Geological disaster automatic identification method based on multi-source data and deep learning

Abstract

The invention relates to a geological disaster automatic identification method based on multi-source data and deep learning. The method is based on multi-source remote sensing data and a deep learning technology, various remote sensing data sources can be fully utilized, an advanced deep learning algorithm is combined, and accurate and efficient recognition of geological disasters is achieved. According to the improved UNet + + semantic segmentation model, the capabilities of UNet + + and Visual Mama are cooperatively used for a new architecture of image segmentation, and the improved UNet + + semantic segmentation model after multi-source data training is adopted to realize rapid and accurate identification of a large-range landslide; by integrating multi-source remote sensing information, the information limitation of a single data source is overcome, the accuracy and efficiency of geological disaster recognition are improved, and powerful technical support is provided for geological disaster prevention and emergency response.

Description

Technical Field

[0001] The present application relates to the technical field of geological disaster identification, and in particular to a method for automatic identification of geological disasters based on multi-source data and deep learning. Background Art

[0002] As global climate change and environmental problems become increasingly severe, small and medium-scale severe convective weather and the geological disasters it causes (such as landslides, mud-rock flows, ground collapse, etc.) not only pose a huge threat to people's lives and property safety, but also have a serious impact on the social economy. Therefore, the monitoring and prevention of geological disasters has become a hot issue of global concern.

[0003] There are many mountains and hills in a certain area, and geological activities are frequent. In addition, the typical subtropical monsoon climate, concentrated rainfall in summer, and abundant surface water, these factors together easily induce geological disasters. Especially after severe convective weather in summer, natural disasters such as landslides and mud-rock flows occur frequently. Therefore, it is particularly important to develop an efficient and accurate automatic identification method for geological disasters.

[0004] Traditional geological disaster monitoring methods mainly rely on on-site investigations and manual analysis, which is not only time-consuming and labor-intensive, but also difficult to achieve real-time and dynamic monitoring of disasters. Most of the existing deep learning-based geological disaster identification methods are based on a single data source. Due to the limitations of a single data source in geological disaster monitoring, the accuracy of geological disaster identification is not high.

[0005] UNet++ has been applied in the semantic segmentation of remote sensing data and has achieved some results, but UNet++ still has some shortcomings. For example, although UNet++ enhances feature fusion through nested skip connections, it is not enough to fully capture features of different scales when dealing with complex scenes, and its multi-scale feature fusion capability is insufficient. UNet++ mainly relies on convolutional layers for feature extraction, which may be insufficient when dealing with certain complex textures and boundaries. In addition, the complex structure of UNet++ increases computational complexity and memory consumption, which may lead to slower training and inference speeds. Summary of the invention

[0006] Based on this, it is necessary to provide an automatic identification method for geological hazards based on multi-source data and deep learning to address the above technical problems.

[0007] A method for automatic identification of geological hazards based on multi-source data and deep learning, the method comprising:

[0008] Acquire multi-source data in a preset area, wherein the multi-source data includes: optical data and multi-source remote sensing data;

[0009] Preprocessing the multi-source data, and fusing the preprocessed multi-source data to obtain a multi-band image;

[0010] Determine the landslide boundary according to the multi-band image and generate a landslide sample;

[0011] The improved UNet++ semantic segmentation model is trained by using the landslide sample to obtain a trained improved UNet++ semantic segmentation model; the improvement of the improved UNet++ semantic segmentation model is in the encoder part, a VSS module based on the Visual Mamba structure is used to replace the second to fourth layer encoding modules, and a bottleneck layer is added after the fourth layer encoding module, and the bottleneck layer is composed of two VSS modules;

[0012] After preprocessing and data fusion, the multi-source data of the area to be identified are input into the trained improved UNet++ semantic segmentation model to obtain the address disaster identification results.

[0013] In one embodiment, the multi-source remote sensing data includes vegetation cover NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data.

