Landslide intelligent remote sensing identification method and system based on Fast-SCNN

Abstract

The invention discloses a Fast-SCNN-based landslide intelligent remote sensing identification method and system. The method comprises the steps of obtaining multi-source data including a high-resolution optical image and DEM data and performing preprocessing; delimiting a landslide-prone area as a region of interest, extracting a high-resolution optical image and DEM data of the region of interest, and constructing a landslide sample set; and constructing a Fast-SCNN-based intelligent landslide remote sensing recognition model, carrying out landslide recognition by using the trained Fast-SCNN-based intelligent landslide remote sensing recognition model, and carrying out post-processing optimization and verification modification on a recognition result. According to the invention, by fusing the high-resolution optical image and the digital elevation model DEM, intelligent landslide identification is carried out based on the Fast-SCNN semantic segmentation algorithm, and the method is suitable for large-range and high-efficiency automatic detection and accurate positioning of landslide disasters.

Description

A Smart Remote Sensing Identification Method and System for Landslides Based on Fast-SCNN Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology for geological disasters, specifically to an intelligent remote sensing identification method and system for landslides based on Fast-SCNN. Background Technology

[0002] Landslides and other surface geological hazards are key geological factors affecting the entire life cycle of railways in typical, complex, and rugged mountainous areas, creating a significant demand for remote sensing interpretation. In recent years, deep learning technology has significantly improved the effectiveness of image feature extraction, but three technical challenges remain:

[0003] First, there are limitations to traditional methods. They rely on manual visual interpretation of remote sensing images, which is inefficient and highly subjective. The interpretation of a single large-scale image can take several days, making it difficult to meet the emergency monitoring needs of sudden geological disasters.

[0004] Secondly, there are bottlenecks in the application of deep learning. Existing landslide recognition models (such as DeeplabV3+ and U-Net) mostly adopt complex encoding and decoding structures, which have high computational overhead and slow inference speed, making them difficult to deploy on edge devices. In addition, there is a lack of high-quality labeled datasets, and the models are not accurate enough in recognizing landslides in complex terrains (such as steep slopes and vegetated areas).

[0005] Thirdly, there is a need to balance multi-source data fusion and efficiency. Single optical images are easily affected by clouds, fog, and vegetation. It is necessary to combine DEM terrain information to improve recognition robustness. However, existing fusion models often neglect efficiency optimization and cannot balance accuracy and speed.

[0006] Meanwhile, satellite and drone aerial photography technologies such as Sentinel-2 (10-meter resolution) and WorldView-3 (0.3-meter resolution) can quickly acquire high-resolution optical images; public data sources such as SRTM and ASTER GDEM provide global DEM data, laying the foundation for multi-source data fusion.

[0007] Fast-SCNN, with its lightweight architecture of "fast downsampling-feature fusion-fast upsampling", maintains high segmentation accuracy while improving inference speed by 5-10 times compared to traditional models. It has been applied in remote sensing tasks such as road extraction and building recognition, but customized optimizations for landslide disasters (such as terrain feature fusion and small sample training) have not yet been reported. Summary of the Invention

[0008] This invention provides a landslide intelligent remote sensing identification method and system based on Fast-SCNN, which integrates high-resolution optical images and digital elevation models (DEMs) and performs intelligent landslide identification based on the Fast-SCNN semantic segmentation algorithm to solve the above-mentioned problems in the prior art. It is suitable for large-scale, high-efficiency automated detection and precise location of landslide disasters, and is especially suitable for real-time monitoring needs in resource-constrained scenarios.

[0009] According to the first aspect, one embodiment provides a landslide intelligent remote sensing identification method based on Fast-SCNN, the method comprising:

[0010] Acquire multi-source data, including high-resolution optical images and DEM data, and perform preprocessing;

[0011] Landslide-prone areas were designated as regions of interest, and high-resolution optical images and DEM data of these regions were extracted to construct a landslide sample set, which was then divided into a training set, a validation set, and a test set.

[0012] A landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM.

[0013] The training set, validation set, and test set were used to train, validate, and test the landslide intelligent remote sensing recognition model based on Fast-SCNN.

[0014] A landslide intelligent remote sensing identification model based on Fast-SCNN, trained and obtained, was used for landslide identification. The identification results were then post-processed, optimized, and verified.

