Erosion gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data

Abstract

The invention provides an erosion gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data, and relates to the field of erosion gully extraction, and the method comprises the steps: obtaining the PlanetScope multi-temporal data and the ALOS DEM data; preprocessing the acquired data and making a data set; manufacturing an erosion gully data set; the selected features are combined, training sample labels are marked, and an erosion gully data set is generated; designing a multi-classification model based on dynamic snakelike convolution to extract erosion gullies, and training the model by using 20-fold cross validation; performing precision evaluation on a model training result; and performing feature importance evaluation on model input features by using a partial dependency graph method, and determining optimal features extracted from the erosion gully according to a precision evaluation result to obtain an optimal result extracted from the erosion gully. According to the invention, large-scale real-time extraction and monitoring of the erosion gully are realized, and data support is provided for making an erosion gully treatment policy.

Description

Technical Field

[0001] The present invention relates to the field of permanent erosion gully extraction, in particular to a permanent erosion gully extraction method based on PlanetScope multi-temporal data and ALOS DEM data. Background Art

[0002] Soil erosion is one of the core drivers of land degradation and poses a serious threat to global food security and ecosystem sustainability. As a major form of soil erosion, current monitoring of gullies is crucial for their management and the protection of black soil resources. Currently, Google Earth imagery is the primary data source for gully extraction research, providing high-resolution (0.5-1.2 m) data with global coverage. However, these studies generally overlook two key issues: First, when mapping large areas, Google Earth imagery is acquired inconsistently, resulting in significant differences in phenological characteristics between images, severely weakening the model's ability to transfer data across time and space. Second, a large portion of these images are historical data, making them difficult to meet current monitoring needs. Furthermore, existing gully extraction networks are mostly simple semantic segmentation models that fail to account for the unique morphological characteristics of gullies. Therefore, there is an urgent need to explore new data sources and models to design an extraction method that is both suitable for large-scale monitoring and effectively considers gully morphological characteristics.

[0003] The networked observation mode of the PlanetScope satellite constellation enables simultaneous acquisition of high-temporal and spatial resolution data. In recent years, scholars have demonstrated its feasibility for gully extraction. The monthly-scale synthetic data provided by Planet Basemap, with its relatively consistent data acquisition time, effectively reduces phenological differences. Furthermore, its 4.77-meter resolution meets the needs of large-scale gully monitoring, making it an ideal data source. Given that gullies typically exhibit elongated tubular morphology, to enhance the model's ability to learn such features, the standard two-dimensional convolutional structure in the network is replaced with a dynamic snake convolution. Dynamic snake convolution was originally designed to improve the accuracy of vascular structure extraction in medical images. As a deformable convolution, it can more effectively learn the features of linear features. Therefore, combining the multi-temporal data source of Planet Basemap and embedding the dynamic snake convolution structure into the gully extraction model is expected to significantly improve the accuracy of gully extraction and expand the scope of research. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the present invention aims to propose a method and system for extracting gullies from PlanetScope multi-temporal data and ALOS DEM data, which can be used for large-scale gully extraction experiments.

[0005] To achieve the above object, the present invention provides the following solutions: The present invention provides a method for extracting gullies based on PlanetScope multi-temporal data and ALOS DEM data, comprising the following steps: Step 1: Acquire multi-temporal remote sensing data and terrain data. The multi-temporal remote sensing data is monthly-scale synthetic data used to provide high temporal and spatial resolution features. Step 2: Preprocessing the multi-temporal remote sensing data and terrain data and preparing a data set, wherein the preprocessing includes data screening and resolution unification, and the data set preparation includes generating an input feature set and marking classification labels; Step 3: Design a multi-classification model based on dynamic convolution to extract landform features. The model uses a deep learning network as the backbone, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results; Step 4: Evaluate the accuracy of the training results of the multi-classification model using a variety of error indicators; Step 5: Determine the optimal input features through the feature importance evaluation method to obtain the final result of landform feature extraction.