[0014] In one embodiment, the multi-source data is preprocessed, and the preprocessed multi-source data is fused to obtain a multi-band image, including:

[0015] Performing radiometric calibration, atmospheric correction, registration, cropping, and panchromatic multispectral fusion on the optical data to generate a fused image with both high spatial resolution and high spectral resolution;

[0016] After cropping the vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data, the data is resampled using the nearest neighbor interpolation method; wherein the resolution of the resampling result is consistent with the resolution of the fused image;

[0017] The fused image and the obtained resampling result are synthesized into a multi-band image containing all multi-source data.

[0018] In one embodiment, determining the landslide boundary according to the multi-band image and generating the landslide sample includes:

[0019] Use remote sensing image interpretation software to visually interpret remote sensing images; outline the boundaries of the landslide based on the characteristics of the landslide on the image;

[0020] The delineated landslide boundary is converted from a surface feature into a raster image to obtain a landslide image containing landslide samples;

[0021] Cutting the multi-band image and the landslide image containing the landslide sample into sub-images of a preset size;

[0022] The landslide samples in the sub-images containing outliers are removed, and the sub-images whose landslide pixel ratio is less than the preset value are deleted to obtain the landslide sample library.

[0023] In one embodiment, the landslide sample includes: a training set and a test set;

[0024] The improved UNet++ semantic segmentation model is trained using the landslide sample to obtain a trained improved UNet++ semantic segmentation model, including:

[0025] The samples in the test set are input into the encoder of the improved UNet++ semantic segmentation model. In the first stage, the input samples are feature extracted by two 3×3 convolution modules to obtain the first intermediate feature. In the second stage, the first intermediate feature is feature extracted by downsampling and three 3×3 convolution modules to obtain the second intermediate feature. In the third stage, the second intermediate feature is processed by a linear latent layer after being separated and divided, and the processed result is subjected to feature learning through two consecutive VSS blocks to obtain the third intermediate feature. In the fourth stage, the third intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks to obtain the fourth intermediate feature. In the fifth stage, the fourth intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks, and then the obtained feature is processed by a bottleneck layer to obtain the fifth intermediate feature.

[0026] The first intermediate feature, the second intermediate feature, the third intermediate feature, the fourth intermediate feature and the fifth intermediate feature are input into the decoder of the improved UNet++ semantic segmentation model, and the resolution of the features of each stage is restored by using a nested decoder based on a dense skip connection structure. Finally, the features of each stage are weightedly fused to obtain the geological disaster identification prediction result;

[0027] According to the geological disaster identification prediction results and the training set, the Adam optimizer is used to update and optimize the network parameters to obtain a trained improved UNet++ semantic segmentation model;

[0028] The verified improved UNet++ semantic segmentation model is tested using the test set, and the accuracy index of model recognition is calculated based on the test results;

[0029] Change different band combinations for training, calculate the accuracy indicators of different band combinations, and test the influencing factors and the effects of different band combinations;

[0030] The generalization of the improved UNet++ semantic segmentation model to different regions and multi-source data is tested to obtain the trained improved UNet++ semantic segmentation model.

[0031] In one embodiment, different band combinations are changed for training, accuracy indicators of different band combinations are calculated, and influencing factors and effects of different band combinations are tested, including:

[0032] Prepare different band combination data sets, use each band combination to train the improved UNet++ semantic segmentation model, evaluate the performance of the trained model on the test set, calculate and compare the accuracy indicators of different band combinations, evaluate the impact of different band combinations on model performance, and select the optimal band combination for subsequent model training and geological disaster identification tasks based on the evaluation results.

[0033] The above-mentioned automatic identification method of geological disasters based on multi-source data and deep learning is based on multi-source remote sensing data and deep learning technology, which can make full use of various remote sensing data sources and combine advanced deep learning algorithms to achieve accurate and efficient identification of geological disasters. The improved UNet++ semantic segmentation model uses the capabilities of UNet++ and Visual Mamba in a new architecture for image segmentation, and uses the improved UNet++ semantic segmentation model trained with multi-source data to achieve rapid and accurate identification of large-scale landslides; by integrating multi-source remote sensing information, it not only overcomes the information limitations of a single data source, but also improves the accuracy and efficiency of geological disaster identification, providing strong technical support for geological disaster prevention and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of a flow chart of a method for automatic identification of geological hazards based on multi-source data and deep learning in one embodiment;

