Landslide refined extraction intelligent method and system based on high-resolution remote sensing image

Abstract

The invention belongs to the technical field of intelligent landslide refined extraction methods, and particularly relates to an intelligent landslide refined extraction method and system based on a high-resolution remote sensing image, and the method comprises the steps: collecting a remote sensing original image of a high-resolution second satellite, combining an artificially interpreted landslide vector label, and uniformly cutting the processed image and the label to form a data set; a UCTransNet is used as a backbone network, a pyramid convolution kernel interaction module is deployed at the tail end of each encoder block comprising two convolution layers in an encoder part of the UCTransNet, a triple attention mechanism is deployed after the encoder blocks are output, after a Transformer module is output and after each time of up-sampling and feature splicing of a decoder, and on a device configured with NVID IA GeForce RTX4070 or a GPU with the same performance, a kernel kernel interaction module is deployed at the tail end of each encoder block comprising two convolution layers. And training the UCTransNet-TPKI model by adopting an Adam optimizer, evaluating the performance of the model through the accuracy rate, the recall rate, the F1 score and the intersection-to-union ratio, simultaneously executing ablation research to verify the independent contribution and the cooperative gain of the PKI module and the TA mechanism, and comparing the performance of the UCTransNet-TPKI model with the performance of the MFFENet model and the performance of the DeepLab v3 + model.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent methods for fine landslide extraction, and particularly relates to an intelligent method and system for fine landslide extraction based on high-resolution remote sensing images. Background Technology

[0002] Landslides, as a frequent geological disaster, pose a serious threat to the lives and property of residents in mountainous areas, the operation of major transportation routes, and the stability of the ecological environment. Accurate identification and cataloging of landslides are core prerequisites for geological disaster risk assessment, hazard investigation, and emergency response. With the development of high-resolution remote sensing technology, satellites such as Gaofen-2 (GF-2) can provide image data in the panchromatic band with a resolution of 0.8 meters and the multispectral band with a resolution of 3.2 meters. Leveraging its advantages of high spatial resolution, wide coverage, and short revisit period, high-resolution remote sensing imagery has become a core data source for large-scale landslide extraction, providing a technical foundation for integrated space-ground geological disaster monitoring.

[0003] However, current landslide extraction technology based on high-resolution remote sensing imagery still faces several bottlenecks, making it difficult to meet the demands for refined, automated, and high-precision applications. Specific problems include:

[0004] For a long time, landslide extraction has relied on manual visual interpretation, that is, professionals manually delineating landslide boundaries on remote sensing images.

[0005] This method has significant drawbacks: on the one hand, it is extremely inefficient, with the interpretation of a single high-resolution image taking several hours or even days, which cannot meet the needs of large-scale (such as county or city) or dynamic landslide monitoring; on the other hand, the interpretation results are greatly affected by personnel experience and subjective judgment, with different interpreters having a 15%-30% difference in defining landslide boundaries (especially small-scale landslides), resulting in insufficient consistency and accuracy of the landslide catalog and making it difficult to support subsequent quantitative risk assessment.

[0006] In recent years, deep learning technologies (such as U-Net, DeepLab v3+, and UCTransNet) have become the mainstream technology for landslide extraction due to their end-to-end feature learning capabilities. However, existing models still have three major drawbacks, making it difficult to adapt to the complex features of landslides in high-resolution images:

[0007] Landslide morphology varies greatly in scale—from small, shallow soil landslides of tens of square meters (such as the landslide type that accounts for 85.68% in the study area of ​​Wushan County, Chongqing) to large rock landslides of tens of thousands of square meters. Existing models (such as baseline UCTransNet and MFFENet) mostly use single-scale or simple cascaded convolutional structures, which either fail to capture the fine edge texture of small landslides (leading to missed detections) or fail to cover the macroscopic morphological structure of large landslides (leading to boundary positioning errors), and cannot simultaneously adapt to the extraction needs of landslides of different scales.

[0008] The spectral features of small landslides highly overlap with those of bare land, temporary construction land, and fallow land. Existing models' attention mechanisms mostly focus on a single dimension (such as only the channel dimension or only the spatial dimension), without considering the cross-dimensional interaction dependency of feature map channels-height-width. This makes it impossible to effectively learn the intrinsic relationship between spectral features and spatial structure, resulting in a persistently high false detection rate.

