Method and system for automatically identifying basins in DEM data based on neural network

Abstract

The invention belongs to the technical field of digital terrain analysis and artificial intelligence crossing, and particularly discloses an automatic identification method and system for basins in DEM data based on a neural network. The method comprises the following steps: adaptively determining a topographic relief amplitude analysis window based on a mean value point change method, and extracting an initial plain area by combining elevation, gradient and topographic relief amplitude; noise is eliminated through morphological optimization (hole filling and edge smoothing), a buffer area is constructed, and key terrain factors such as elevation and gradient change rate are screened to generate multi-channel feature data; dividing landform types by using ISODATA dynamic clustering, and outputting binary basin data; an improved ResUNet + + network (integrating multiple encoders, a cavity space pyramid and a channel attention mechanism) is adopted for end-to-end training, and refined segmentation of a complex boundary is achieved. The method supports cross-regional generalization testing, can realize high-precision basin identification in drought and moist landform scenes, and can be widely applied to the fields of resource exploration, ecological protection and disaster prevention and control.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of digital terrain analysis and artificial intelligence, and particularly relates to a fully automatic identification method for basin landforms in DEM (Digital Elevation Model) data based on a deep neural network, and more particularly to an adaptive extraction of basin boundaries and multi - scale feature fusion technology under complex terrains. Background Art

[0002] Landform is the comprehensive product of the long - term action of surface materials, tectonic movements and climate. Its type classification is the cornerstone of geographical research. As a typical tectonic landform unit, the basin has important strategic significance: its enclosed terrain forms a unique climate micro - environment (such as the agricultural ecology of the Sichuan Basin), the internal sedimentary layer is rich in oil, gas and salt lake resources (such as a certain basin), and its flat bottom is conducive to the development of urban agglomerations (such as the Guanzhong Basin). However, the basin boundary is often blurred due to tectonic activities and erosion effects (such as the mountain - plain transition zone), and it is difficult to accurately extract by traditional manual interpretation or semi - automatic methods.

[0003] The current automatic landform classification mainly relies on digital terrain analysis (DTA) technology, which classifies by setting thresholds of terrain factors (such as slope, relief). For example, Li Bingyuan et al. (2008) classified the landforms of China into 28 types based on altitude and relief. However, due to the significant differences in the three - step landforms in China, it is necessary to manually adjust local thresholds, making it difficult to adapt to complex terrain scenarios. The basin boundary extraction method proposed by Liu Dandan et al. (2023), which combines ISODATA clustering and morphological operations, although achieving semi - automation, still has significant limitations: this method requires presetting the number of ISODATA clusters, merging / splitting thresholds, and relying on empirical adjustment of morphological operations (such as the number of erosion and dilation times), resulting in high parameter sensitivity and poor generalization ability; at the same time, the calculation of terrain factors is easily interfered by DEM errors (such as pseudo - depressions, noise in mountain valleys), and the risk of boundary misjudgment is prominent; more critically, rule - based terrain indices (such as relief) cannot capture the texture features (such as the strike of fault zones) and context semantics (such as multi - stage sedimentary boundaries) of the basin edge, making it difficult to describe the non - linear spatial patterns of complex basins.

[0004] The above technical bottleneck stems from the "manual feature design + shallow model" paradigm of traditional methods: at the feature level, relying on expert knowledge to define terrain factors cannot achieve end-to-end learning of complex geomorphic semantics; at the model level, traditional clustering (such as ISODATA) and morphological operations only complete pixel-level classification and lack the semantic understanding ability of geomorphic boundaries, resulting in missed detection of small basins or over-segmentation of fragmented areas. Existing technologies are limited by manual parameter dependence, shallow feature expression, and insufficient anti-noise ability. There is an urgent need for a fully automatic solution that can adapt to multi-source terrain features and deeply integrate multi-scale context information to break through the technical bottleneck of fine extraction of basin boundaries in complex scenarios. Summary of the Invention

[0005] To solve the problems of strong dependence on manual parameters, limited feature expression, and prominent noise interference in existing technologies, the present invention provides an automatic identification method for basins in DEM data based on neural networks, which realizes high-precision and fully automatic extraction of basin boundaries through multi-feature fusion and deep learning technologies.