[0035] Figure 2 A schematic diagram of an improved UNet++ semantic segmentation model in another embodiment. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0037] In one embodiment, Figure 1 As shown, a method for automatic identification of geological hazards based on multi-source data and deep learning is provided, and the method comprises the following steps:

[0038] Step 100: Acquire multi-source data in a preset area, wherein the multi-source data includes: optical data and multi-source remote sensing data;

[0039] Step 102: preprocessing the multi-source data, and fusing the preprocessed multi-source data to obtain a multi-band image;

[0040] Step 104: Determine the landslide boundary according to the multi-band image and generate a landslide sample.

[0041] Specifically, the landslide boundary is delineated, landslide samples are generated, the images are cut into sub-images, samples are screened, and a landslide sample library is created.

[0042] Step 106: using the landslide sample to train the improved UNet++ semantic segmentation model to obtain a trained improved UNet++ semantic segmentation model; the improvement of the improved UNet++ semantic segmentation model is in the encoder part, using a VSS module based on the Visual Mamba structure to replace the second to fourth layer encoding modules, and adding a bottleneck layer after the fourth layer encoding module, wherein the bottleneck layer is composed of two VSS modules;

[0043] Specifically, the improved UNet++ semantic segmentation model is a new architecture that uses the capabilities of UNet++ and Visual Mamba for image segmentation. The improved UNet++ model is used to train a landslide recognition model for multi-source data to achieve rapid and accurate recognition of large-scale landslides.

[0044] The improved UNet++ semantic segmentation model includes an encoder and a decoder. The VSS module of Visual Mamba enhances the feature extraction and fusion capabilities, so the second to fourth encoding modules and the bottleneck layer in the encoder in this model are built based on the Visual Mamba block. The encoder and decoder are connected using a dense skip connection (DSC) structure; the nested skip connection and multi-scale feature fusion structure of UNet++ are more effective in feature fusion and detail processing methods, so a nested decoder based on DSC is used to restore the resolution of features at each stage. Finally, the features of each stage are weightedly fused to obtain the final result. This improvement not only retains the advantages of UNet++ in feature fusion, but also utilizes the multi-scale feature extraction capabilities of the VSS module.

[0045] The improved UNet++ semantic segmentation model in this application only uses VSS modules in the encoder and bottleneck layers, retaining the decoder part of UNet++, making the model more flexible in structure. Such a design can better adapt to different task requirements and achieve higher task adaptability by adjusting some module parameters.

[0046] The advantages of the improved UNet++ semantic segmentation model in feature fusion, computational efficiency and flexibility make the improved UNet++ semantic segmentation model in this application more suitable for solving the technical problems in this application and can better achieve the expected beneficial effects.

[0047] The improved UNet++ semantic segmentation model is as follows Figure 2 shown.

[0048] Visual Mamba is a newly proposed deep learning network architecture that can be used for image segmentation. UNet++ is a relatively classic image segmentation model. UNet++ has enhanced the feature extraction capability to a certain extent through nested skip connections and multi-scale feature fusion. After the introduction of the VSS module, the multi-scale feature extraction capability can be further enhanced, making the model perform better when processing images containing information of different scales. The complex structure of UNet++ performs well in feature extraction and fusion, but it also increases the computational complexity. The efficient design of the VSS module can reduce this computational burden, improve the overall efficiency of the model, and make training and reasoning faster and more efficient.

[0049] The reasons for replacing the third to fifth stage encoding modules with Visual Mamba's VSS module in the UNet++ encoder are: (1) Although UNet++ enhances feature fusion through nested skip connections, it may not be enough to fully capture features of different scales when dealing with complex scenes. The multi-scale feature fusion capability of the Visual Mamba module can make up for this deficiency and enhance the capture of image details and global information; (2) UNet++ mainly relies on convolutional layers for feature extraction, which may be insufficient when dealing with certain complex textures and boundaries. The innovative convolution operation introduced by the Visual Mamba module can enhance the feature extraction capability and enable the model to better handle complex image content. (3) The complex structure of UNet++ increases computational complexity and memory consumption, resulting in slower training and inference speeds. The lightweight design and efficient training strategy of the Visual Mamba module can reduce computational complexity and improve training and inference efficiency.