[0009] Small landslides, due to their shallow sliding surface and small damage area, only exhibit subtle texture changes (such as interrupted vegetation cover and blurred boundaries of exposed soil) on high-resolution images. Existing models often employ dimensionality reduction operations in their spatial attention branches, leading to the loss of edge feature information and further reducing the recall rate for identifying small landslides. Experimental data shows that existing models are less effective for landslides with areas <500m². 2 The recall rate for landslides is generally below 75%, and a large number of potential landslides have not been included in the catalog.

[0010] In summary, existing landslide extraction technologies have significant shortcomings in terms of automation, precision, interference suppression, and practicality, making it difficult to meet the needs of high-precision extraction and standardized cataloging of small-scale landslides under high-resolution remote sensing images. There is an urgent need for an intelligent extraction technology that can adaptively capture multi-scale landslide features, has strong interference suppression capabilities, and a high degree of automation to support the scientific prevention and risk management of geological disasters. Summary of the Invention

[0011] The purpose of this invention is to address the aforementioned technical problems by providing an intelligent method and system for refined landslide extraction based on high-resolution remote sensing images.

[0012] In view of this, the present invention provides an intelligent method and system for fine-grained landslide extraction based on high-resolution remote sensing images. Step 1: Collect original remote sensing images from the Gaofen-2 satellite, combine them with manually interpreted landslide vector labels, and uniformly crop the processed images and labels to form a dataset.

[0013] Step 2: Using UCTransNet as the backbone network, in the encoder part of UCTransNet, a pyramid convolution kernel interaction module is deployed at the end of each encoder block containing two convolutional layers. After the encoder block output, after the Transformer module output, and after each upsampling and feature concatenation of the decoder, a triplet attention mechanism is deployed.

[0014] Step 3: Train the UCTransNet-TPKI model on a device equipped with an NVIDIA GeForce RTX4070 or equivalent GPU using the Adam optimizer;

[0015] Model performance was evaluated using precision, recall, F1 score, and crossover ratio. Ablation studies were performed to verify the independent contributions and synergistic gains of the PKI module and the TA mechanism. The performance of the UCTransNet-TPKI model was compared with that of MFFENet and DeepLab v3+ models.

[0016] Step 4: Input the preprocessed test set images into the trained UCTransNet-TPKI model to output pixel-level landslide segmentation maps;

[0017] The binarization mask of the segmentation image is processed;

[0018] Step 5: Finally, call the GDAL geographic data processing library to convert it into Shapefile or GeoJSON standard geospatial vector format, generating a landslide catalog that can be used for geographic information system analysis.

[0019] Preferably, the image in step one includes a 0.8-meter resolution panchromatic band and a 3.2-meter resolution multispectral band;

[0020] The spatial resolution of the multispectral bands was improved to 0.8 meters through full-color sharpening and fusion technology;

[0021] The dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0022] Preferably, in step two, the backbone network adopts the Channel Transformer (CTrans) module; in the encoder part of the UCTransNet, a pyramid convolution kernel interaction module is deployed at the end of each encoder block containing two convolutional layers, and a triplet attention mechanism is deployed after the encoder block output, after the Transformer module output, and after each upsampling and feature concatenation of the decoder.

[0023] The PKI module employs four depthwise separable convolutional branches with kernel sizes increasing from 5×5 to 11×11, set in parallel, and one identity mapping branch. It first extracts basic texture features through a 3×3 standard convolution, and then weights and aggregates the outputs of each branch through a 1×1 convolution.

[0024] The TA mechanism consists of three parallel branches that capture the cross-dimensional interactions of the feature map's channel-width, channel-height, and height-width, respectively. Each branch generates attention weights through Z-Pool operations, standard convolutions, and the Sigmoid function. The final output is the weighted average of the results from the three branches.

[0025] Preferably, step four also includes calling the GDAL geographic data processing library to convert it into Shapefile or GeoJSON standard geospatial vector format, generating a landslide catalog that can be used for geographic information system analysis.

[0026] Preferably, the data acquisition and preprocessing in step one specifically includes:

[0027] The first submodule operation: Perform full-color sharpening and fusion processing on the acquired GF-2 satellite remote sensing raw images to uniformly improve the spatial resolution of the multispectral bands to 0.8 meters;

[0028] The second submodule operation: Based on the manually drawn landslide vector ground value, the image with improved resolution and its corresponding binary mask are cropped into non-overlapping slices of 256×256 pixels to form data units that pair images with labels, thereby constructing a dataset containing 1000 samples.

[0029] Preferably, the encoder of the UCTransNet-TPKI model in step two adopts a five-level downsampling structure, and each encoder block sequentially performs two standard convolutional layer operations, multi-scale feature extraction operations of the PKI module, and feature refinement operations of the TA mechanism.