[0015] Furthermore, multi-source data, including high-resolution optical images and DEM data, are acquired and preprocessed, specifically including:

[0016] The high-resolution optical images were acquired by satellite remote sensing or drone aerial photography, with a spatial resolution of 0.3-10 meters, including red, green and blue channels, a unified coordinate system of WGS84, and a geometric correction error of ≤1 pixel.

[0017] The DEM data uses SRTM 1 arcsecond or ASTER GDEM 2 publicly available data. Noise is removed by Gaussian filtering, and three types of derived topographic factors are generated: slope, aspect, and topographic relief. A 4-channel DEM feature set containing the original elevation plus the three types of derived factors is constructed and resampled to the same resolution as the high-resolution optical image.

[0018] Furthermore, landslide-prone areas were designated as regions of interest (ROIs), and high-resolution optical images and DEM data of these ROIs were extracted to construct a landslide sample set, which specifically includes:

[0019] Based on historical landslide logging data and geological hazard-prone area distribution maps, landslide-prone areas are designated as Regions of Interest (ROIs) within the target area to ensure that the sample includes landslide instances of different scales and topographic conditions.

[0020] The landslide boundary was manually marked and drawn to generate a single-channel binary mask. The landslide area was marked as 1 and the background was marked as 0. The consistency of the marking was verified by the Kappa coefficient.

[0021] The 3-channel optical image of the ROI region is superimposed with the 4-channel DEM feature set to form 7-channel data. The data is then segmented according to a preset pixel size, and the corresponding binary mask is extracted simultaneously to construct a sample pair of "7-channel data + binary mask" to obtain the landslide sample set.

[0022] The landslide sample set is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is augmented with random flipping, scaling, and Gaussian blur to alleviate the sample imbalance problem.

[0023] Furthermore, a landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. This model includes a fast feature extraction module (FFEN), a simplified pyramid pooling module (S-PPM), and a fast upsampling module (FUM), specifically comprising:

[0024] The fast feature extraction module FFEN includes three convolutional blocks with kernel sizes of 3×3, 5×5, and 3×3, and a stride of 2. It is used to compress 7-channel input data into 64-channel feature maps with a downsampling rate of 8 times.

[0025] Furthermore, a landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. This model includes a fast feature extraction module (FFEN), a simplified pyramid pooling module (S-PPM), and a fast upsampling module (FUM), specifically comprising:

[0026] The simplified pyramid pooling module S-PPM includes two parallel branches: 1×1 convolution and 3×3 convolution, which are used to perform multi-scale feature aggregation on the feature map output by FFEN to enhance the perception of landslides of different scales; and introduces a DEM feature attention mechanism to adaptively adjust the weight ratio of DEM and optical image features through 1×1 convolution.

[0027] Furthermore, a landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. This model includes a fast feature extraction module (FFEN), a simplified pyramid pooling module (S-PPM), and a fast upsampling module (FUM), specifically comprising:

[0028] The fast upsampling module FUM uses a combination of transposed convolution and pixel shuffle to upsample the feature map to the original resolution.

[0029] Furthermore, the training set, validation set, and test set are used to train, validate, and test the landslide intelligent remote sensing recognition model based on Fast-SCNN, specifically including:

[0030] The model training uses a hybrid loss function, which includes cross-entropy loss and Dice loss, with a cross-entropy loss coefficient of 0.7 and a Dice loss coefficient of 0.3.

[0031] The AdamW optimizer was used with an initial learning rate of 0.001. A cosine annealing strategy was employed, and training was iterated for 100 epochs. When the intersection and union of the validation set showed no improvement for 10 consecutive epochs, an early stopping strategy was used to terminate the training.

[0032] By using transfer learning to load pre-trained weights from publicly available remote sensing images, freezing the parameters of the first two convolutional blocks of FFEN, and training only subsequent modules, the model's generalization ability in small sample scenarios is improved.

[0033] Furthermore, the landslide intelligent remote sensing identification model based on Fast-SCNN, trained for this purpose, is used for landslide identification. The identification results are then post-processed for optimization and verification, specifically including:

[0034] The preprocessed 7-channel data of the target area is input, and a landslide probability map is generated by the trained landslide intelligent remote sensing recognition model based on Fast-SCNN. Adaptive threshold segmentation is used to obtain the binarized recognition result.

[0035] Further post-processing optimization is performed, with morphological operations used to remove isolated noise points, and connected component analysis used to merge broken landslide patches, outputting a preliminary landslide extent vector map.

[0036] Finally, geological experts verified and modified the model output results, supplementing missed landslides, deleting misjudged areas, adjusting boundary positions by circling with the mouse, recording modification logs, and finally outputting high-precision landslide range results.