[0006] To achieve the above object, the technical solution provided by the present invention is: an automatic identification method for basins in DEM data based on neural networks, including the following steps:

[0007] 1) Based on DEM data, extract initial plain vector surface data;

[0008] 2) Perform mathematical morphological operations on the initial plain vector surface data to generate optimized plain boundary vector data;

[0009] 3) Construct a buffer along the plain boundary, calculate multi-terrain factors in the buffer, screen the feature combination with the highest information content based on the Theil entropy value method, and generate multi-channel feature data through normalization and band synthesis;

[0010] 4) Perform dynamic clustering on the multi-channel feature data, divide the plain and mountainous areas, and mosaically fuse with the initial plain data to output binary basin data;

[0011] 5) Crop boundary region slices based on the binary data to generate an RGB three-channel format elevation and slope data set;

[0012] 6) Construct a neural network and train it;

[0013] 7) Preprocess the DEM data of the area to be measured and input it into the trained network, and obtain the basin range result through inference and splicing.

[0014] Further, in step 1), based on the DEM data, the mean change point method is used to adaptively calculate the optimal window size for terrain undulation analysis, and the terrain undulation is calculated with this window; the slope data is calculated synchronously; combining the three types of data of elevation, slope and terrain undulation, each grid point is reclassified according to the preset landform classification standard, and the initial plain vector surface data is extracted.

[0015] Further, the mathematical morphology operations in step 2) include: area threshold setting, hole filling, and edge smoothing.

[0016] Further, the multi-topographic factors in step 3) include 8 topographic factors: elevation, slope, slope change rate, total accumulated curvature, terrain undulation, elevation coefficient of variation, surface roughness, and surface cutting depth. The She entropy method is used to evaluate the information content of the 8 factors, and the combination of the 5 items with the highest She entropy value is selected. The formula of the She entropy method is:

[0017]

[0018] where n represents the dimension, S h is the entropy of the n-dimensional data, and M S is the variance.

[0019] Further, in step 4), the ISODATA algorithm is used to dynamically cluster the multi-channel feature data, and the parameters of the ISODATA algorithm are set: the expected number of clusters, the inter-class distance threshold, the minimum number of samples, and the maximum number of iterations.

[0020] Further, the neural network is an improved ResUNet++ network, which is composed of an encoding module, a feature fusion module and a decoding module as a whole, and finally generates a binary classification result consistent with the input resolution and size;

[0021] Among them, the encoding module consists of two independent encoders, one for processing DEM data and the other for processing slope data. Each encoder contains an input layer and multiple downsampling layers; in the encoder for processing DEM data, the feature map after each downsampling is directly transmitted to the decoding module through a skip connection;

[0022] The processing process of the feature fusion module is: the encoded features from the two encoders are spliced, and feature mapping is performed through a residual module. Then, the features are processed by an atrous spatial pyramid pooling ASPP module and then passed into the decoding module;

[0023] The decoding module adopts a step-by-step upsampling strategy to gradually restore the spatial resolution of the feature map, obtaining the decoded high-level features. Finally, the decoded high-level features enter the output integration layer. The number of upsampling layers is the same as the number of downsampling layers. The processing process of the output integration layer is as follows: First, it passes through the Atrous Spatial Pyramid Pooling module (ASPP). Subsequently, the features processed by ASPP undergo channel compression through a 1×1 convolution. Finally, normalization is performed through the Sigmoid activation function to make the output result suitable for the binary classification task.

[0024] Furthermore, the input layer of each encoder is responsible for performing preliminary feature extraction on the input data, which consists of two 3×3 convolutional layers, batch normalization, and the ReLU activation function.

[0025] After the input layer, the encoder conducts deep feature learning through four downsampling layers. Each downsampling layer consists of multiple residual convolutional blocks. Each residual block contains two layers of 3×3 convolution, batch normalization, and the ReLU activation function. The input layer and the downsampling layers both adopt residual connections to alleviate the problem of gradient disappearance, and an SE module is set between each layer.

[0026] Furthermore, the encoding module is equipped with four upsampling layers. Before each upsampling layer, the input features are weighted through an attention mechanism to enhance the attention to important regions. Subsequently, the spatial resolution of the feature map is restored through an upsampling operation, and it is concatenated with the corresponding encoded features along the channel dimension.