[0006] Furthermore, the pre-processing of the multi-temporal remote sensing data and terrain data and the preparation of data sets include: The multi-temporal remote sensing data are screened to remove data from months with large phenological differences and low image quality, and data from spring plowing and autumn harvest periods are retained; the terrain data are resampled to make its resolution consistent with that of the multi-temporal remote sensing data; The filtered multi-temporal remote sensing data and the resampled terrain data are combined to generate an input feature set; For the input feature set, landform features, roads and other category labels are annotated to generate a training sample set.

[0007] Furthermore, the design is based on a multi-classification model of dynamic convolution, including: Constructing a deep learning model with U-Net as the backbone, wherein the model input is multi-temporal remote sensing features and terrain features; The ordinary two-dimensional convolution in the U-Net is replaced by a dynamic snake convolution structure, which consists of ordinary convolution and dynamic convolution in the x and y directions; The Dice loss function is used to optimize the multi-classification model and output three classification results of landform features, roads and other categories; The multi-classification model was trained by 20-fold cross validation. Multiple data sets were randomly selected from the study area, one of which was used as a test set and the rest as training sets.

[0008] Furthermore, the accuracy evaluation of the training results of the multi-classification model includes: The volume error was evaluated using Dice coefficient, overall accuracy, precision, and recall metrics; The morphological error is evaluated using the Hausdorff distance metric based on persistent homology; The SSIM index is used to evaluate the structural error; The extraction accuracy of the model is comprehensively judged based on the error indicators.

[0009] Furthermore, determining the optimal input feature by using a feature importance evaluation method includes: Using partial dependency graph method to evaluate the feature importance of the input feature set; By assigning the feature of a certain month to zero, the change in the accuracy of the statistical model is calculated; Judging the validity of the features based on the change in accuracy, and retaining valid features; The optimal input feature set is generated based on the retained features to determine the final result of landform feature extraction.

[0010] Furthermore, the screening of the multi-temporal remote sensing data includes: Snow cover period data were eliminated based on the consistency of phenological characteristics to avoid interference from spatial heterogeneity; Eliminate low-quality images during cloudy and rainy periods and retain high-definition images; Determine the monthly data corresponding to the spring plowing period and the autumn harvest period as the screening results; Convert the format of the filtered data to generate input features in a unified format.

[0011] Furthermore, the multi-classification model is trained by 20-fold cross validation, comprising: Multiple multi-temporal remote sensing data were randomly selected from the study area; Generate equal samples for each data set and divide them into training set and test set; Iteratively optimizing the multi-classification model using a training set; The generalization ability of the model is verified through the test set to generate training results.

[0012] Another object of the present invention is to provide a gully extraction system for PlanetScope multi-temporal data and ALOS DEM data, comprising: Data acquisition unit: acquires multi-temporal remote sensing data and terrain data. The multi-temporal remote sensing data is monthly-scale synthetic data used to provide high temporal and spatial resolution features; Data preprocessing and data set preparation unit: preprocessing the multi-temporal remote sensing data and terrain data and preparing data sets, wherein the preprocessing includes data screening and resolution unification, and the data set preparation includes generating input feature sets and marking classification labels; Model training unit: Design a multi-classification model based on dynamic convolution to extract landform features. The model uses a deep learning network as the backbone, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results; Accuracy evaluation unit: performs accuracy evaluation on the training results of the multi-classification model, using multiple error indicators for evaluation; Feature importance evaluation unit: Determine the optimal input features through the feature importance evaluation method to obtain the final result of landform feature extraction.

