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

Abstract

The present application provides an erosion gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data, relates to the field of erosion gully extraction, obtains PlanetScope multi-temporal data and ALOS DEM data; the obtained data is pretreated and a data set is made; an erosion gully data set is made; selected features are combined, training sample labels are labeled, and an erosion gully data set is generated; a multi-classification model based on dynamic snake convolution is designed to extract the erosion gully, 20-fold cross-validation is used to train the model; the model training result is evaluated in terms of accuracy; the importance of the features of the model input features is evaluated using the method of partial dependence plot, the optimal features of the erosion gully extraction are determined according to the accuracy evaluation result, and the optimal result of the erosion gully extraction is obtained. The present application realizes real-time extraction and monitoring of large-scale erosion gully, and provides data support for policy making for erosion gully treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of permanent gully extraction, in particular to a method for extracting permanent gully based on PlanetScope multi-temporal data and ALOS DEM data. BACKGROUND

[0002] Soil erosion is one of the core driving factors of land degradation, which is seriously threatening global food security and sustainable development of the ecological system. As the main form of soil erosion, the present monitoring of gullies is particularly important for the management of gullies and the protection of black soil resources. At present, Google Earth images are the main data source for gully extraction research, as they can provide global coverage of high-resolution (0.5m-1.2m) data. However, these studies generally ignore two key issues: first, when mapping large areas, the acquisition time of Google Earth images is inconsistent, resulting in significant differences in phenological characteristics between images, which seriously weakens the spatiotemporal transferability of the model; second, a large part of the images are historical data, which makes it difficult to meet the needs of present monitoring. In addition, existing gully extraction networks are mostly simple semantic segmentation models, which lack consideration of the unique morphological characteristics of gullies. Therefore, it is urgent to explore new data sources and models and design an extraction method that can be applied to large-scale monitoring and effectively consider the morphological characteristics of gullies.

[0003] The networking observation mode of the PlanetScope satellite constellation can synchronously acquire high spatiotemporal resolution data. In recent years, scholars have demonstrated its feasibility in gully extraction. The monthly scale synthesis data provided by Planet Basemap effectively reduces the phenological differences due to the relatively consistent data acquisition time; at the same time, its 4.77-meter resolution can meet the needs of large-scale gully monitoring, making it an ideal data source. Given that gullies usually exhibit a long and tubular shape, to enhance the model's ability to learn such features, the standard two-dimensional convolution structure is replaced by a dynamic snake convolution in the network. Dynamic snake convolution was originally designed to improve the accuracy of blood vessel structure extraction in medical images. As a kind of deformable convolution, it can more effectively learn the features of linear features. Therefore, the combined use of Planet Basemap multi-temporal data sources and the embedding of dynamic snake convolution structure into the gully extraction model is expected to significantly improve the accuracy of gully extraction and expand the scope of the study. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a method and system for extracting gullies based on 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 application provides the following scheme:

[0006] The present application provides an erosion gully extraction method based on PlanetScope multi-temporal data and ALOS DEM data, comprising the following steps:

[0007] Step one: obtain multi-temporal remote sensing data and terrain data, the multi-temporal remote sensing data is monthly scale synthesis data, which is used to provide high spatio-temporal resolution features;

[0008] Step two: pre-process and data set making of the multi-temporal remote sensing data and terrain data, the pre-processing includes data screening and resolution unification, the data set making includes generating input feature set and labeling classification label;

[0009] Step three: design a multi-classification model based on dynamic convolution for geomorphic feature extraction, the model takes a deep learning network as a skeleton, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results;

[0010] Step four: precision evaluation of the training results of the multi-classification model, using multiple error indicators for evaluation;

[0011] Step five: determine the optimal input feature through feature importance evaluation method, and obtain the final result of geomorphic feature extraction.

[0012] Further, the pre-processing and data set making of the multi-temporal remote sensing data and terrain data comprises:

[0013] Screening the multi-temporal remote sensing data, eliminating the monthly data with large phenology difference and low image quality, and retaining the spring planting period and autumn harvest period data; resampling the terrain data to make its resolution consistent with the multi-temporal remote sensing data;

[0014] Combining the screened multi-temporal remote sensing data and the resampled terrain data to generate an input feature set;

[0015] Labeling the geomorphic features, roads and other category labels for the input feature set to generate a training sample set.