[0027] The concatenated features are successively processed through two rounds of batch normalization, ReLU activation, and a 3×3 convolutional layer to obtain enhanced features. Finally, through a feature addition operation, the enhanced features are fused with the concatenated features.

[0028] Furthermore, the loss function of the neural network is the Binary Cross Entropy (BCE), and the formula is:

[0029]

[0030] where y i is the true label, p i is the probability value predicted by the neural network, and N is the number of samples.

[0031] The present invention also provides an automatic identification system for basins in DEM data based on a neural network, including:

[0032] A processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an automatic identification method for basins in DEM data based on a neural network as described in the above technical solution.

[0033] The present invention has the following beneficial effects:

[0034] 1) Advantage of feature fusion: By jointly extracting the initial plain using multiple factors such as elevation, slope, and terrain undulation, the problem of misclassification and missed classification of a single terrain factor (such as undulation) is overcome, the quality of training data is improved, and the subsequent computational complexity is reduced.

[0035] 2) Full-process automation: From data preprocessing, feature screening to neural network inference, no manual parameter intervention is required throughout the process, significantly improving the generalization ability of the algorithm.

[0036] 3) Fine-grained segmentation ability: Through the multi-scale feature fusion and attention mechanism of ResUNet++, it can effectively identify fuzzy boundaries (such as the mountain-plain transition zone) and suppress the interference of DEM noise (pseudo depressions, mountain valleys).

[0037] 4) Cross-scene adaptability: It supports the input of DEM data with different resolutions and can be extended to the automatic division of other geomorphic units (such as plateaus, hills). BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described in detail below with reference to the accompanying drawings.

[0039] Figure 1 is the flowchart of the present invention.

[0040] Figure 2 is the schematic diagram of the improved ResUNet++ network structure.

[0041] Figure 3 is the comparison chart of the initial plain extraction effect.

[0042] Figure 4 is the schematic diagram of the multi-feature fusion effect.

[0043] Figure 5 is the clustering analysis result chart of a certain basin buffer zone.

[0044] Figure 6 is the identification result chart of a certain basin.

[0045] Figure 7 is the identification result chart of a certain lake basin in the embodiment of the present invention, where (a) is the neural network inference result chart of a certain lake basin, (b) is the result chart of mosaicking the inference result with the initial plain, (c) is the result chart of hole filling and burr elimination, and (d) is the identification result chart of a certain lake basin. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0047] The automatic identification method of basins in DEM data based on neural network according to the present invention includes the following steps:

[0048] 1) Initial plain area extraction: Based on DEM data, the mean change point method is used to adaptively calculate the optimal window size for terrain undulation analysis, and the terrain undulation is calculated with this window; the slope data is calculated synchronously; combining the three types of data of elevation, slope and terrain undulation, each grid point is reclassified according to the preset landform classification standard (plain / non-plain), and the initial plain vector surface data is extracted;

[0049] 2) Data denoising and optimization: Perform mathematical morphological operations (including hole filling and edge smoothing) on the initial plain data, eliminate noise patches with an area smaller than the threshold, and remove irrelevant areas through spatial clipping;

[0050] 3) Buffer zone feature construction and screening: Generate a buffer zone with a preset width along the boundary of the initial plain, and calculate 8 terrain factors within the buffer zone, including elevation, slope, slope change rate, terrain undulation, total cumulative curvature, elevation coefficient of variation, surface roughness, and surface cutting depth; Based on the Shannon Entropy method, select the combination of 5 feature factors with the highest information entropy, and generate five-channel feature data through standardization and band synthesis;

[0051] 4) Landform type clustering analysis: Use the ISODATA unsupervised clustering algorithm to classify the five-channel feature data, set the threshold of the clustering center distance, dynamically merge or split clusters, and finally divide the grid points within the buffer zone into two landform types: mountain and plain;

[0052] 5) Basin data fusion and binarization: Spatially mosaic the plain area generated by clustering with the initial plain data, and after morphological denoising and hole filling, output the binarized basin data (plain is 1, non-plain is 0);