[0013] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data provided by the present invention can realize the extraction and monitoring of gully potential over a large range, providing necessary data support for the formulation of gully control policies. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 A schematic diagram illustrating a flow chart of a method for extracting gullies based on PlanetScope multi-temporal data and ALOS DEM data provided by an embodiment of the present invention is shown; Figure 2 A schematic diagram showing the feature importance results of 20 models based on partial dependence graphs according to an embodiment of the present invention; Figure 3 Figures showing the extraction results of the model ablation experiment and the comparative experiment of the embodiment of the present invention, where (a)-(f) are test set samples, white represents the correct classification results of different models, blue represents missed classifications, and red represents incorrect classifications; Figure 4 A schematic diagram of the structure of an erosion gully extraction system based on PlanetScope multi-temporal data and ALOS DEM data provided by an embodiment of the present invention is shown; Figure 5 A schematic diagram of the network structure for erosion gully extraction provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] The purpose of the present invention is to provide a method and system for extracting erosion gullies based on PlanetScope multi-temporal data and ALOS DEM data, which can be used for large-scale extraction experiments of erosion gullies.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown in FIG, a method for extracting gullies based on PlanetScope multi-temporal data and ALOS DEM data includes: Step 1: Acquire multi-temporal remote sensing data and terrain data. The multi-temporal remote sensing data is monthly-scale synthetic data used to provide high temporal and spatial resolution features. Step 2: Preprocessing the multi-temporal remote sensing data and terrain data and preparing a data set, wherein the preprocessing includes data screening and resolution unification, and the data set preparation includes generating an input feature set and marking classification labels; Step 3: Design a multi-classification model based on dynamic convolution to extract landform features. The model uses a deep learning network as the backbone, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results; Step 4: Evaluate the accuracy of the training results of the multi-classification model using a variety of error indicators; Step 5: Determine the optimal input features through the feature importance evaluation method to obtain the final result of landform feature extraction.

[0020] Specifically, in one embodiment, step 1: obtain PlanetScope multi-temporal data and ALOS DEM data; wherein the PlanetScope multi-temporal data selects the monthly-scale synthetic data of Planet Basemap.

[0021] Specifically, step 2: preprocess the acquired data and create a dataset; preprocessing includes data screening and resampling, deleting monthly features with poor image quality and large differences in ground objects and phenology in the Planet Basemap monthly-scale synthetic data, and unifying the resolution of the DEM data through resampling methods; creating an erosion gully dataset; combining the selected features, annotating the training sample labels, and generating an erosion gully dataset, where the labels include erosion gullies, roads, and other categories.

[0022] In one embodiment, the data preprocessing described in step 2 above includes: screening the Planet Basemap monthly composite data, selecting monthly data based on phenological consistency and data quality. First, data from the snowy period from November to March are removed to avoid spatial heterogeneity, and low-quality imagery from the cloudy and rainy period in July is also excluded. Finally, data from April to June and August to October (covering the spring plowing and autumn harvest periods) are retained. ALOS data are resampled to maintain the same resolution as the Planet data.

[0023] In one embodiment, the dataset creation described in step 2 above involves combining the filtered Planet Basemap monthly-scale synthetic data with the resampled ALOS DEM data to generate an input feature set, while also labeling the input feature set with corresponding labels. Label types include gully, road, and other categories. Finally, each sample feature set and label are combined to complete the dataset creation.

[0024] Specifically, step three: design a multi-classification model based on dynamic snake convolution to extract erosion gullies, and use 20-fold cross validation to train the model.

[0025] In one embodiment, the multi-classification model design structure based on dynamic snake convolution described in step 3 above (such as Figure 5 The model input features are Planet multi-temporal features and DEM features; the model uses U-Net as the backbone, in which the ordinary two-dimensional convolution is replaced by a dynamic snake convolution structure, which consists of a common two-dimensional convolution and dynamic snake convolutions in the x and y directions; the model output is a three-category result, including: erosion gullies, roads, and other categories; the model loss function is Dice loss.

[0026] In one embodiment, the 20-fold cross-validation described in step 3 above involves randomly selecting 20 different Planet data sets from the study area. The same number of samples are generated for each data set. One sample set is used as the test set, and the remaining 19 samples are used as the training set.

[0027] Specifically, step four: evaluate the accuracy of the model training results. The evaluation indicators include volume error, morphological error, and structural error.

[0028] In one embodiment, the accuracy evaluation indicators in step 4 above include: volume error, morphological error, and structural error. For volume error, the Dice coefficient, overall accuracy, precision, and recall rate are selected; for morphological error, the Hausdorff distance metric based on persistent homology is selected; and for structural error, the SSIM metric is selected.

[0029] Specifically, step five: use the partial dependence graph method to evaluate the importance of the model input features, determine the optimal features for permanent erosion gully extraction based on the accuracy evaluation results, and obtain the optimal results of erosion gully extraction.