[0016] Further, the design of the multi-classification model based on dynamic convolution comprises:

[0017] Building a deep learning model with U-Net as a skeleton, the model input is multi-temporal remote sensing features and terrain features;

[0018] Replace the ordinary two-dimensional convolution in the U-Net with a dynamic snake convolution structure, the dynamic snake convolution structure is composed of ordinary convolution and x, y direction dynamic convolution;

[0019] The multi-classification model is optimized by using a Dice loss function, and outputs three classification results of landform features, roads and other categories.

[0020] The multi-classification model is trained by 20-fold cross-validation, and multiple data are randomly selected from the study area, one of which is used as the test set and the rest are used as the training set.

[0021] Further, the training result of the multi-classification model is evaluated in precision, including:

[0022] The volume error is evaluated by using the Dice coefficient, overall accuracy, precision and recall rate indicators;

[0023] The morphological error is evaluated by using the Hausdorff distance indicator based on the continuous coherence;

[0024] The structural error is evaluated by using the SSIM indicator;

[0025] The extraction accuracy of the model is comprehensively judged according to the error indicators.

[0026] Further, the optimal input features are determined by the feature importance evaluation method, including:

[0027] The feature importance of the input feature set is evaluated by using the partial dependence plot method;

[0028] By assigning a certain month feature to zero, the change of model accuracy is counted;

[0029] According to the accuracy change, the effectiveness of the feature is judged, and the effective feature is retained;

[0030] According to the retained features, the optimal input feature set is generated, and the final result of landform feature extraction is determined.

[0031] Further, the multi-temporal remote sensing data is screened, including:

[0032] Based on the consistency of the phenological features, the snow period data is removed to avoid spatial heterogeneity interference;

[0033] Remove low-quality images in cloudy and rainy periods and retain high-definition images;

[0034] Determine the month data corresponding to the spring ploughing period and the autumn harvest period as the screening result;

[0035] The data after screening is converted into a unified format input feature.

[0036] Further, the multi-classification model is trained by 20-fold cross-validation, including:

[0037] Multiple multi-temporal remote sensing data are randomly selected from the study area;

[0038] Equal samples are generated for each data, and are divided into a training set and a test set;

[0039] The training set is used to iteratively optimize the multi-classification model;

[0040] The generalization ability of the model is verified by the test set, and a training result is generated.

[0041] Another object of the present application is to provide a PlanetScope multi-temporal data and ALOS DEM data erosion gully extraction system, comprising:

[0042] Data acquisition unit: acquiring multi-temporal remote sensing data and terrain data, the multi-temporal remote sensing data is monthly scale synthesis data, used to provide high spatio-temporal resolution features;

[0043] Data preprocessing and data set making unit: preprocessing and data set making of the multi-temporal remote sensing data and terrain data, the preprocessing includes data screening and resolution unification, the data set making includes generating input feature set and labeling classification label;

[0044] Model training unit: designing a multi-classification model based on dynamic convolution for geomorphic feature extraction, the model uses a deep learning network as a skeleton, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results;

[0045] Precision evaluation unit: precision evaluation of the training result of the multi-classification model, using multiple error indicators for evaluation;

[0046] Feature importance evaluation unit: determining the optimal input feature through a feature importance evaluation method to obtain the final result of geomorphic feature extraction.

[0047] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0048] The erosion gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data provided by the present application can realize large-scale erosion gully real-time extraction and monitoring, and provide necessary data support for the formulation of erosion gully management policies. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0050] Figure 1 Fig. 1 shows a flowchart of the erosion gully extraction method based on PlanetScope multi-temporal data and ALOS DEM data according to an embodiment of the present application;

[0051] Figure 2 Fig. 6 shows a diagram of feature importance results of 20 models based on partial dependence plots according to an embodiment of the present application;

[0052] Figure 3 Fig. 7 shows a diagram of model ablation experiment and comparative experiment extraction results according to an embodiment of the present application, wherein (a)-(f) are test set samples, white is correct classification result of different models, blue is missed points, and red is wrong points;

[0053] Figure 4 Fig. 8 shows a structural diagram of the erosion gully extraction system based on PlanetScope multi-temporal data and ALOS DEM data according to an embodiment of the present application;

[0054] Figure 5 Fig. 9 shows a diagram of the erosion gully extraction network structure according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0056] The purpose of the present application is to provide an erosion gully extraction method and system based on PlanetScope multi-temporal data and ALOS DEM data, which is used for large-scale extraction experiment of erosion gully.