[0053] 6) Neural network dataset construction: Based on the binarized basin data, crop the boundary area slices with a 256×256 pixel sliding window (step size 128), synchronously generate the normalized elevation and slope data at the corresponding positions, synthesize the training samples according to the RGB three channels, and divide them into a training set and a validation set according to a 9:1 ratio;

[0054] 7) Improved ResUNet++ model training

[0055] The present invention proposes an improved ResUNet++ structure, which combines residual connection, Squeeze-and-Excitation (SE) channel attention mechanism, atrous spatial pyramid pooling (ASPP), and attention mechanism to enhance the ability to extract multi-scale features, improve the effectiveness of feature fusion, and optimize the feature recovery effect in the decoding stage. The overall network consists of an encoding module, a feature fusion module, and a decoding module, and finally generates a binary classification result consistent with the input resolution and size.

[0056] a) Encoding module

[0057] In the encoding stage, the network consists of two independent encoders, one for processing digital elevation model (DEM) data and the other for processing slope data. Each encoder contains an input layer and four downsampling layers.

[0058] Input layer: The input layer of each encoder is responsible for preliminary feature extraction of the input data. This layer consists of two 3×3 convolutional layers, batch normalization, and ReLU activation function, aiming to perform shallow feature extraction on the original input. This process enhances the network's perception ability of local details and provides a more stable initialization for deep feature learning. The output of the input layer is directly added to the input data after 3×3 convolution to form preliminary feature enhancement, so as to retain more original information and slow down information loss.

[0059] Downsampling layer: After the input layer, the encoder performs deep feature learning through four downsampling layers. Each downsampling layer consists of multiple residual convolutional blocks (Residual Conv Block). Each residual block contains two layers of 3×3 convolution, batch normalization, and ReLU activation function, and uses residual connection to alleviate the gradient disappearance problem and improve the feature expression ability.

[0060] Each downsampling layer introduces a Squeeze-and-Excitation (SE) module as an attention enhancement component, which adaptively adjusts the weights of different channels through global information, thereby enhancing the ability to express key features. The output of the SE module is fused with the feature map passing through the residual block through element-wise addition as the output feature of the current downsampling layer, further strengthening the response to key regions while retaining the original information.

[0061] Feature Transfer and Fusion: In the encoder for processing DEM data, the feature maps after downsampling in each layer are directly transferred to the decoder through skip connections to retain low-dimensional detailed features, which helps to effectively recover spatial detail information during the decoding stage.

[0062] b) Feature Fusion: In the feature fusion stage, the encoded features from different input sources are first concatenated and then feature-mapped through a residual module to enhance the fusion ability of different input modalities. Then, the features pass through an Atrous Spatial Pyramid Pooling (ASPP) module, which uses convolutional kernels with different dilation rates (dilation rates of 6, 12, and 18) to capture information at different scales, thereby improving the network's parsing ability for complex terrains or natural scenes. The multi-scale feature extraction mechanism of ASPP ensures that the network can still retain local detailed features under a large receptive field, improving the network's adaptability to targets at different scales.

[0063] c) In the decoding stage, the network adopts a progressive upsampling strategy to gradually restore the spatial resolution of the feature maps. A total of 4 upsampling layers are set. By introducing an attention mechanism, the model's ability to screen key features is enhanced, and the accuracy of detail restoration is improved.

[0064] Upsampling Layers: Each upsampling layer first weights the features through an attention mechanism to enhance the attention to important regions. Subsequently, the spatial resolution of the feature map is restored through an upsampling operation, and it is concatenated with the corresponding encoded features along the channel dimension. This concatenation helps to retain the spatial detail information in the low-level features and reduce information loss.

[0065] The fused features are successively processed through two batch normalizations, ReLU activations, and a 3×3 convolutional layer to extract more discriminative feature representations. Finally, the decoded features obtained after processing are fused with the skip connection features after channel concatenation through addition, and the result is used as the input for the next upsampling layer to ensure the continuous transfer of information and the effective restoration of details.

[0066] Output Integration Layer: In the final output stage, the decoded high-level features enter the output integration layer. First, through an Atrous Spatial Pyramid Pooling (ASPP) module, multi-scale atrous convolutions are used to capture context information under different receptive fields, thereby enhancing the network's perception ability for complex structures. Subsequently, the features processed by ASPP undergo 1×1 convolution for channel compression to reduce the computational amount and improve the generalization ability of the model. Finally, normalization is performed through a Sigmoid activation function to make the output results suitable for binary classification tasks.