[0030] In one embodiment, the method of using the partial dependence graph to evaluate the importance of the model input features in step 5 is to assign a value of 0 to the features of a certain month after the model training is completed, and then calculate the change in the accuracy of the model. Figure 2 As shown in the figure, after assigning a certain feature value to 0, all accuracy evaluation indicators deteriorate, taking the median value of all models as the reference standard. This shows that in this example, both the DEM features and the monthly features are effective for accurately extracting gullies. They are all retained as valid monthly features and used as the input features for the final model selection.

[0031] In this embodiment of the present invention, ablation and comparative experiments based on the target model were conducted to systematically verify the effectiveness of the network design. The ablation experiments used a baseline U-Net architecture without the dynamic snake convolution structure and compared it with the complete model (including the dynamic snake convolution structure). The dynamic snake convolution structure consists of a standard 3×3 convolution in parallel with dynamic snake convolutions in the x and y directions. A learnable offset Δ∈[-1,1] is used to enhance sensitivity to tubular features. Four representative segmentation models (SegNet, LinkNet, R2U-Net, and RNN-based models) were selected for comparison. The ablation results show that the introduction of the dynamic snake convolution structure significantly improves the model's extraction accuracy by 11.2%. Compared to the precision metric, the recall metric decreases more significantly, resulting in more severe under-detection. This is primarily because the dynamic snake convolution structure better captures gully details, resulting in more complete extraction of the terminal or thinner parts of the gully structure and a lower under-detection rate. A comparative analysis of experimental results from mainstream semantic segmentation models, including R2U-Net, SegNet, and LinkNet, revealed that DSCNet achieved improvements of 13.2%, 19.5%, and 27.5% in the Dice metric, respectively. CNN extraction accuracy far outperformed RNN, reflecting the task's focus on the spatial structure of gullies rather than temporal dependencies, validating the inherent advantages of CNNs in spatial feature extraction. Among CNN models, RU2-Net achieved relatively high extraction accuracy, primarily due to its integration of a recurrent neural network and residual connections with U-Net. However, the accuracy difference between the two models for this task was not significant, with a difference of less than 3%. SegNet's accuracy was inferior, primarily due to differences in their feature fusion methods. SegNet used index pooling to restore spatial information, which, while saving space, resulted in a significant loss of detail. U-Net, on the other hand, used skip connections to concatenate feature maps with the corresponding decoder layer, integrating deep semantic features with shallower detail features, resulting in better boundary segmentation accuracy. LinkeNet achieved the worst accuracy, particularly in the precision metric. The main reason is that LinkeNet fuses features through element-wise addition, which reduces the number of bands while retaining less information than U-Net splicing. To further illustrate the difference in model extraction results, Figure 3 The extraction results of different networks for some erosion gullies are shown, with blue indicating missed classifications and red indicating misclassifications. Figure 3 The blue box in (a) indicates the end of the erosion gully and the thinner part. DSCNet's extraction results are more consistent with the real results, while the other networks have varying degrees of under-classification. This shows that the dynamic snake convolutional structure can better capture the geometric characteristics of the erosion gully. Figure 3 In (bd), each network has different degrees of misclassification, especially on the ridge (b) and road (cd), which are also the types of objects that the network easily misclassifies. However, DSCNet has fewer misclassification phenomena and higher extraction accuracy. Figure 3 In the results of (ef), DSCNet extracts erosion grooves more accurately, while other networks have varying degrees of misclassification and omission. In addition, the extraction results of RNN are more fragmented. This is because RNN has a strong ability to learn temporal features, but a poor ability to capture spatial features. This is also because the present invention uses less temporal feature data, and RNN is not suitable for this problem. Table 1 shows the accuracy evaluation results of the model ablation experiment and the comparative experiment of different models in the embodiment of the present invention.

[0032] Table 1

[0033] In addition, in another embodiment of the present invention, a gully extraction system based on PlanetScope multi-temporal data and ALOS DEM data is provided. Figure 4 , which includes: Data acquisition unit: acquires PlanetScope multi-temporal data and ALOS DEM data; among them, PlanetScope multi-temporal data selects the monthly scale synthetic data of Planet Basemap.