[0057] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0058] As shown in Fig. 1, an erosion gully extraction method based on PlanetScope multi-temporal data and ALOS DEM data comprises: Figure 1 Step 1: Obtain multi-temporal remote sensing data and terrain data, wherein the multi-temporal remote sensing data is monthly scale synthetic data, which is used to provide high spatio-temporal resolution features;

[0059]

[0060] ​Step two: preprocessing and dataset making of the multi-temporal remote sensing data and terrain data, the preprocessing including data screening and resolution unification, the dataset making including generating input feature set and labeling classification label;

[0061] Step three: designing a multi-classification model based on dynamic convolution for geomorphic feature extraction, the model taking a deep learning network as a skeleton, replacing ordinary convolution with a dynamic convolution structure, and outputting multi-class classification results;

[0062] Step four: precision evaluation of the training results of the multi-classification model, using multiple error indicators for evaluation;

[0063] Step five: determining the optimal input feature through feature importance evaluation method to obtain the final result of geomorphic feature extraction.

[0064] Specifically, in one embodiment, step one: obtaining PlanetScope multi-temporal data and ALOS DEM data; wherein the PlanetScope multi-temporal data selects the monthly scale synthesis data of Planet Basemap.

[0065] Specifically, step two: preprocessing and dataset making of the obtained data; the preprocessing includes data screening and resampling, deleting the months of Planet Basemap monthly scale synthesis data with poor image quality and large differences in feature and phenology, and unifying the resolution of DEM data through resampling method; making an erosion gully dataset; combining the selected features, labeling the training sample label, and generating an erosion gully dataset, wherein the label includes erosion gully, road and other categories.

[0066] In one embodiment, the data preprocessing of the above-mentioned step two includes: data screening of Planet Basemap monthly scale synthesis data, based on the consistency of phenological features and data quality, screening the monthly data: first, eliminate the snow period data from November to the next March to avoid spatial heterogeneity interference, and exclude low-quality images in the cloudy and rainy period in July; finally, retain the data from April to June and August to October (this period covers the spring planting period and the autumn harvest period). Resample the ALOS data to make its resolution consistent with that of the Planet data.

[0067] In one embodiment, the dataset making of the above-mentioned step two includes combining the screened Planet Basemap monthly scale synthesis data and the resampled ALOS DEM data to generate an input feature set, and labeling the corresponding label at the same time, the label type including: erosion gully, road and other categories. Finally, combine each sample feature set and label together to complete the dataset making.

[0068] Specifically, step three: design a multi-classification model based on dynamic snake convolution for erosion gully extraction, and use 20-fold cross-validation for model training.

[0069] In one embodiment, the multi-classification model based on dynamic snake convolution described in step three above has a design structure (as shown in Figure 5 The input features of the model are Planet multi-temporal features and DEM features. The model uses U-Net as the skeleton, and the ordinary two-dimensional convolution is replaced by a dynamic snake convolution structure composed of an ordinary two-dimensional convolution and dynamic snake convolutions in x and y directions. The output of the model is a three-classification result, including erosion gully, road, and other categories. The loss function of the model is the Dice loss.

[0070] In one embodiment, the 20-fold cross-validation described in step three above randomly selects different 20 Planet data for the study area. The same number of samples are made for each data. One sample is used as the test set, and the other 19 samples are used as the training set.

[0071] Specifically, step four: precision evaluation of the model training results, including volume error, morphological error, and structural error.

[0072] In one embodiment, the precision evaluation indicators described in step four above include volume error, morphological error, and structural error. The volume error selects Dice coefficient, overall accuracy, precision, and recall rate indicators. The morphological error selects the Hausdorff distance based on persistent homology indicators. The structural error selects the SSIM indicator.

[0073] Specifically, step five: using the partial dependence plot method to evaluate the importance of the input features of the model, and according to the precision evaluation results, determine the optimal features for permanent erosion gully extraction, and obtain the optimal results of erosion gully extraction.