[0030] The decoder of the model adopts a U-Net structure symmetrical to the encoder. It restores the resolution by progressively upsampling through transposed convolution. Each level of the decoder receives the splicing and fusion results of the upsampled features of the previous level, the output features of the Transformer module, and the corresponding level encoder features. After fusion, the features are refined through the TA mechanism. In the intermediate connection part of the model, the encoder features refined by the TA mechanism are input into the Transformer (CTrans) module to capture the long-distance dependencies between features.

[0031] Preferably, the mathematical expression of the PKI module is:

[0032]

[0033] Where L represents the basic features extracted by 3x3 convolution;

[0034] Z (m) For the output of each multi-scale branch;

[0035] This is a 1×1 convolution operation.

[0036] Preferably, in step two, each cross-dimensional branch of the TA mechanism first compresses the feature map through Z-Pool operation, then adjusts the number of channels through standard convolution, and then activates the attention weights through the Sigmoid function. The attention weights of the three branches are applied to the feature maps of the corresponding dimensions, and the results are weighted and averaged to obtain the final output of the TA mechanism.

[0037] The ablation study described in step three specifically involves testing the F1 scores of the baseline UCTransNet model, the UCTransNet model integrating only the PKI module, the UCTransNet model integrating only the TA module, and the UCTransNet-TPKI model integrating both the PKI module and the TA mechanism on the dataset.

[0038] Preferably, in step four, the landslide segmentation map generation utilizes the TensorRT inference engine to load the trained UCTransNet-TPKI model weights, and the landslide segmentation map is generated on a GPU device.

[0039] When converting a binary mask to a vector format using the GDAL library, the geographic coordinate information of the image is preserved simultaneously.

[0040] A landslide refinement system based on high-resolution remote sensing imagery includes a data acquisition and preprocessing module: used to perform the data acquisition and preprocessing operations described in step one to obtain a standardized dataset;

[0041] UCTransNet-TPKI Model Building Module: Used to build the UCTransNet-TPKI model described in step two, integrating the PKI module and the TA mechanism;

[0042] Model training and evaluation module: used to perform the model training and evaluation operations described in step 3, optimize model parameters and verify model performance;

[0043] Landslide Extraction and Result Generation Module: This module is used to perform the landslide extraction and result generation operations described in step four, and outputs a landslide catalog list in standard vector format.

[0044] The beneficial effects of this invention are:

[0045] This invention integrates a pyramidal kernel interaction (PKI) module into the UCTransNet backbone network to construct a feature learning mechanism that combines basic texture extraction with multi-scale branch parallel capture. First, the basic texture of the image is extracted using a 3×3 standard convolution. Then, depth-separable convolution branches with increasing sizes from 5×5 to 11×11 are used to simultaneously capture the fine edges of small-scale landslides (such as vegetation interruption textures and shallow soil exposure boundaries) and the macroscopic morphology of large-scale landslides (such as the overall outline of rock landslides). This perfectly adapts to scenarios with extremely large differences in landslide morphology scale (such as the coexistence of small shallow soil landslides and large rock landslides, which account for 85.68% of the study area in Wushan County, Chongqing).

[0046] On a dataset of 1000 samples from Wushan County, the UCTransNet-TPKI model of this invention achieved an F1 score of 0.9008 and an IoU of 0.8252, representing a 1.35% improvement over the baseline UCTransNet model (F1 = 0.8888); this is particularly evident for areas <500m². 2 For extremely small-scale landslides, the recall rate has been increased from generally <75% in existing models to over 89%, effectively solving the problem of traditional methods "catching the big and missing the small" and ensuring that no potential landslides are missed.

[0047] To address the spectral overlap problem between small-scale landslides and disturbing features such as bare land and fallow land, this invention innovatively introduces a triplet attention mechanism. Through three parallel branches across dimensions—channel-width (CW), channel-height (CH), and height-width (HW)—the model is forced to learn the intrinsic relationship between spectral features and spatial structure. Each branch generates attention weights through Z-Pool compression, standard convolution parameter tuning, and Sigmoid activation. The final weighted output achieves the effect of "focusing on landslide features and suppressing interference signals."

[0048] The UCTransNet model with the TA module integrated alone improved its F1 score from 0.8888 to 0.8919. When the TA mechanism works in conjunction with the PKI module, the false positive rate for bare land decreased from 18.7% in the baseline UCTransNet to below 9.2%, completely solving the industry pain point of "false detection of spectrally similar land cover" and ensuring the accuracy of landslide cataloging.