[0050] Advantages of replacing the third to fifth stage encoding modules with Visual Manba's VSS module in the UNet++ encoder:

[0051] 1) Improved multi-scale feature extraction capabilities: Traditional encoding layers usually use fixed-size convolution kernels when processing images, which may limit the model's ability to capture information at different scales. Visual Mamba's VSS module uses variable scale sampling technology to more flexibly capture features at different scales. This helps the model perform better when processing images that contain details and large-scale information.

[0052] 2) Improve model flexibility and efficiency: Traditional convolutional coding layers require a lot of computation when extracting features, especially in deep networks, where computational complexity and memory consumption are high. The VSS module uses a more efficient sampling and convolution strategy that can reduce computation and memory usage while maintaining or even improving the quality of feature extraction.

[0053] 3) The traditional coding layer structure is fixed and difficult to flexibly adjust according to different task requirements. The design of the VSS module is more flexible and can adapt to the needs of different tasks. By adjusting the sampling and convolution parameters, better task adaptability can be achieved.

[0054] Step 108: After preprocessing and data fusion, the acquired multi-source data of the area to be identified is input into the trained improved UNet++ semantic segmentation model to obtain the address disaster identification result.

[0055] Specifically, multi-source data of the area to be identified are collected, including optical data, vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, slope aspect data and geological lithology data, etc. The area to be identified can be distinguished from the preset area.

[0056] Preprocess the multi-source data, including radiation correction, geometric correction, atmospheric correction, etc. Deploy the improved UNet++ semantic segmentation model trained with the multi-source data set into the geological disaster identification system. Process and analyze the multi-source data of the identification area, and use the model to identify landslides. Apply the identification results to the geological disaster warning and prevention work in Hunan Province, and provide decision support for relevant departments.

[0057] In the above-mentioned automatic identification method of geological disasters based on multi-source data and deep learning, the method is based on multi-source remote sensing data and deep learning technology, which can make full use of various remote sensing data sources and combine advanced deep learning algorithms to achieve accurate and efficient identification of geological disasters. The improved UNet++ semantic segmentation model uses the capabilities of UNet++ and Visual Mamba in a new architecture for image segmentation, and uses the improved UNet++ semantic segmentation model trained with multi-source data to achieve rapid and accurate identification of large-scale landslides; by integrating multi-source remote sensing information, it not only overcomes the information limitations of a single data source, but also improves the accuracy and efficiency of geological disaster identification, providing strong technical support for geological disaster prevention and emergency response.

[0058] In one embodiment, the multi-source remote sensing data in step 100 includes vegetation cover NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data.

[0059] Specifically, multi-source remote sensing data in the preset area is obtained, including vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data. At the same time, high-resolution optical image data is required, such as data from different periods of Gaofen-1, Gaofen-2 and Gaofen-6.

[0060] In one embodiment, step 102 specifically includes the following steps:

[0061] Step 200: Perform radiation calibration, atmospheric correction, registration, cropping, and panchromatic multi-spectral fusion on the optical data to generate a fused image with both high spatial resolution and high spectral resolution.

[0062] Specifically, the original multispectral data of several scenes of high-resolution data were radiometrically calibrated, and the original DN (Digital Number) values ​​observed by the remote sensor were converted into radiation brightness values ​​or reflectivity values ​​with physical significance; atmospheric correction was performed to eliminate the influence of the atmosphere on the surface reflectivity or radiation brightness to obtain real surface information; the images were orthorectified using DEM data; Xupu County, Hunan Province was selected as the study area and the images were cropped; the original panchromatic images of several scenes of high-resolution data were radiometrically calibrated and orthorectified; the spatial resolution of the panchromatic image was combined with the spectral information of the multispectral image to generate a fused image with both high spatial resolution and high spectral resolution.

