Ti RAM 32GB operating system Windows 10

[0044] An improved YOLOv8-FMC network model was trained using a remote sensing image sample set. The number of training epochs was set, and the model was continuously optimized through multiple iterations. In each training epoch, the improved YOLOv8-FMC network model used gradient descent and its optimizer to calculate the loss and automatically updated the weight parameters between the layers of the neural network to reduce classification and boundary regression losses and improve the model's detection accuracy. After multiple iterations, the improved YOLOv8-FMC network model with the highest accuracy and best visual effect was selected for the identification, segmentation, and output of slope instability boundary boxes from remote sensing images of the study area.

[0045] The feature fusion enhancement network uses the mask generation module MGB to generate CAM heatmaps of different scales from feature maps F1 to F3, and performs corresponding-scale multiplication, upsampling, and stepwise fusion processing on these heatmaps to obtain feature maps N1 to N3 of different scales. The structure of the mask generation module MGB is as follows: Figure 5 As shown, the mask generation module MGB processes the input feature map (including feature maps F1 to F3) through 1×1 convolution, full average pooling (calculating the mean of each channel), and a fully connected layer to obtain a CAM heatmap of the slope instability area features (the task of the mask generation module MGB is to highlight the most important feature regions and only determine the presence of slope instability areas).

[0046] In some embodiments, such as Figure 2 As shown, the feature fusion enhancement network multiplies feature map F3 with the CAM heatmap corresponding to feature map F3 and concatenates them to obtain feature map M1 (B×C1×H / 32×W / 32), and upsamples feature map M1; the feature fusion enhancement network multiplies feature map F2 with the CAM heatmap corresponding to feature map F2 and upsamples feature map M1 and concatenates them, and then processes them through the feature extraction module SDASPP to obtain feature map M2 (B×C2×H / 16×W / 16), and upsamples feature map M2; the feature fusion enhancement network multiplies feature map F1 with the CAM heatmap corresponding to feature map F1 and upsamples feature map M2 and concatenates them, and then processes them through the feature extraction module SDASPP to obtain feature map M3 (B×(C1+C2)×H / 8×W / 8), thus realizing multi-layer feature fusion and attention enhancement.

[0047] like Figure 2 As shown, the feature fusion enhancement network processes feature map M3 through the feature extraction module SDASPP to obtain feature map N1 (B×C1×H / 8×W / 8). Feature map N1 is then processed by dimensionality downsampling and channel adjustment CBS (preferably, the stride of dimensionality downsampling and channel adjustment CBS is S=2, and the convolution kernel is K=3) and concatenated with feature map M2. It is then processed again by the feature extraction module SDASPP to obtain feature map N2 (B×C2×H / 16×W / 16). Feature map N2 is then processed by dimensionality downsampling and channel adjustment CBS (preferably, the stride of dimensionality downsampling and channel adjustment CBS is S=2, and the convolution kernel is K=3) and concatenated with feature map M1. It is then processed again by the feature extraction module SDASPP to obtain feature map N3 (B×C2×H / 32×W / 32).

[0048] S3, the object detection head includes the Convolutional Block Attention (CBAM) module, such as... Figure 6As shown, the Convolutional Block Attention Module (CBAM) includes a channel attention module and a spatial attention module. The channel attention module extracts channel attention weights (B×C×1×1) from feature maps N1 to N3, weights them, and then inputs them into the spatial attention module to obtain spatial attention weights (B×1×H×W). Preferably, feature maps N1 to N3 are processed by parallel dimensionality downsampling and channel adjustment (CBS) (stride S=2, convolution kernel K=3) and then concatenated before being input into the CBAM of the target detection head. The convolutional module outputs the slope instability category and the slope instability bounding box. In some embodiments, such as Figure 7 As shown, the channel attention module obtains two feature maps through two parallel adaptive global max pooling layers and an adaptive global average pooling layer. These two feature maps are then input into a multilayer perceptron, and the output feature maps are summed and processed using a sigmoid activation function to generate channel attention weights for each channel. See also... Figure 8 The spatial attention module generates feature maps through two parallel adaptive max pooling layers and average pooling layers, which are then concatenated and convolved. Finally, the spatial attention weights are obtained by passing the sigmoid activation function.