[0049] This invention constructs an end-to-end automated chain from "data preprocessing - model training - landslide extraction - result generation", completely eliminating the dependence on manual intervention and improving efficiency by tens of times compared with traditional methods:

[0050] Automatic full-color sharpening and fusion is performed on GF-2 images (unifying the resolution of multispectral bands to 0.8 meters), and image-label pairing units are generated by non-overlapping cropping at 256×256 pixels. The construction of a dataset of 1,000 samples can be completed in 2 hours (traditional manual cropping takes 1-2 days).

[0051] Automated model training and inference: Based on the NVIDIA GeForce RTX 4070 GPU, the Adam optimizer (initial learning rate 1e-4) automatically completes 500 training rounds. Combined with the TensorRT inference engine to load model weights, landslide segmentation of a single 10km×10km image takes only 8-10 minutes (traditional manual interpretation takes 4-6 hours).

[0052] The system automatically calls the GDAL library to convert the binary mask of the pixel-level segmentation map into GIS standard vector formats such as Shapefile / GeoJSON, without requiring manual secondary processing, and directly generates a landslide catalog that can be used for geological hazard analysis. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the steps of a deep learning-based method for refined landslide extraction according to an embodiment of this application.

[0054] Figure 2 This is an overall technical flowchart of a deep learning-based method for refined landslide extraction, as described in the application embodiment.

[0055] Figure 3 This is the architecture diagram of the UCTransNet-TPKI model;

[0056] Figure 4 This is a structural block diagram of a landslide refinement extraction device based on deep learning, according to an embodiment of this application. Detailed Implementation

[0057] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0058] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0059] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0060] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0061] This invention proposes a novel deep learning model called UCTransNet-TPKI. Based on the UCTransNet architecture, this model improves upon it by constructing a technical framework that combines a multi-layered feature fusion network with an attention mechanism. This effectively enhances the feature discriminability of small-scale landslides and improves their ability to distinguish them from disturbing features such as bare land.

[0062] The core innovation of this invention lies in the introduction of two key modules:

[0063] Pyramid Kernel Interaction (PKI) module:

[0064] Inspired by the Inception architecture, this module employs parallel, multi-branch depthwise separable convolutions to adaptively fuse multi-scale local contextual information.

[0065] By using convolution kernels of various sizes in parallel (such as 5x5 to 11x11), this module can simultaneously capture the fine edge textures of small landslides and the macroscopic morphological structure of large landslides, thus effectively addressing the problem of scale variability in landslides.

[0066] Triplet Attention (TA) Mechanism:

[0067] This mechanism aims to refine feature representations by capturing the interactive dependencies between feature maps in the three dimensions of channel, height, and width.

[0068] Its innovative cross-dimensional interactive design can force the model to learn the intrinsic relationship between spectral features and spatial structure, thereby effectively suppressing the interference of spectrally similar land features such as bare land. At the same time, its non-dimensionality-reduction spatial attention branch can enhance the perception of weak boundaries of small landslides.

[0069] This invention employs a refined module integration strategy, deploying PKI modules at the end of each level of the encoder to achieve hierarchical multi-scale feature extraction. Simultaneously, it strategically deploys the TA mechanism at key information flow nodes after the encoder and Transformer modules and in the decoder for feature refinement and attention guidance.

[0070] This synergistic effect of broad-based and then precise segmentation enables the model to see both the whole picture (adapting to multiple scales) and the details (resisting interference), ultimately achieving a significant breakthrough in segmentation accuracy.

[0071] In another aspect of the present invention, an electronic device is provided.

[0072] The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory exchange data with each other through the communication bus;

[0073] The memory stores computer programs, and when the program is executed by the processor, the following complete technical process is implemented:

[0074] First, the forward propagation of the UCTransNet-TPKI model is accelerated through GPU parallel computing. The CUDA kernel is used to efficiently perform multi-branch depthwise separable convolution operations in the PKI module, and tensor rotation and Z-Pool operations in the TA module are optimized. Next, geographic data processing libraries such as GDAL or OpenCV are called to perform automated preprocessing on the input GF-2 satellite imagery, including pan-color sharpening to improve spatial resolution to 0.8 meters and performing 256×256 pixel non-overlapping cropping to generate image patches. Then, the trained UCTransNet-TPKI model weights are loaded based on the TensorRT inference engine, enabling efficient landslide segmentation map generation on devices equipped with NVIDIA GeForce RTX4070 or equivalent GPUs. Finally, the binary landslide segmentation map mask is converted into a standard geospatial vector format (such as Shapefile or GeoJSON) using the GDAL library, generating a landslide catalog that can be used for Geographic Information System (GIS) analysis, thus completing a fully automated, high-precision landslide extraction process.