[0074] In one embodiment, the method of using partial dependence plot to evaluate the importance of the input features of the model described in step five above refers to assigning a certain month feature to 0 after the model training is completed, and then calculating the change of model precision. The results of the partial dependence plot are as shown in Figure 2 After assigning a certain feature to 0, all model precision evaluation indicators are worse than the median value of all models, which indicates that in this embodiment, DEM features and each month feature are effective for accurate extraction of erosion gully. They are all effective month features and are retained as the final input features of the model.

[0075] In the embodiments of the present application, the design ablation experiment and the comparative experiment system based on the target model are used to verify the effectiveness of the network design. The ablation experiment sets the benchmark U-Net architecture by removing the dynamic snake convolution structure, and compares it with the complete model (containing the dynamic snake convolution structure). The dynamic snake convolution structure is composed of a standard 3x3 convolution and an x and y direction dynamic snake convolution in parallel, and the feature sensitivity to the tubular ground object is enhanced through a learnable offset Δ ∈ [-1, 1]. The comparative experiment selects four typical segmentation models (SegNet, LinkNet, R2U-Net, and RNN-based model) as the comparative model. The ablation experiment results show that the introduction of the dynamic snake convolution structure significantly improves the extraction accuracy of the model by 11.2%. Compared with the precision index, the recall rate index decreases more, and the missing score phenomenon is more serious. This is mainly because the dynamic snake convolution structure is more likely to capture the details of the erosion ditch, and the extraction of the end or thinner part of the erosion ditch structure is more complete, and the missing rate is lower. By comparing the experimental results of the mainstream semantic segmentation models such as R2U-Net, SegNet and LinkNet, it is found that the DSCNet improves the Dice index by 13.2%, 19.5% and 27.5% respectively. The extraction accuracy of CNN is much higher than that of RNN, which reflects that the focus of this task is to pay attention to the spatial structure information of the erosion ditch, rather than the time sequence dependence, and verifies the inherent advantages of CNN in spatial feature extraction. In the CNN model, the extraction accuracy of RU2-Net is relatively high, mainly because the model combines the structure of recurrent neural network and residual connection based on U-Net, but the accuracy of the two is not much different, only less than 3%. The accuracy of SegNet is second, mainly because of the difference in feature fusion method between SegNet and U-Net. SegNet uses index pooling to restore spatial information, which makes the network save space while losing a lot of details. U-Net uses a jump connection to splice (Concatenation) the feature map and the corresponding layer of the decoder, which integrates deep semantic features and shallow detail features, and has better boundary segmentation accuracy. The accuracy of LinkeNet is the worst, especially the precision index. The main reason is that LinkeNet uses element-wise addition to fuse features, which is weaker in information retention than the splicing operation of U-Net while reducing the number of bands. In order to further illustrate the difference in model extraction results, Figure 3 The extraction results of different networks for part of the erosion ditch are shown. Blue represents missing, and red represents misclassification. Figure 3 The blue box in (a) is the end and thinner part of the erosion ditch. Except for the extraction result of DSCNet, which is more consistent with the true result, the other networks all have different degrees of missing. This shows that the dynamic snake convolution structure can better capture the geometric features of the erosion ditch.Figure 3 In the middle (b-d), each network has different degrees of misclassification, especially on the ridge (b) and the road (c-d), which are also the ground objects that are prone to misclassification, but DSCNet has less misclassification and higher extraction accuracy. Figure 3 In the results of the middle (e-f), DSCNet is more accurate in extracting erosion gullies, and other networks have different degrees of misclassification and omission. In addition, the extraction results of RNN are more fragmented, because RNN has strong learning ability of time sequence characteristics, but poor ability to capture spatial characteristics. This is because the time sequence characteristic data used in the present application is less, and RNN is not suitable for this problem. Table 1 is the precision evaluation results of the model ablation experiment and the comparison experiment of different models of the embodiment of the present application.

[0076] Table 1

[0077]

[0078] In addition, in another embodiment of the present application, a PlanetScope multi-temporal data and ALOS DEM data erosion gully extraction system is also provided, which is described with reference to Figure 4 which comprises:

[0079] The data acquisition unit acquires PlanetScope multi-temporal data and ALOS DEM data, wherein the PlanetScope multi-temporal data selects the monthly scale synthesis data of Planet Basemap.