[0037] According to a second aspect, one embodiment provides a landslide intelligent remote sensing identification system based on Fast-SCNN, the system comprising:

[0038] The multi-source data acquisition and preprocessing module is used to acquire and preprocess multi-source data, including high-resolution optical images and DEM data.

[0039] The landslide sample set construction module is used to delineate landslide-prone areas as regions of interest, extract high-resolution optical images and DEM data of the regions of interest, construct the landslide sample set, and divide it into training set, validation set and test set;

[0040] The Fast-SCNN model building module is used to build a landslide intelligent remote sensing identification model based on Fast-SCNN. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM.

[0041] The Fast-SCNN model training and optimization module is used to train, validate, and test the landslide intelligent remote sensing identification model based on Fast-SCNN using the training set, validation set, and test set.

[0042] The landslide reasoning and identification and result verification module is used to identify landslides using the trained Fast-SCNN-based intelligent remote sensing identification model, and to perform post-processing optimization and verification modifications on the identification results.

[0043] According to a third aspect, one embodiment provides an electronic device, the device comprising: a processor and a memory;

[0044] The memory is used to store one or more program instructions;

[0045] The processor is configured to run one or more program instructions to perform the steps of a landslide intelligent remote sensing identification method based on Fast-SCNN as described in any of the preceding claims.

[0046] According to a fourth aspect, one embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a landslide intelligent remote sensing identification method based on Fast-SCNN as described in any of the preceding claims.

[0047] This invention provides a landslide intelligent remote sensing identification method and system based on Fast-SCNN, which has the following beneficial effects:

[0048] (1) Efficiency improvement: The inference speed of the Fast-SCNN model is more than 8 times faster than that of the traditional DeeplabV3+. The inference time for a single 10km×10km image (10m resolution) is ≤30 seconds (based on NVIDIA RTX 4090), and it supports batch processing of large areas.

[0049] (2) Accuracy optimization: Through DEM feature fusion and attention mechanism, the landslide identification mIoU ≥ 82% and the boundary extraction error ≤ 3 pixels. The accuracy of identifying vegetation-covered areas and small landslides is improved by 15%-20% compared with the single optical image model.

[0050] (3) High practicality: The lightweight architecture can be deployed on edge computing devices (such as drone onboard computers) and can support rapid identification in the field. Attached Figure Description

[0051] Figure 1 is a flowchart of a landslide intelligent remote sensing identification method based on Fast-SCNN provided in an embodiment of the present invention;

[0052] Figure 2 is an optimized Fast-SCNN model architecture diagram in a landslide intelligent remote sensing identification method based on Fast-SCNN provided in an embodiment of the present invention;

[0053] Figure 3 is a comparison of the recognition results of different models in a landslide intelligent remote sensing recognition method based on Fast-SCNN provided in an embodiment of the present invention (the dataset is the publicly available Bijie landslide dataset). Detailed Implementation

[0054] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0055] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0056] The first embodiment of this invention provides a landslide intelligent remote sensing identification method based on Fast-SCNN. Through a standardized multi-source data processing procedure, customized sample set construction, and an optimized Fast-SCNN model, it achieves rapid and high-precision automatic identification of landslide areas, meeting the real-time and accuracy requirements of large-scale landslide monitoring. A detailed description is provided below with reference to Figure 1.

[0057] As shown in Figure 1, in step S100, multi-source data including high-resolution optical images and DEM data are acquired and preprocessed.

[0058] The above steps specifically include:

[0059] S110: High-resolution optical images are acquired through satellite remote sensing (Sentinel-2, WorldView-3) or drone aerial photography, with a spatial resolution of 0.3-10 meters. They include three channels: red, green, and blue (RGB). Geometric correction (error ≤ 1 pixel), radiometric calibration, atmospheric correction, and cloud removal are performed sequentially, and the unified coordinate system is WGS84.

[0060] S120: DEM data uses publicly available SRTM 1 arcsecond (approximately 30 meters) or ASTER GDEM 2 (approximately 15 meters) data. Noise is removed by Gaussian filtering (standard deviation σ=1.5), and three types of derived topographic factors are generated: slope, aspect, and topographic relief. A DEM feature set (4 channels: original elevation + 3 types of derived factors) is constructed and resampled to the same resolution as the high-resolution optical image.

[0061] As shown in Figure 1, in step S200, the landslide-prone area is defined as the region of interest, high-resolution optical images and DEM data of the region of interest are extracted, a landslide sample set is constructed, and it is divided into a training set, a validation set and a test set.