[0075] The technical solution provided by this invention can achieve high-precision and automated extraction of landslide targets, especially small-scale landslides, providing efficient and robust technical support for generating a more comprehensive and detailed landslide catalog, and has important practical application value for geological hazard assessment and risk management.

[0076] To make the technical problems, solutions, and advantages of this invention clearer, a detailed explanation is provided below using an example of Gaofen-2 (GF-2) satellite image processing. Addressing the core technical challenges of refined landslide extraction from high-resolution remote sensing images—the low recognition rate of small-scale landslides due to weak features, and severe spectral interference from background features such as bare land—this invention proposes a complete technical approach based on the UCTransNet-TPKI deep learning model.

[0077] This embodiment selects Wushan County, Chongqing Municipality, China (geographical coordinates 30°46′ to 31°28′N, 109°33′ to 110°11′E) as the study area. This region has a complex geological structure, with a predominantly mountainous and hilly terrain, making it a high-risk area for geological disasters. Statistics show that small, shallow soil landslides account for as much as 85.68% of historical landslides in this region. These landslides have indistinct morphological boundaries, and their spectral characteristics are highly similar to those of bare land and fallow land, leading to a large number of false positives and false negatives in the extraction results of traditional deep learning models (such as the baseline UCTransNet).

[0078] To this end, this invention designs the UCTransNet-TPKI architecture, achieving breakthroughs through two core technological innovations:

[0079] Pyramid Kernel Interaction (PKI) module: Employs parallel, multi-branch depthwise separable convolutions with kernel sizes increasing from 5x5 to 11x11 to capture multi-scale landslide features ranging from fine edges to macroscopic morphology.

[0080] Triplet Attention Mechanism:

[0081] By constructing cross-dimensional interactive branches in three dimensions—CW, CH, and HW—the model is forced to learn spectral-space correlation, thereby effectively suppressing background interference such as bare land and enhancing the perception of weak landslide edges.

[0082] The following, in conjunction with the accompanying drawings, provides a detailed description of the landslide refinement extraction scheme based on the UCTransNet-TPKI model, through specific embodiments and application scenarios.

[0083] As attached Figure 1 As shown, the landslide refinement extraction method based on deep learning in this application includes the following steps:

[0084] Step 101: Data Acquisition and Preprocessing;

[0085] Using Wushan County as the study area, we acquired raw GF-2 satellite imagery collected in 2022. This imagery includes a panchromatic band with a resolution of 0.8 meters and a multispectral band with a resolution of 3.2 meters. First, we used pan-sharpening fusion technology to enhance the spatial resolution of the multispectral imagery to 0.8 meters, providing a high-quality data foundation for refined extraction. Subsequently, combined with manually interpreted landslide vector labels, the imagery and labels were uniformly cropped into non-overlapping image patches of 256×256 pixels. The resulting Wushan dataset contains 1000 samples, divided into training, validation, and test sets in a 7:2:1 ratio.

[0086] Step 102: Construct the UCTransNet-TPKI model;

[0087] The UCTransNet-TPKI model proposed in this study uses UCTransNet as its backbone network, which replaces the skip connections of the traditional U-Net with channel Transformer (CTrans) modules. The core improvement of this invention lies in:

[0088] Encoder: At the end of each encoder block (containing two convolutional layers), a PKI module is deployed to perform hierarchical multi-scale feature extraction.

[0089] Attention mechanism: A Triplet Attention mechanism is deployed after each encoder block output, after the Transformer module output, and after each upsampling and feature concatenation in the decoder to guide and refine the key information flow.

[0090] PKI module structure: The module first extracts basic texture features through a 3x3 standard convolution; then, it captures multi-scale context in parallel through four parallel depthwise separable convolutional branches with kernel sizes of 5x5, 7x7, 9x9, and 11x11 and an identity mapping branch; finally, it performs a weighted aggregation of the outputs of all branches through a 1x1 convolution.

[0091] TA Module Structure: This module contains three parallel branches that capture cross-dimensional interactions of CW, CH, and HW, respectively. Each branch uses Z-Pool operations, standard convolutions, and the sigmoid function to generate attention weights, and the final output is a weighted average of the results from the three branches.

[0092] Step 103: Model training and evaluation;

[0093] Model training was performed on an NVIDIA GeForce RTX4070 GPU platform. The batch size was set to 6, the Adam optimizer was used for parameter updates, and the initial learning rate was set to 1e-4.