[0049] S4. Input the remote sensing images of the slopes in the study area into the trained YOLOv8-FMC network model (selecting the model with the highest accuracy and best visual effect) and identify whether slope instability exists in the remote sensing images. In remote sensing images showing slope instability, identify and segment the slope instability bounding boxes and output them. Preferably, the target detection head uses a 1×1 convolutional module to predict and identify slope instability categories and output them; the target detection head uses another 1×1 convolutional module to identify, segment, and output the slope instability bounding boxes in remote sensing images showing slope instability. In some embodiments, the overall loss function expression of the YOLOv8-FMC network model is as follows:

[0050] ,in The overall loss of the model (composed of multiple parts) Classification loss for the target For bounding box regression loss, For distributed regression loss, The loss is handled by the mask generation module MGB.

[0051] All feature extraction modules SDASPP involved in this invention adopt an improved network structure, as follows: Figure 3As shown, the feature extraction module SDASPP comprises five branches. The first branch includes a 1×1 convolution module and a Sobel-DCNv1 module; the second branch includes a 3×3 convolution module with a dilation rate of 6 and a Sobel-DCNv1 module; the third branch includes a 3×3 convolution module with a dilation rate of 12 and a Sobel-DCNv1 module; the fourth branch includes a 3×3 convolution module with a dilation rate of 18 and a Sobel-DCNv1 module; and the fifth branch includes a global average pooling module and a Sobel-DCNv1 module. The Sobel-DCNv1 module includes parallel Sobel computation modules and deformable convolution DCNv1 (see [link to documentation]). Figure 3 (See the lower structural diagram). The Sobel computation module calculates the gradient magnitude G of the input feature map using the Sobel operator, outputs feature map T2, and concatenates it with the feature map T1 processed by deformable convolution DCNv1 (feature fusion is performed through 1×1 convolution). The expression for the gradient magnitude G is as follows: , This represents the gradient detected by the Sobel operator in the x-direction. This represents the gradient detected by the Sobel operator in the y-direction.

[0052] The feature extraction module SDASPP has five branches that extract features at different scales. The feature outputs from the five branches are then concatenated, convolved, and fused to finally output a feature map. SDASPP achieves comprehensive feature extraction by combining multi-scale information, edge features, and morphological awareness.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting a slope instability region target based on remote sensing image deep learning, characterized in that: The method comprises: S1, constructing a remote sensing image sample set, the slope instability region of the remote sensing image in the remote sensing image sample set is labeled; S2, inputting the remote sensing image sample set into the improved YOLOv8-FMC network model for model learning and training, the YOLOv8-FMC network model comprising a feature extraction core network, a feature fusion enhancement network and a target detection head, the feature extraction core network extracting multi-scale feature maps F1~F3 layer by layer; the feature fusion enhancement network uses a mask generation module MGB to generate CAM heat maps of different scales for the feature maps F1~F3 respectively, and performs corresponding scale multiplication, upsampling and step-by-step fusion processing on the feature maps F1~F3 to obtain feature maps N1~N3 of different scales; S3, the target detection head comprises a convolution block attention module CBAM, the convolution block attention module CBAM comprising a channel attention module and a spatial attention module, the channel attention module respectively extracting channel attention weights for the feature maps N1~N3 and inputting the channel attention weights into the spatial attention module to obtain spatial attention weights, and outputting a slope instability class and a slope instability bounding box through a convolution module; S4, inputting a remote sensing image of a study area into the trained YOLOv8-FMC network model and identifying whether the remote sensing image of the slope exists slope instability, and identifying and segmenting out a slope instability bounding box in the remote sensing image of the slope instability and outputting.

2. The method of claim 1, wherein the method is characterized in that: The slope instability region of the remote sensing image in the remote sensing image sample set is labeled by using a LabelMe software to generate a JSON file corresponding to the remote sensing image, the JSON file is converted into a TXT label, the TXT label is used to make a labeled label data corresponding to the remote sensing image, and the remote sensing image sample set is constructed by associating the labeled label data with the remote sensing image. 3.The slope instability region target detection method based on remote sensing image deep learning according to claim 1, characterized in that: The feature extraction core network comprises a first layer feature extraction network, a second layer feature extraction network and a third layer feature extraction network, the first to third layer feature extraction networks all comprising an improved feature extraction module SDASPP and a maximum pooling module, the first layer feature extraction network extracting features to obtain a feature map F1, the feature map F1 being input into the second layer feature extraction network to obtain a feature map F2, and the feature map F2 being input into the third layer feature extraction network to obtain a feature map F3.