[0063] Step 202: After cropping the vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data, the data is resampled using the nearest neighbor interpolation method; wherein the resolution of the resampling result is consistent with the resolution of the fused image.

[0064] For other multi-source remote sensing data, the specific methods are as follows:

[0065] Xupu County, Hunan Province was selected as the research area, and vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data were clipped respectively. Commonly used resampling methods include nearest neighbor interpolation, bilinear interpolation and cubic convolution interpolation. The nearest neighbor interpolation method was selected to resample the above multi-source data to a resolution consistent with the optical image. The specific method of the nearest neighbor interpolation method is:

[0066] Determine distance: First, determine the distance between the target pixel and its surrounding pixels.

[0067] Select nearest neighbor: Then, the cell closest to the target cell is selected from the surrounding cells.

[0068] Assignment: Assign the value of the nearest neighbor pixel to the target pixel as the new value of the pixel.

[0069] Step 204: synthesize the fused image and the obtained resampling result into a multi-band image containing all multi-source data.

[0070] Specifically, use remote sensing image processing software (such as ENVI, ArcGIS, etc.) to perform band synthesis operations to synthesize the data into a multi-band image containing all the above data. Perform a quality check on the synthesized multi-band image to ensure that all bands are loaded correctly, check the data quality and clarity of each band, and pay attention to whether there are problems such as missing data, misalignment, or outliers.

[0071] In one of the embodiments, step 104 includes: visually interpreting the remote sensing image using remote sensing image interpretation software; outlining the boundary of the landslide according to the characteristics of the landslide on the image; converting the outlined landslide boundary from a surface element into a raster image to obtain a landslide image containing landslide samples; cropping the multi-band image and the landslide image containing the landslide samples into sub-images of a preset size; removing the landslide samples of the sub-image containing abnormal values, and deleting the sub-images whose landslide pixel ratio is less than a preset value to obtain a landslide sample library.

[0072] Specifically, (1) landslide samples are prepared from the synthesized multi-band images for subsequent model training and verification. The specific method is as follows: visually interpret the remote sensing images using professional remote sensing image interpretation software. The boundaries of the landslide are delineated based on the features of the landslide on the image (such as morphology, texture, color tone, etc.).

[0073] (2) Convert the outlined landslide boundary from a surface feature into a raster image. In the GIS software, use the rasterization tool to convert vector data into raster data for subsequent processing and analysis. At the same time, it is also necessary to convert the background area (i.e., non-landslide area) into a raster image as a negative sample for model training.

[0074] (3) In order to reduce the complexity of model training and improve computational efficiency, the landslide image containing the landslide sample and the multi-source remote sensing data synthesized in step 102 are cropped into sub-images of size 256×256. During the cropping process, it is necessary to ensure that each sub-image contains sufficient landslide and background information so that the model can learn effective features. At the same time, a certain degree of overlap is set to increase the diversity and robustness of the samples.

[0075] (4) Screen the initially prepared landslide samples and remove abnormal data. For sub-images containing outliers, directly remove them from the sample library to prevent these outliers from having a negative impact on model training. At the same time, remove sub-images with a very small proportion of landslide pixels, because such samples provide limited information in model training and may have an adverse effect on the generalization ability of the model.

[0076] (5) The samples in the landslide sample library are divided into training set, validation set and test set in a ratio of 8:1:1. The training set is the main data source for model training and is used to guide the model to learn the characteristics and laws of landslide identification. During the training process, the model adjusts and optimizes parameters according to the samples in the training set to improve its recognition accuracy. The validation set is used to evaluate the performance of the model during the training process, identify possible problems in the model and make corresponding adjustments. The test set is used to evaluate the generalization ability of the model after the model training is completed, that is, the model's ability to recognize unknown data. The results of the test set can be used to understand the performance of the model in practical applications, and the model can be further optimized and improved accordingly.