[0062] The above steps specifically include:

[0063] S210: Based on historical landslide logging data and geological hazard susceptibility distribution maps, landslide-prone areas are delineated as regions of interest (ROIs) within the study area to ensure that the sample includes landslide instances of different sizes (small: <10,000 m², medium: 10,000-100,000 m², large: >100,000 m²) and different topographic conditions (bare rock areas, vegetated areas, residential areas).

[0064] S220: Manual annotation is performed by two or more senior geological engineers using professional software such as ENVI and ArcGIS to manually delineate the landslide boundary and generate a single-channel binary mask (the landslide area is marked as 1 and the background is marked as 0). The consistency of the annotation is verified by the Kappa coefficient (Kappa≥0.85).

[0065] S230: The optical image (3 channels) of the ROI region is superimposed with the DEM feature set (4 channels) to form 7-channel data, which is then segmented into 512×512 pixel sizes. The corresponding binary mask is extracted simultaneously to construct a sample pair of "7-channel data + binary mask". The sample is divided into training set, validation set and test set in a 7:2:1 ratio. The training set is augmented by random flipping (horizontal / vertical), scaling (0.8-1.2 times), and Gaussian blur (σ=0-0.5) to alleviate the sample imbalance problem.

[0066] As shown in Figure 1, in step S300, a landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM.

[0067] In this embodiment, as shown in Figure 2, the Fast-SCNN network architecture includes the following three core modules:

[0068] ① Fast Feature Extraction Network (FFEN): It uses 3 convolutional blocks (Conv-BN-ReLU) to achieve fast downsampling. The kernel sizes are 3×3, 5×5 and 3×3 respectively, and the stride is 2. It compresses the input 7-channel data into a 64-channel feature map, with a downsampling rate of 8 times and reduced computational cost.

[0069] ② Simplified Pyramid Pooling Module (S-PPM): Two parallel branches (1×1 convolution + 3×3 convolution) are designed to perform multi-scale feature aggregation on the FFEN output feature map, enhancing the ability to perceive landslides of different scales; a DEM feature attention mechanism is introduced, which adjusts the weight ratio of DEM and optical image features through 1×1 convolution (initial weight ratio 1:1, adaptively optimized during training).

[0070] ③ Fast Upsampling Module (FUM): It uses a combination of transposed convolution (stride 2) and pixel shuffle to upsample the feature map to the original resolution, avoiding the boundary blurring caused by traditional interpolation.

[0071] As shown in Figure 1, in step S400, the training set, validation set, and test set are used to train, validate, and test the landslide intelligent remote sensing recognition model based on Fast-SCNN.

[0072] The above steps specifically include:

[0073] S410: The model training uses a hybrid loss function (cross-entropy loss coefficient of 0.7 and Dice loss coefficient of 0.3) to alleviate the training bias caused by the low proportion of landslide samples (usually < 5%).

[0074] S420: Use the AdamW optimizer (weight decay coefficient 0.001), initial learning rate 0.001, cosine annealing strategy (T_max=50), iterate for 100 epochs, and terminate training by early stopping strategy when the intersection-union ratio (mIoU) of the validation set has no improvement for 10 consecutive epochs.

[0075] S430: Employs transfer learning to load pre-trained weights from publicly available remote sensing images (initialized based on the publicly available ImageNet-R dataset), freezes the parameters of the first two convolutional blocks of FFEN, and trains only subsequent modules to improve the model's generalization ability in small sample scenarios.

[0076] As shown in Figure 1, in step S500, the landslide intelligent remote sensing identification model based on Fast-SCNN, which has been trained, is used to identify landslides, and the identification results are post-processed, optimized, and verified.

[0077] The above steps specifically include:

[0078] S510: First, perform automated inference. Input the preprocessed 7-channel data (optical image + DEM feature set) of the study area. Generate a landslide probability map (pixel value 0-1) through the trained Fast-SCNN model. Use adaptive threshold segmentation (automatically determine the threshold based on the Otsu algorithm) to obtain the binarized recognition result.

[0079] S520: Further post-processing optimization is performed. Isolated noise points (area < 50 pixels) are removed through morphological operations (erosion followed by dilation, kernel size 3×3). Connectivity analysis is used to merge broken landslide patches and output a preliminary landslide range vector map.