[0094] The model was trained for 500 epochs to achieve full convergence. Performance was evaluated using four standard metrics: precision, recall, F1 score, and intersection-over-union (IoU).

[0095] Step 104: Landslide Extraction and Results Generation;

[0096] The preprocessed test set images are input into the trained UCTransNet-TPKI model, which outputs pixel-level landslide segmentation maps (SegMaps). The output binarized mask is processed and converted to standard vector formats such as Shapefile or GeoJSON using geographic information processing libraries like GDAL, ultimately generating a refined landslide catalog map of the study area. Experimental results show that the model achieves an F1 score of 0.9008 and an IoU of 0.8252 on the Wushan dataset. The generated landslide catalog not only closely matches historical records but also discovers a large number of previously unrecorded new landslides.

[0097] The following is combined Figure 2 , Figure 3The method for refined landslide extraction based on deep learning, as described in this application, will be illustrated with a specific example.

[0098] Data acquisition and preprocessing stage:

[0099] This stage inputs raw GF-2 satellite imagery (Wushan County, Chongqing, China, acquired in 2022). First, a panchromatic fusion algorithm is used to fuse the 0.8-meter panchromatic band with the 3.2-meter multispectral band, outputting a four-band imagery with a resolution of 0.8 meters. Then, based on a manually drawn landslide vector map (Ground Truth), the corrected imagery and its corresponding binary mask (landslide = 1, non-landslide = 0) are cropped into 256×256 pixel non-overlapping slices, forming image + label paired data units.

[0100] Model building phase:

[0101] As attached Figure 3 As shown, the UCTransNet-TPKI network architecture of the present invention adopts an encoder-decoder symmetrical structure.

[0102] The encoder section is based on five levels of downsampling. Each encoder block contains two standard convolutional layers, followed by a PKI module and a TA module.

[0103] The mathematical expression of the PKI module is:

[0104]

[0105] Where L represents the basic features extracted by 3x3 convolution;

[0106] Z (m) This is the output of each multi-scale branch.

[0107] The TA module outputs a weighted average of the results from the three branches.

[0108] Intermediate connection section: The encoder features refined by the TA module are input into a channel Transformer (CTrans) module to capture long-distance dependencies.

[0109] The decoder section employs a U-Net symmetric structure, restoring resolution through progressive upsampling via transposed convolutions. Each stage receives and fuses the upsampled features from the previous stage, the Transformer output features, and the corresponding encoder layer features. After each fusion, the TA module is used for feature refinement.

[0110] Model evaluation phase:

[0111] The model's performance was comprehensively evaluated on a separate test set.

[0112] Ablation study: On the Wushan dataset, the baseline UCTransNet achieved an F1 score of 0.8888; adding the PKI module alone improved it to 0.8925; adding the TA module alone improved it to 0.8919; while the UCTransNet-TPKI model, which integrates both modules, achieved the highest F1 score of 0.9008. This trend was also consistent on the Palu dataset, demonstrating that the synergistic effect of the two modules ("1+1>2") is key to the performance improvement.

[0113] Model Comparison: Compared with other advanced models such as MFFENet and DeepLab v3+, UCTransNet-TPKI demonstrates competitive performance in key metrics such as F1 and IoU.

[0114] Results analysis phase:

[0115] The trained model was applied to the entire Wushan County study area to generate the final landslide catalog. Spatial analysis showed that the model's identification results highly overlapped with historical landslide records, while also effectively supplementing a large number of previously unrecorded new landslide areas, thus verifying the reliability and practical value of the method presented in this invention.

[0116] Figure 4 The structural block diagram of a landslide refinement extraction device based on deep learning is shown in this embodiment of the invention.

[0117] The landslide extraction device provided in this embodiment of the invention includes the following functional modules:

[0118] The data acquisition and preprocessing module is used to acquire data from high-resolution satellite remote sensing imagery and perform standardized processing for use by the model. Specifically, this module performs panchromatic sharpening on the original imagery to improve spatial resolution and combines manually interpreted label data to crop the image, generating image patches with a uniform format.

[0119] The UCTransNet-TPKI model building module is used to build a deep learning network model for refined landslide identification.

[0120] Specifically, it includes: using UCTransNet as the backbone network, its core innovation lies in integrating the Pyramid Kernel Interaction (PKI) module to capture multi-scale landslide features, and introducing the triplet attention (TA) mechanism to enhance the model's ability to identify landslide features and suppress background interference.