[0077] In one embodiment, the landslide samples include: a training set and a test set; step 106 includes: inputting the samples in the test set into the encoder of the improved UNet++ semantic segmentation model, extracting features of the input samples through two 3×3 convolution modules in the first stage to obtain a first intermediate feature, and extracting features of the first intermediate feature through downsampling and three 3×3 convolution modules in the second stage to obtain a second intermediate feature; in the third stage, the second intermediate feature is processed by a linear latent layer after being separated and divided, and the processed result is subjected to feature learning through two consecutive VSS blocks to obtain a third intermediate feature; in the fourth stage, the third intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks to obtain a fourth intermediate feature; in the fifth stage, the fourth intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks, and then the obtained feature is processed by a bottleneck layer to obtain a fifth intermediate feature; The first intermediate feature, the second intermediate feature, the third intermediate feature, the fourth intermediate feature and the fifth intermediate feature are input into the decoder of the improved UNet++ semantic segmentation model, and a nested decoder based on a dense skip connection structure is used to restore the resolution of the features of each stage. Finally, the features of each stage are weighted fused to obtain a geological disaster identification prediction result; according to the geological disaster identification prediction result and the training set, the Adam optimizer is used to update and optimize the network parameters to obtain a trained improved UNet++ semantic segmentation model; the verified improved UNet++ semantic segmentation model is tested with a test set, and the accuracy index of the model recognition is calculated according to the test results; different band combinations are changed for training, the accuracy index of different band combinations is calculated, and the influencing factors and the effects of different band combinations are tested; the generalization of the improved UNet++ semantic segmentation model to different regions and multi-source data is tested to obtain a trained improved UNet++ semantic segmentation model.

[0078] Specifically, the specific workflow of the improved UNet++ semantic segmentation model is as follows: First, since patch slicing will lead to information loss, thus affecting the accuracy, a two-layer CNN structure is designed at a shallower level to supplement the local feature extraction capability of the model. In the first stage, the first intermediate feature F1∈R is obtained through two 3×3 convolution modules. H×W×C In the second stage, the second intermediate feature F2∈R is obtained through downsampling and three 3×3 convolution modules. (H / 2)×(W / 2)×2C. Afterwards, in the third to fifth stages of the encoder, the c-dimensional tokenized input with reduced resolution is subjected to two consecutive VSS blocks for feature learning, maintaining the dimension and resolution. The VSS network block is mainly based on the Visual Mamba structure. Three patch merging is used as a downsampling process in the encoder of Mamba-UNet++, which reduces 12 token counts and increases the feature dimension by 2 times by splitting the input into quadrants, concatenating them, and then normalizing the dimension by layer norm each time. In the decoder part, the resolution of the features of each stage is restored by using the nested decoder of DSC. Finally, the features of each stage are weighted fused to obtain the final result. In addition, two VSS blocks are used in the bottleneck layer of the improved UNet++ semantic segmentation model. Each encoding module and the decryption module of the corresponding layer use dense skip connections to mix multi-scale features and upscaled outputs, that is, not only short connections are used, but also long connections are added to the network to enhance spatial details by merging shallow and deep layers. The subsequent linear layer maintains the dimension of this integrated feature set to ensure consistency with the upscaled resolution.

[0079] (1) Randomly select 80% of the images in the landslide sample set to train the deep learning Mamba-UNet++ model, and use the other 10% of the data as the test set. According to the test results, the accuracy index of the model recognition is calculated to quantify the segmentation effect. The models of this embodiment are trained and tested on a computer configured with an NVIDIA GeForce RTX 3060 graphics card. Pytorch1.13 (GPU version) is used as the deep learning framework of the experiment, and CUDA 11.6 and CuDNN 8.3 are selected to realize GPU parallel computing and GPU acceleration.

[0080] (2) Data augmentation is applied to the landslide sample set to increase the amount and diversity of data, thereby improving the generalization ability and robustness of the model. The specific data augmentation methods include: rotating 90°, 180°, 270°, or no rotation with a probability of 1 / 4; flipping from top to bottom, left to right, or no flipping with a probability of 1 / 3; random cropping with a probability of 0.8; and randomization with a probability of 0.4.