[0080] S530: Finally, geological experts verify and modify the model output results. They use the mouse to circle and draw to supplement missed landslides, delete misjudged areas, adjust boundary positions, record modification logs, and finally output high-precision landslide range results.

[0081] The comparison of recognition results of different models is shown in Table 1 and Figure 3.

[0082] Table 1 Comparison of recognition results of different models

[0083]

[0084] Corresponding to the aforementioned disclosed method for intelligent remote sensing identification of landslides based on Fast-SCNN, this invention also discloses an intelligent remote sensing identification system for landslides based on Fast-SCNN, which specifically includes:

[0085] The multi-source data acquisition and preprocessing module is used to acquire and preprocess multi-source data, including high-resolution optical images and DEM data.

[0086] The landslide sample set construction module is used to delineate landslide-prone areas as regions of interest, extract high-resolution optical images and DEM data of the regions of interest, construct the landslide sample set, and divide it into training set, validation set and test set;

[0087] The Fast-SCNN model building module is used to build a landslide intelligent remote sensing identification model based on Fast-SCNN. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM.

[0088] The Fast-SCNN model training and optimization module is used to train, validate, and test the landslide intelligent remote sensing identification model based on Fast-SCNN using the training set, validation set, and test set.

[0089] The landslide reasoning and identification and result verification module is used to identify landslides using the trained Fast-SCNN-based intelligent remote sensing identification model, and to perform post-processing optimization and verification modifications on the identification results.

[0090] It should be noted that for a detailed description of the landslide intelligent remote sensing identification system based on Fast-SCNN provided in the embodiments of the present invention, please refer to the relevant description of the landslide intelligent remote sensing identification method based on Fast-SCNN provided in the embodiments of the present invention, which will not be repeated here.

[0091] In addition, embodiments of the present invention also provide an electronic device, the device comprising: a processor and a memory; the memory being used to store one or more program instructions; the processor being used to run one or more program instructions to perform the steps of a landslide intelligent remote sensing identification method based on Fast-SCNN as described in any of the preceding embodiments.

[0092] It should be noted that for a detailed description of the electronic device provided in the embodiments of the present invention, please refer to the relevant description of the landslide intelligent remote sensing identification method based on Fast-SCNN provided in the embodiments of the present invention, which will not be repeated here.

[0093] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the landslide intelligent remote sensing identification method based on Fast-SCNN as described in any of the preceding claims.

[0094] It should be noted that for a detailed description of the computer-readable storage medium provided in the embodiments of the present invention, please refer to the relevant description of the landslide intelligent remote sensing identification method based on Fast-SCNN provided in the embodiments of the present invention, which will not be repeated here.

[0095] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0096] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A landslide intelligent remote sensing identification method based on Fast-SCNN, characterized in that, The method includes: acquiring multi-source data including high-resolution optical images and DEM data and preprocessing them; delineating landslide-prone areas as regions of interest, extracting high-resolution optical images and DEM data from the regions of interest, constructing a landslide sample set, and dividing it into a training set, a validation set, and a test set; constructing a landslide intelligent remote sensing identification model based on Fast-SCNN, the model including a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM; training, validating, and testing the landslide intelligent remote sensing identification model based on Fast-SCNN using the training set, validation set, and test set; using the trained landslide intelligent remote sensing identification model based on Fast-SCNN to identify landslides, and performing post-processing optimization and verification modifications on the identification results.

2. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 1, characterized in that, The process involves acquiring and preprocessing multi-source data, including high-resolution optical imagery and DEM data. Specifically, the high-resolution optical imagery is acquired via satellite remote sensing or UAV aerial photography, with a spatial resolution of 0.3-10 meters, containing red, green, and blue channels, using a unified coordinate system of WGS84, and a geometric correction error of ≤1 pixel. The DEM data uses publicly available SRTM 1 arcsecond or ASTER GDEM 2 data, removes noise through Gaussian filtering, generates three types of derived topographic factors (slope, aspect, and topographic relief), constructs a 4-channel DEM feature set containing the original elevation plus the three types of derived factors, and resamples it to the same resolution as the high-resolution optical imagery.

3. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 1, characterized in that, Landslide-prone areas were designated as Regions of Interest (ROIs). High-resolution optical images and DEM data of the ROIs were extracted to construct a landslide sample set. Specifically, this involved: defining landslide-prone areas as ROIs within the target area based on historical landslide logging data and geological hazard-prone area distribution maps, ensuring that the samples included landslide instances of different scales and terrain conditions; manually marking and delineating landslide boundaries to generate single-channel binary masks, with landslide areas marked as 1 and background as 0, and verifying the consistency of the markings using the Kappa coefficient; superimposing the 3-channel optical images of the ROIs with the 4-channel DEM feature set to obtain 7-channel data, dividing it according to a preset pixel size, and simultaneously extracting the corresponding binary masks to construct "7-channel data + binary mask" sample pairs, thus obtaining the landslide sample set; and dividing the landslide sample set into training, validation, and test sets according to a preset ratio. The training set underwent data augmentation through random flipping, scaling, and Gaussian blurring to alleviate the sample imbalance problem.

4. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 1, characterized in that, A landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM. Specifically, the fast feature extraction module FFEN includes three convolutional blocks with kernel sizes of 3×3, 5×5, and 3×3, and a stride of 2. It is used to compress 7-channel input data into 64-channel feature maps, achieving a downsampling rate of 8 times.

5. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 4, characterized in that, A landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM. Specifically, the simplified pyramid pooling module S-PPM includes two parallel branches: 1×1 convolution and 3×3 convolution, which are used to perform multi-scale feature aggregation on the feature map output by FFEN to enhance the perception capability of landslides of different scales. A DEM feature attention mechanism is introduced, which adaptively adjusts the weight ratio of DEM and optical image features through 1×1 convolution.

6. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 5, characterized in that, A landslide intelligent remote sensing identification model based on Fast-SCNN is constructed. The model includes a fast feature extraction module FFEN, a simplified pyramid pooling module S-PPM, and a fast upsampling module FUM. Specifically, the fast upsampling module FUM uses a combination of transposed convolution and pixel shuffle to upsample the feature map to the original resolution.

7. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 1, characterized in that, The Fast-SCNN-based intelligent remote sensing model for landslide identification was trained, validated, and tested using the training, validation, and test sets. Specifically, the training employed a hybrid loss function, including cross-entropy loss and Dice loss, with a cross-entropy loss coefficient of 0.7 and a Dice loss coefficient of 0.

3. The AdamW optimizer was used with an initial learning rate of 0.001, employing a cosine annealing strategy for 100 epochs. Training was terminated early when the cross-union ratio of the validation set showed no improvement for 10 consecutive epochs. Transfer learning was used to load pre-trained weights from publicly available remote sensing images, freezing the parameters of the first two convolutional blocks of FFEN and training only subsequent modules to improve the model's generalization ability in small-sample scenarios.

8. The landslide intelligent remote sensing identification method based on Fast-SCNN as described in claim 3, characterized in that, Landslide identification was performed using a trained Fast-SCNN-based intelligent remote sensing model. The identification results were then post-processed, optimized, and validated. Specifically, this involved: inputting pre-processed 7-channel data of the target area; generating a landslide probability map using the trained Fast-SCNN-based model; and obtaining binarized identification results through adaptive threshold segmentation. Post-processing optimization was then performed, including morphological operations to remove isolated noise points, and connected component analysis to merge fragmented landslide patches, resulting in a preliminary landslide extent vector map. Finally, geological experts validated and modified the model's output, using mouse drawing to supplement missed landslides, delete misidentified areas, adjust boundary positions, and record modification logs. The final result was a high-precision landslide extent assessment.

9. A landslide intelligent remote sensing identification system based on Fast-SCNN, characterized in that, The system includes: a multi-source data acquisition and preprocessing module for acquiring and preprocessing multi-source data, including high-resolution optical images and DEM data; a landslide sample set construction module for delineating landslide-prone areas as regions of interest, extracting high-resolution optical images and DEM data from these regions of interest, constructing a landslide sample set, and dividing it into a training set, a validation set, and a test set; a Fast-SCNN model construction module for constructing a Fast-SCNN-based intelligent remote sensing landslide recognition model, which includes a Fast Feature Extraction (FFEN) module, a Simplified Pyramid Pooling (S-PPM) module, and a Fast Upsampling (FUM) module; a Fast-SCNN model training and optimization module for training, validating, and testing the Fast-SCNN-based intelligent remote sensing landslide recognition model using the training set, validation set, and test set; and a landslide inference, recognition, and result verification module for using the trained Fast-SCNN-based intelligent remote sensing landslide recognition model to identify landslides and performing post-processing optimization and verification modifications on the recognition results.

10. An electronic device, characterized in that, The device includes: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to perform the steps of a landslide intelligent remote sensing identification method based on Fast-SCNN as described in any one of claims 1 to 8.

Images