[0080] The data preprocessing and dataset making unit pre-processes and makes datasets of the acquired data; the preprocessing includes data screening and resampling, deleting the month features of the Planet Basemap monthly scale synthesis data with poor image quality and large differences in the phenology of ground objects, and unifying the resolution of the DEM data by the resampling method; making an erosion gully dataset; combining the selected features, labeling the training sample labels, and generating the erosion gully dataset, wherein the labels include erosion gully, road and other categories.

[0081] The model training unit designs a multi-classification model based on dynamic snake convolution for erosion gully extraction, and uses 20-fold cross-validation to train the model.

[0082] The precision evaluation unit evaluates the model training results, and the evaluation index types include: volume error, shape error and structure error.

[0083] The feature importance evaluation unit uses the method of partial dependence plot to evaluate the importance of the features of the model input, determines the optimal features for permanent erosion gully extraction according to the precision evaluation results, and obtains the optimal results of erosion gully extraction.

[0084] The erosion gully extraction system of the PlanetScope multi-temporal data and the ALOS DEM data, the data preprocessing and the data set manufacturing unit include: data screening is carried out to the Planet Basemap monthly scale synthesis data, and the monthly data is screened based on the consistency of the phenological characteristics and the data quality: first, the data of the snow period from November to the following March is removed to avoid the interference of spatial heterogeneity, and the low-quality image of the cloudy and rainy period in July is excluded; finally, the data of April to June and August to October (the period covers the spring farming period and the autumn harvest period) is retained. The ALOS data is resampled to make the resolution consistent with the Planet data. The screened Planet Basemap monthly scale synthesis data and the resampled ALOS DEM data are combined to generate an input feature set, and the corresponding label is labeled, and the label type includes: erosion gully, road and other categories. Finally, each sample feature set and label are combined to complete the manufacturing of the data set.

[0085] The model training unit of the erosion gully extraction system of the PlanetScope multi-temporal data and the ALOS DEM data includes: the model input feature is the Planet multi-temporal feature and the DEM feature; the model takes U-Net as the skeleton, wherein the ordinary two-dimensional convolution is replaced by a dynamic snake convolution structure composed of an ordinary two-dimensional convolution, x and y two-direction dynamic snake convolution; the output of the model is a three-classification result, including: erosion gully, road and other categories; the loss function of the model is the Dice loss. The model is subjected to 20-fold cross-validation, and the 20-fold cross-validation is that different 20 pieces of Planet data are randomly selected in the study area. The same number of samples are manufactured for each data. Among them, 1 sample is used as a test set, and the other 19 samples are used as a training set.

[0086] The precision evaluation unit of the erosion gully extraction system of the PlanetScope multi-temporal data and the ALOS DEM data includes: volume error, shape error and structure error. Among them, the volume error selects the Dice coefficient, the overall accuracy, the precision and the recall rate index; the shape error selects the Hausdorff distance index based on the continuous coherence; and the structure error selects the SSIM index.

[0087] The feature importance evaluation unit of the erosion gully extraction system of the PlanetScope multi-temporal data and the ALOS DEM data includes: the method of using the partial dependence diagram is used to evaluate the importance of the features of the model input features, that is, after the model training is completed, a certain month feature is assigned to 0, and the change of the model precision is counted. The improvement (decrease) of the model precision is used to judge whether the feature is redundant (effective). The effective month feature is retained as the final selected input feature of the model.

[0088] Those skilled in the art will understand that, although the present application has been described in terms of a number of embodiments, not every embodiment according to the application will completely satisfy every object or requirement set forth in the description. The description has been presented for purposes of illustration and description only and is not intended to limit the scope of the application to the precise form described. It will be apparent to those skilled in the art that various modifications can be made to the form described without departing from the scope of the application as claimed.

[0089] Although the present application has been described in connection with the embodiments thereof with reference to the drawings, it is to be noted that the application is not limited to those precise embodiments and applications and that various changes and modifications can be made to the described embodiments by those skilled in the art without departing from the scope of the application as defined by the claims.