[0121] The model training and evaluation module is used to optimize the constructed model and verify its performance. Specifically, it includes iterative training of the model using the Adam optimizer and setting hyperparameters such as the initial learning rate. This module also includes comprehensive performance evaluation functions, systematically verifying the model's effectiveness and generalization ability by calculating quantitative metrics such as F1 score and Intersection over Union (IoU), and combining ablation experiments with state-of-the-art (SOTA) models.

[0122] The landslide extraction and result generation module is used to infer the landslide distribution from the trained UCTransNet-TPKI model and generate the final landslide distribution results.

[0123] Specifically, this includes: generating pixel-level landslide segmentation maps using network prediction results, converting them into vector format, and finally outputting a refined landslide catalog that can be used for geographic information system analysis.

[0124] The data acquisition and preprocessing module includes:

[0125] The first submodule is used to acquire remote sensing images from the Gaofen-2 (GF-2) satellite and improve the spatial resolution to 0.8 meters through full-color sharpening and fusion processing.

[0126] The second submodule is used to crop the images and their corresponding landslide labels into non-overlapping image blocks of 256×256 pixels, and divide them into training, validation and test sets in a ratio of 7:2:1.

[0127] The UCTransNet-TPKI model construction module includes:

[0128] The third submodule is used to build the PKI module, which uses parallel, depthwise separable convolutions with kernel sizes increasing from 5x5 to 11x11 to effectively adapt to the morphological characteristics of landslides of different scales.

[0129] The fourth submodule is used to integrate the TA mechanism, which enhances the perception of weak landslide edges and suppresses interference from spectrally similar features such as bare land by capturing the interaction between the three dimensions of channel, height and width.

[0130] The model training and evaluation module includes:

[0131] The fifth submodule is used to set the hyperparameters for model training, including using the Adam optimizer, setting the initial learning rate to 1e-4, and the batch size to 6.

[0132] The sixth submodule is used to perform ablation studies to quantify the independent contributions and synergistic gains of the PKI module and the TA mechanism to model performance.

[0133] The seventh submodule is used to compare the performance of the model of this invention with other advanced models such as MFFENet and DeepLab v3+ to verify its technical superiority.

[0134] The landslide fine extraction device provided in this invention combines efficient PKI multi-scale feature extraction capabilities with TA cross-dimensional attention mechanism to achieve high-precision identification of landslides, especially small-scale landslides, in high-resolution remote sensing images. It provides strong and reliable technical support for automated landslide cataloging and has significant scientific research value and practical application potential.

[0135] The embodiments of the present invention provide Figure 4 The landslide extraction device shown can realize the various specific processes in the method embodiments proposed in this invention, and will not be described again here to avoid redundancy.

[0136] Furthermore, this invention also provides a corresponding computer device, including a processor, a communication interface, and a memory, which implements the aforementioned landslide extraction method by executing a computer program stored in the memory. The memory can be random access memory (RAM) or other non-volatile storage media, such as a solid-state drive (SSD). The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA), providing efficient data processing and computing capabilities.

[0137] The methods and apparatus in the embodiments of the present invention can be implemented wholly or partially through software, hardware, or a combination thereof. The computer program can be stored in a computer-readable storage medium and remotely deployed and upgraded via network transmission to adapt to the needs of different application environments, exhibiting strong flexibility and scalability.

[0138] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A smart method for refined landslide extraction based on high-resolution remote sensing imagery, characterized in that: Step 1: Collect raw remote sensing images from the Gaofen-2 satellite, combine them with manually interpreted landslide vector labels, and then crop the processed images and labels to form a dataset. Step 2: Using UCTransNet as the backbone network, in the encoder part of UCTransNet, a pyramid convolution kernel interaction module is deployed at the end of each encoder block containing two convolutional layers. After the encoder block output, after the Transformer module output, and after each upsampling and feature concatenation of the decoder, a triplet attention mechanism is deployed. Step 3: Train the UCTransNet-TPKI model on a device equipped with an NVIDIA GeForce RTX4070 or equivalent GPU using the Adam optimizer; Model performance was evaluated using precision, recall, F1 score, and crossover ratio. Ablation studies were performed to verify the independent contributions and synergistic gains of the PKI module and the TA mechanism. The performance of the UCTransNet-TPKI model was compared with that of MFFENet and DeepLabv3+ models. Step 4: Input the preprocessed test set images into the trained UCTransNet-TPKI model to output pixel-level landslide segmentation maps; The binarization mask of the segmentation image is processed; Step 5: Finally, call the GDAL geographic data processing library to convert it into the Shapefile or GeoJSON standard geospatial vector format, generating a landslide catalog that can be used for geographic information system analysis.

2. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 1, characterized in that: The image described in step one includes a 0.8-meter resolution panchromatic band and a 3.2-meter resolution multispectral band; The spatial resolution of the multispectral bands was improved to 0.8 meters through full-color sharpening and fusion technology; The dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

3. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 1, characterized in that: In step two, the backbone network adopts the Channel Transformer (CTrans) module; in the encoder part of the UCTransNet, a pyramid convolution kernel interaction module is deployed at the end of each encoder block containing two convolutional layers, and a triplet attention mechanism is deployed after the encoder block output, after the Transformer module output, and after each upsampling and feature concatenation of the decoder. The PKI module employs four depthwise separable convolutional branches with kernel sizes increasing from 5×5 to 11×11, set in parallel, and one identity mapping branch. It first extracts basic texture features through a 3×3 standard convolution, and then weights and aggregates the outputs of each branch through a 1×1 convolution. The TA mechanism consists of three parallel branches that capture the cross-dimensional interactions of the feature map's channel-width, channel-height, and height-width, respectively. Each branch generates attention weights through Z-Pool operations, standard convolutions, and the Sigmoid function. The final output is the weighted average of the results from the three branches.

4. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 1, characterized in that: Step four also includes calling the GDAL geographic data processing library to convert it into the Shapefile or GeoJSON standard geospatial vector format, generating a landslide catalog that can be used for geographic information system analysis.

5. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 2, characterized in that: The data acquisition and preprocessing described in step one specifically includes: The first submodule operation: Perform full-color sharpening and fusion processing on the acquired GF-2 satellite remote sensing raw images to uniformly improve the spatial resolution of the multispectral bands to 0.8 meters; The second submodule operation: Based on the manually drawn landslide vector ground value, the image with improved resolution and its corresponding binary mask are cropped into non-overlapping slices of 256×256 pixels to form data units that pair images with labels, thereby constructing a dataset containing 1000 samples.

6. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 3, characterized in that: The encoder of the UCTransNet-TPKI model described in step two adopts a five-level downsampling structure. Each encoder block sequentially performs two standard convolutional layer operations, multi-scale feature extraction operations of the PKI module, and feature refinement operations of the TA mechanism. The decoder of the model adopts a U-Net structure symmetrical to the encoder. It restores the resolution by progressively upsampling through transposed convolution. Each level of the decoder receives the splicing and fusion results of the upsampled features of the previous level, the output features of the Transformer module, and the corresponding level encoder features. After fusion, the features are refined through the TA mechanism. In the intermediate connection part of the model, the encoder features refined by the TA mechanism are input into the Transformer (CTrans) module to capture the long-distance dependencies between features.

7. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 1, characterized in that: The mathematical expression of the PKI module is as follows: Where L represents the basic features extracted by 3x3 convolution; Z (m) For the output of each multi-scale branch; This is a 1×1 convolution operation.

8. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 3, characterized in that: In step two, each cross-dimensional branch of the TA mechanism first compresses the feature map through Z-Pool operation, then adjusts the number of channels through standard convolution, and then activates the attention weights through the Sigmoid function. The attention weights of the three branches are applied to the feature maps of the corresponding dimensions, and the results are weighted and averaged to obtain the final output of the TA mechanism. The ablation study described in step three specifically involves testing the F1 scores of the baseline UCTransNet model, the UCTransNet model integrating only the PKI module, the UCTransNet model integrating only the TA module, and the UCTransNet-TPKI model integrating both the PKI module and the TA mechanism on the dataset.

9. The intelligent method for refined landslide extraction based on high-resolution remote sensing imagery according to claim 1, characterized in that: The landslide segmentation map generation described in step four utilizes the TensorRT inference engine to load the trained UCTransNet-TPKI model weights, and implements the landslide segmentation map generation on a GPU device. When converting a binary mask to a vector format using the GDAL library, the geographic coordinate information of the image is preserved simultaneously.

10. A landslide refinement extraction system based on high-resolution remote sensing imagery, based on the intelligent method for refinement extraction of landslides based on high-resolution remote sensing imagery as described in any one of claims 1-9, characterized in that: Includes a data acquisition and preprocessing module: used to perform the data acquisition and preprocessing operations described in step one to obtain a standardized dataset; UCTransNet-TPKI Model Building Module: Used to build the UCTransNet-TPKI model described in step two, integrating the PKI module and the TA mechanism; Model training and evaluation module: used to perform the model training and evaluation operations described in step 3, optimize model parameters and verify model performance; Landslide Extraction and Result Generation Module: This module is used to perform the landslide extraction and result generation operations described in step four, and outputs a landslide catalog list in standard vector format.

Images