[0034] Data preprocessing and dataset creation unit: The acquired data are preprocessed and datasets are created. Preprocessing includes data screening and resampling, removing monthly features with poor image quality and large differences in ground features and phenology from the Planet Basemap monthly-scale synthetic data, and unifying the resolution of the DEM data through resampling methods. An erosion gully dataset is created. The selected features are combined and training sample labels are annotated to generate an erosion gully dataset, where labels include erosion gullies, roads, and other categories.

[0035] Model training unit: Design a multi-classification model based on dynamic snake convolution to extract erosion gullies, and use 20-fold cross validation to train the model.

[0036] Accuracy evaluation unit: performs accuracy evaluation on model training results. Evaluation indicators include volume error, morphological error, and structural error.

[0037] Feature importance evaluation unit: Use the partial dependence graph method to evaluate the importance of the model input features. According to the accuracy evaluation results, the optimal features for permanent erosion gully extraction are determined to obtain the optimal results of erosion gully extraction.

[0038] The data preprocessing and dataset creation unit of the gully extraction system for PlanetScope multi-temporal data and ALOS DEM data includes: data screening of Planet Basemap monthly-scale synthetic data, and screening of monthly data based on phenological consistency and data quality. First, data from the snow period from November to March are removed to avoid spatial heterogeneity, and low-quality images from the cloudy and rainy period in July are excluded. Finally, data from April to June and August to October (covering the spring plowing and autumn harvest periods) are retained. ALOS data are resampled to maintain the same resolution as Planet data. The filtered Planet Basemap monthly-scale synthetic data and the resampled ALOS DEM data are combined to generate an input feature set, which is labeled with corresponding labels. Label types include gully, road, and other categories. Finally, each sample feature set and label are combined to complete the dataset creation.

[0039] The described system for extracting gullies from PlanetScope multi-temporal data and ALOS DEM data includes a model training unit comprising: model input features consisting of Planet multi-temporal features and DEM features; a U-Net framework in which conventional two-dimensional convolutions are replaced with a dynamic snake convolution structure consisting of a conventional two-dimensional convolution and dynamic snake convolutions in the x and y directions; the model output is a three-category classification result, including gullies, roads, and other categories; and the model loss function is Dice loss. The model is subjected to 20-fold cross-validation, which randomly selects 20 different Planet data sets from the study area. The same number of samples are generated for each data set. One sample is used as a test set, and the remaining 19 samples are used as training sets.

[0040] The accuracy evaluation units for the gully extraction system using PlanetScope multi-temporal data and ALOS DEM data include volumetric error, morphological error, and structural error. The volumetric error uses the Dice coefficient, overall accuracy, precision, and recall metrics; the morphological error uses the Hausdorff distance metric based on persistent homology; and the structural error uses the SSIM metric.

[0041] The feature importance evaluation unit of the gully extraction system for PlanetScope multi-temporal data and ALOS DEM data includes a method for evaluating the importance of model input features using a partial dependence graph. This method involves assigning a monthly feature value of 0 after model training is complete and statistically analyzing changes in model accuracy. Based on the improvement (decrease) in model accuracy, the feature is determined to be redundant (valid). The valid monthly feature is retained as the final input feature selected by the model.

[0042] Those skilled in the art should understand that although the present invention is described in terms of multiple embodiments, not each embodiment contains only one independent technical solution. This description is provided for clarity only. Those skilled in the art should understand the description as a whole and consider the technical solutions involved in each embodiment as being combinable into different embodiments to understand the scope of protection of the present invention.

[0043] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. An erosion gully extraction method based on PlanetScope multi-temporal data and ALOS DEM data is characterized by: The following steps are involved: Step 1: Acquire multi-temporal remote sensing data and terrain data. The multi-temporal remote sensing data is monthly-scale synthetic data used to provide high temporal and spatial resolution features. Step 2: Preprocessing the multi-temporal remote sensing data and terrain data and preparing a data set, wherein the preprocessing includes data screening and resolution unification, and the data set preparation includes generating an input feature set and marking classification labels; Step 3: Design a multi-classification model based on dynamic convolution to extract landform features. The model uses a deep learning network as the backbone, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results; Step 4: Evaluate the accuracy of the training results of the multi-classification model using a variety of error indicators; Step 5: Determine the optimal input features through the feature importance evaluation method to obtain the final result of landform feature extraction.