[0081] (3) During the training process, the Adam optimizer was first used to update and optimize the network parameters, and the initial learning rate of the training was set to 1e-4. In order to more effectively adjust the learning rate, the cosine annealing algorithm was used as a dynamic adjustment method for the learning rate. The weight initialization method used the Normal method, and the initial values ​​were assigned to the various weight parameters of the model before the training began. After experiments, the mini-batch size and training iteration number of all models were set to 8 and 150, respectively.

[0082] (4) The optimizer can also choose Adamw optimizer and SGD optimizer, and the weight initialization method can also choose Kaiming method in addition to Normal method. After changing the optimizer and weight initialization method, retrain the model and evaluate its performance on the validation set. Compare the performance differences under different configurations to determine which combination is most suitable for the task and dataset.

[0083] (5) The main indicator used to evaluate the performance of deep learning models is mean Intersection over Union (mIoU). In semantic segmentation tasks, mIoU is often used to measure the similarity between the pixel-level classification results predicted by the model and the true label. Intersection over Union (IoU) is the ratio of the intersection to the union of two sets. In semantic segmentation tasks, these two sets are usually the segmentation map predicted by the model and the segmentation map of the true label (ground truth).

[0084] For each category, IoU can be calculated by the following formula:

[0085]

[0086] Among them: TP (True Positives) is the true positive example, that is, the number of pixels predicted by the model as positive and the true label is also positive; FP (False Positives) is the false positive example, that is, the number of pixels predicted by the model as positive but the true label is negative; FN (False Negatives) is the false negative example, that is, the number of pixels predicted by the model as negative but the true label is positive;

[0087] In order to obtain the mIoU of the entire dataset, the IoU of all categories is averaged, and the formula is:

[0088]

[0089] Where k is the total number of categories. The higher the mIoU value, the closer the segmentation map predicted by the model is to the true label, and the better the performance of the model.

[0090] In addition, Matthews Correlation Coefficient (MCC) and Precision are used as auxiliary indicators, and the formula is:

[0091]

[0092]

[0093] Among them, TN is the true negative example, that is, the number of pixels predicted by the model as negative and the true label is also negative. MCC takes into account all four values ​​(TP, TN, FP, FN), so it is suitable for datasets with class imbalance, where one class may be more common than another. In particular, when you want to evaluate model performance without being affected by class distribution, MCC provides an evaluation criterion that is not affected by class distribution by calculating the correlation between predicted and actual classification. Therefore, even when one class is far more common than another, MCC can provide a relatively fair evaluation of model performance.

[0094] In one of the embodiments, different band combinations are replaced for training, the accuracy indicators of different band combinations are calculated, and the influencing factors and effects of different band combinations are tested, including: preparing different band combination data sets, using each band combination to train an improved UNet++ semantic segmentation model, evaluating the performance of the trained model on the test set, calculating and comparing the accuracy indicators of different band combinations, evaluating the impact of different band combinations on model performance, and selecting the optimal band combination for subsequent model training and geological disaster identification tasks based on the evaluation results.

[0095] Specifically, test the impact of different band combinations on model training results and find the optimal band combination to improve the accuracy of automatic identification of geological hazards. Specifically, prepare different band combination data sets, use improved UNet++ to train each band combination, evaluate the performance of the trained model on the test set, and calculate and compare the accuracy indicators of different band combinations. Analyze the impact of different band combinations on model performance, especially the changes in accuracy indicators. According to the evaluation results, select the optimal band combination for subsequent model training and geological hazard identification tasks.

[0096] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0097] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for automatic identification of geological hazards based on multi-source data and deep learning, characterized in that: The method comprises: Acquire multi-source data in a preset area, wherein the multi-source data includes: optical data and multi-source remote sensing data; Preprocessing the multi-source data, and fusing the preprocessed multi-source data to obtain a multi-band image; Determine the landslide boundary according to the multi-band image and generate a landslide sample; The improved UNet++ semantic segmentation model is trained by using the landslide sample to obtain a trained improved UNet++ semantic segmentation model; the improvement of the improved UNet++ semantic segmentation model is in the encoder part, a VSS module based on the Visual Mamba structure is used to replace the second to fourth layer encoding modules, and a bottleneck layer is added after the fourth layer encoding module, and the bottleneck layer is composed of two VSS modules; After preprocessing and data fusion, the multi-source data of the area to be identified are input into the trained improved UNet++ semantic segmentation model to obtain the address disaster identification results.