Claims

1. A method for extracting erosion trenches based on PlanetScope multi-temporal data and ALOSDEM data, characterized in that, Includes 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 spatiotemporal resolution features. Step 2: Preprocess the multi-temporal remote sensing data and terrain data and create a dataset. The preprocessing includes data filtering and resolution unification. The dataset creation includes generating an input feature set and labeling it with classification tags. 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 multiple error metrics. Step 5: Determine the optimal input features using a feature importance assessment method to obtain the final result of geomorphological feature extraction; Step two, the preprocessing and dataset creation of the multi-temporal remote sensing data and terrain data, includes: The multi-temporal remote sensing data were filtered, and the monthly data were filtered based on the consistency of phenological characteristics and data quality, retaining the data for the spring plowing period and the autumn harvest period; The terrain data is resampled to make its resolution consistent with the multi-temporal remote sensing data; The filtered multi-temporal remote sensing data and the resampled terrain data are combined to generate the input feature set; For the input feature set, landform features, roads, and other category labels are labeled to generate a training sample set; In step three, the design is based on a dynamic convolutional multi-classification model, including: A deep learning model with U-Net as its backbone is constructed, and the input of the model is multi-temporal remote sensing features and terrain features; The ordinary two-dimensional convolution in the U-Net is replaced with a dynamic serpentine convolution structure, which consists of ordinary convolution and dynamic convolution in the x and y directions; The multi-classification model is optimized using the Dice loss function, and the output results are classified into three categories: landform features, roads, and others. The multi-classification model was trained using 20-fold cross-validation. Multiple datasets were randomly selected from the study area, with one dataset used as the test set and the rest as the training set.

2. The method as described in claim 1, characterized in that, Step four, which involves evaluating the accuracy of the training results of the multi-classification model, includes: The Dice coefficient, overall precision, accuracy, and recall were used to evaluate volumetric error. The morphological error is evaluated using the Hausdorff distance index based on continuous cohomology. SSIM index is used to evaluate structural error; The extraction accuracy of the model is judged comprehensively based on the aforementioned error indicators.

3. The method as described in claim 1, characterized in that, Step five, which involves determining the optimal input features using a feature importance evaluation method, includes: The feature importance of the input feature set is evaluated using a partial dependency graph method; The change in the accuracy of the statistical model by assigning a feature value of zero for a certain month; The validity of the feature is determined based on the change in precision, and valid features are retained; The optimal input feature set is generated based on the retained features, and the final result of the landform feature extraction is determined.

4. The method as described in claim 1, characterized in that, The filtering of the multi-temporal remote sensing data includes: Data on snow cover periods were removed based on the consistency of phenological characteristics to avoid spatial heterogeneity interference; Low-quality images from cloudy and rainy periods are removed, while high-resolution images are retained; The monthly data corresponding to the spring planting season and the autumn harvest season were used as the filtering results; The filtered data is formatted to generate input features in a uniform format.

5. The method as described in claim 1, characterized in that, The training of the multi-class classification model using 20-fold cross-validation includes: Multiple multi-temporal remote sensing data were randomly selected from the study area; For each data set, generate an equal number of samples and divide them into a training set and a test set; The multi-classification model is iteratively optimized using the training set; The model's generalization ability is verified using a test set, and training results are generated.

6. A system for extracting erosion trenches from PlanetScope multi-temporal data and ALOSDEM data, characterized in that, The method described in any one of claims 1-5 comprises: Data acquisition unit: acquires multi-temporal remote sensing data and terrain data, wherein the multi-temporal remote sensing data is monthly scale synthetic data, used to provide high spatiotemporal resolution features; Data preprocessing and dataset creation unit: preprocesses the multi-temporal remote sensing data and terrain data and creates datasets. The preprocessing includes data filtering and resolution unification. The dataset creation includes generating input feature sets and labeling them with classification tags. 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 skeleton, replaces ordinary convolution with a dynamic convolution structure, and outputs multi-class classification results. Accuracy evaluation unit: Evaluates the accuracy of the training results of the multi-classification model, using multiple error metrics for assessment; Feature Importance Assessment Unit: The optimal input features are determined through feature importance assessment methods to obtain the final result of geomorphological feature extraction.

Images