4. The method of claim 3, wherein the method further comprises: The feature extraction module SDASPP is an improved network structure, and the feature extraction module SDASPP includes five branches, the first branch includes a 1*1 convolution module and a Sobel-DCNv1 module, the second branch includes a 3*3 convolution module with an expansion rate of 6 and a Sobel-DCNv1 module, the third branch includes a 3*3 convolution module with an expansion rate of 12 and a Sobel-DCNv1 module, the fourth branch includes a 3*3 convolution module with an expansion rate of 18 and a Sobel-DCNv1 module, and the fifth branch includes a global average pooling module and a Sobel-DCNv1 module; the Sobel-DCNv1 module includes a Sobel calculation module and a deformable convolution DCNv1 in parallel with each other, the Sobel calculation module calculates the gradient amplitude G of the input feature map through the Sobel operator to output a feature map T2 and splices and fuses the feature map T1 processed by the deformable convolution DCNv1; The five branches of the feature extraction module SDASPP perform feature extraction of different scales respectively, and the features output by the five branches are spliced, convolved and fused, and finally a feature map is output. 5.The slope instability region target detection method based on remote sensing image deep learning according to claim 1, characterized in that: The feature fusion enhancement network multiplies the feature map F3 and the CAM heat map corresponding to the feature map F3 and splices to obtain a feature map M1; The feature fusion enhancement network multiplies the feature map F2 and the CAM heat map corresponding to the feature map F2, splices the up-sampled feature map M1, and obtains a feature map M2 through feature extraction processing of the feature extraction module SDASPP; The feature fusion enhancement network multiplies the feature map F1 and the CAM heat map corresponding to the feature map F1, splices the up-sampled feature map M2, and obtains a feature map M3 through feature extraction processing of the feature extraction module SDASPP.

6. The method of claim 5, wherein the method further comprises: The feature fusion enhancement network obtains a feature map N1 through feature extraction processing of the feature extraction module SDASPP on the feature map M3; the feature map N1 is processed through dimension down-sampling and channel adjustment CBS and spliced with the feature map M2, and then a feature map N2 is obtained through processing of the feature extraction module SDASPP; the feature map N2 is processed through dimension down-sampling and channel adjustment CBS and spliced with the feature map M1, and then a feature map N3 is obtained through processing of the feature extraction module SDASPP. 7.The slope instability region target detection method based on remote sensing image deep learning according to claim 1, characterized in that: The channel attention module obtains two feature maps through two adaptive global maximum pooling layers and two adaptive global average pooling layers in parallel with each other, inputs the two feature maps into a multilayer perceptron respectively, adds the output feature maps, and generates a channel attention weight of each channel through a Sigmoid activation function; the spatial attention module generates feature maps through two adaptive maximum pooling layers and two average pooling layers in parallel with each other respectively, and then performs channel splicing and convolution operation, and finally obtains a spatial attention weight through a Sigmoid activation function. 8.The slope instability region target detection method based on remote sensing image deep learning according to claim 1 or 7, characterized in that: The feature maps N1 to N3 are input into the convolution block attention module CBAM of the target detection head after being processed through parallel dimension down-sampling and channel adjustment CBS and splicing. 9.The slope instability region target detection method based on remote sensing image deep learning according to claim 1 or 7, characterized in that: The target detection head performs slope instability category prediction, identification and output through a 1*1 convolution module; and the target detection head performs slope instability bounding box identification, segmentation and output on the remote sensing image with slope instability through another 1*1 convolution module. 10.The slope instability region target detection method based on remote sensing image deep learning according to claim 1, characterized in that: The overall loss function expression of the YOLOv8-FMC network model is as follows: wherein is a model overall loss, is a target classification loss, is a bounding box regression loss, is a distributed regression loss, is a mask generation block MGB processing loss.

Images