2. The method according to claim 1, characterized in that The multi-source remote sensing data includes vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, slope aspect data and geological lithology data.

3. The method according to claim 1, characterized in that Preprocessing the multi-source data and fusing the preprocessed multi-source data to obtain a multi-band image, including: Performing radiometric calibration, atmospheric correction, registration, cropping, and panchromatic multispectral fusion on the optical data to generate a fused image with both high spatial resolution and high spectral resolution; After cropping the vegetation coverage NDVI data, precipitation data, land cover type data, geological elevation data, slope data, aspect data and geological lithology data, the data is resampled using the nearest neighbor interpolation method; wherein the resolution of the resampling result is consistent with the resolution of the fused image; The fused image and the obtained resampling result are synthesized into a multi-band image containing all multi-source data.

4. The method according to claim 1, characterized in that: Determine the landslide boundary according to the multi-band image and generate a landslide sample, including: Use remote sensing image interpretation software to visually interpret remote sensing images; outline the boundaries of the landslide based on the characteristics of the landslide on the image; The delineated landslide boundary is converted from a surface feature into a raster image to obtain a landslide image containing landslide samples; Cutting the multi-band image and the landslide image containing the landslide sample into sub-images of a preset size; The landslide samples in the sub-images containing outliers are removed, and the sub-images whose landslide pixel ratio is less than the preset value are deleted to obtain the landslide sample library.

5. The method according to claim 1, characterized in that The landslide samples include: a training set and a test set; The improved UNet++ semantic segmentation model is trained using the landslide sample to obtain a trained improved UNet++ semantic segmentation model, including: The samples in the test set are input into the encoder of the improved UNet++ semantic segmentation model. In the first stage, the input samples are feature extracted by two 3×3 convolution modules to obtain the first intermediate feature. In the second stage, the first intermediate feature is feature extracted by downsampling and three 3×3 convolution modules to obtain the second intermediate feature. In the third stage, the second intermediate feature is processed by a linear latent layer after being separated and divided, and the processed result is subjected to feature learning through two consecutive VSS blocks to obtain the third intermediate feature. In the fourth stage, the third intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks to obtain the fourth intermediate feature. In the fifth stage, the fourth intermediate feature is processed by a block merging layer and then subjected to feature learning through two consecutive VSS blocks, and then the obtained feature is processed by a bottleneck layer to obtain the fifth intermediate feature. The first intermediate feature, the second intermediate feature, the third intermediate feature, the fourth intermediate feature and the fifth intermediate feature are input into the decoder of the improved UNet++ semantic segmentation model, and the resolution of the features of each stage is restored by using a nested decoder based on a dense skip connection structure. Finally, the features of each stage are weightedly fused to obtain the geological disaster identification prediction result; According to the geological disaster identification prediction results and the training set, the Adam optimizer is used to update and optimize the network parameters to obtain a trained improved UNet++ semantic segmentation model; The verified improved UNet++ semantic segmentation model is tested using the test set, and the accuracy index of model recognition is calculated based on the test results; Change different band combinations for training, calculate the accuracy indicators of different band combinations, and test the influencing factors and the effects of different band combinations; The generalization of the improved UNet++ semantic segmentation model to different regions and multi-source data is tested to obtain the trained improved UNet++ semantic segmentation model.

6. The method according to claim 5, characterized in that Change different band combinations for training, calculate the accuracy indicators of different band combinations, and test the influencing factors and effects of different band combinations, including: Prepare different band combination data sets, use each band combination to train the improved UNet++ semantic segmentation model, evaluate the performance of the trained model on the test set, calculate and compare the accuracy indicators of different band combinations, evaluate the impact of different band combinations on model performance, and select the optimal band combination for subsequent model training and geological disaster identification tasks based on the evaluation results.

Images