2. The method according to claim 1, wherein In step 2, the multi-temporal remote sensing data and terrain data are pre-processed and a data set is prepared, including: The multi-temporal remote sensing data is screened, and the monthly data is screened based on the consistency of phenological characteristics and data quality, and the data of the spring plowing period and the autumn harvest period are retained; the terrain data is resampled so that its resolution is consistent with the multi-temporal remote sensing data; The filtered multi-temporal remote sensing data and the resampled terrain data are combined to generate an input feature set; For the input feature set, landform features, roads and other category labels are annotated to generate a training sample set.

3. The method according to claim 1, wherein In step 3, the design is based on a multi-classification model of dynamic convolution, including: Constructing a deep learning model with U-Net as the backbone, wherein the model input is multi-temporal remote sensing features and terrain features; The ordinary two-dimensional convolution in the U-Net is replaced by a dynamic snake convolution structure, which consists of ordinary convolution and dynamic convolution in the x and y directions; The Dice loss function is used to optimize the multi-classification model and output three classification results of landform features, roads and other categories; The multi-classification model was trained by 20-fold cross validation. Multiple data sets were randomly selected from the study area, one of which was used as a test set and the rest as training sets.

4. The method according to claim 1, wherein In step 4, the accuracy evaluation of the training results of the multi-classification model includes: The volume error was evaluated using Dice coefficient, overall accuracy, precision, and recall metrics; The morphological error is evaluated using the Hausdorff distance metric based on persistent homology; The SSIM index is used to evaluate the structural error; The extraction accuracy of the model is comprehensively judged based on the error indicators.

5. The method according to claim 1, wherein In step 5, determining the optimal input features by using a feature importance evaluation method includes: Using partial dependency graph method to evaluate the feature importance of the input feature set; By assigning the feature of a certain month to zero, the change in the accuracy of the statistical model is calculated; Judging the validity of the features based on the change in accuracy, and retaining valid features; The optimal input feature set is generated based on the retained features to determine the final result of landform feature extraction.

6. The method according to claim 2, wherein The screening of the multi-temporal remote sensing data includes: Snow cover period data were eliminated based on the consistency of phenological characteristics to avoid interference from spatial heterogeneity; Eliminate low-quality images during cloudy and rainy periods and retain high-definition images; Determine the monthly data corresponding to the spring plowing period and the autumn harvest period as the screening results; Convert the format of the filtered data to generate input features in a unified format.

7. The method according to claim 3, wherein The multi-classification model is trained by 20-fold cross validation, comprising: Multiple multi-temporal remote sensing data were randomly selected from the study area; Generate equal samples for each data set and divide them into training set and test set; Iteratively optimizing the multi-classification model using a training set; The generalization ability of the model is verified through the test set to generate training results.

8. A gully extraction system for PlanetScope multi-temporal data and ALOS DEM data, characterized by: Applying the method according to any one of claims 1 to 7, comprising: Data acquisition unit: acquires multi-temporal remote sensing data and terrain data. The multi-temporal remote sensing data is monthly-scale synthetic data used to provide high temporal and spatial resolution features; Data preprocessing and data set preparation unit: preprocessing the multi-temporal remote sensing data and terrain data and preparing data sets, wherein the preprocessing includes data screening and resolution unification, and the data set preparation includes generating input feature sets and marking classification labels; Model training unit: Design a multi-classification model based on dynamic convolution to extract landform features. The model uses a deep learning network as the backbone, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results; Accuracy evaluation unit: performs accuracy evaluation on the training results of the multi-classification model, using multiple error indicators for evaluation; Feature importance evaluation unit: Determine the optimal input features through the feature importance evaluation method to obtain the final result of landform feature extraction.

Images