[0067] The following are specific embodiments:

[0068] As Figure 1 shown, the automatic identification method of basins in DEM data based on neural network according to the present invention includes the following steps:

[0069] 1) Initial plain area extraction and geomorphic classification

[0070] Based on the DEM data, the mean change point method is used to adaptively determine the optimal window size for terrain undulation analysis. The specific operations are as follows:

[0071] a1) Optimal window calculation: Calculate the terrain undulation within the window range of 2×2 to 50×50, obtain the average terrain undulation of each window and divide it by the window area to get the unit terrain undulation; Based on the logarithmic characteristic curve of the terrain undulation, determine the inflection point through the mean change point method, and the corresponding window is the optimal analysis window.

[0072] a2) Slope calculation: Calculate the slope raster based on the DEM data.

[0073] a3) Geomorphic type classification: According to the geomorphic classification standard in Table 1 (plain when elevation < 3500m, slope < 8°, and terrain undulation < 70m), reclassify each grid point, assign 1 to the plain area and 0 to the non-plain area.

[0074] Table 1

[0075]

[0076] a4) Initial plain extraction: Conduct raster overlay analysis on the reclassification results of elevation, slope, and terrain undulation, and take the area with the largest numerical sum as the initial plain raster data.

[0077] 2) Data denoising and optimization

[0078] Perform morphological optimization on the initial plain raster data:

[0079] b1) Area threshold setting: Refer to the "Specifications for 1:1,000,000 Geomorphic Map of China", set the minimum mapping area of the plain within the mountain as 4 km 2 , and the strip-shaped plain as 6 km 2 ; Convert it to the pixel number threshold according to the DEM resolution.

[0080] b2) Hole filling: Traverse the binary data contour through Python script to eliminate the patches with an area smaller than the threshold.

[0081] b3) Edge smoothing: Perform multiple mathematical morphological operations on the binary image to eliminate the spike-shaped noise.

[0082] b4) Data export: Export the optimized data in raster and vector surface formats.

[0083] 3) Buffer Feature Construction and Screening

[0084] c1) Buffer Generation: Convert the initial plain vector surface into boundary line data, expand it inward and outward by a preset width (such as 30 km) to generate a buffer, and clip the DEM data within this area.

[0085] c2) Feature Factor Calculation: Calculate 8 topographic factors including elevation, slope, slope change rate, total accumulated curvature, terrain undulation degree, elevation variation coefficient, surface cutting depth, and surface roughness based on the clipped DEM.

[0086] c3) Feature Screening: Use the Theil entropy method to evaluate the information content of the 8 factors, and select the combination of the 5 factors with the highest entropy value (such as elevation, slope, slope change rate, total accumulated curvature, terrain undulation degree). The formula of the Theil entropy method is:

[0087]

[0088] where n represents the dimension, S h is the entropy of the n-dimensional data, and M S is the variance.

[0089] c4) Data Synthesis: Normalize the 5 features (Min-Max normalization) and synthesize five-band feature data. The normalization formula is:

[0090]

[0091] where O(i,j) represents the output gray value, I(i,j) represents the gray value of the i-th row and j-th column of the image, I mim represents the minimum gray value of the image, and I max represents the maximum gray value of the image.

[0092] 4) ISODATA Geomorphic Clustering Analysis

[0093] d1) Parameter Initialization: Set the expected number of clusters to 2, the minimum sample threshold to 20, the inter-class distance threshold to 5, and the maximum number of iterations to 10.

[0094] d2) Dynamic Clustering: Divide the buffer data into two categories, plain / mountain, through operations such as sample assignment, cluster deletion, center update, and split and merge, and remove the abnormal classes in the classification results.

[0095] 5) Basin Data Fusion and Binarization

[0096] Mosaic the clustered plain data with the initial plain, perform hole filling and noise filtering operations, and output binarized basin data (plain = 1, others = 0), which is stored in the GeoTIFF format.

[0097] 6) Neural Network Dataset Construction

[0098] e1) Boundary Slice Cropping: Use a 256×256 sliding window (stride 128) to traverse the binary data. When both 0 and 1 values exist in the window, crop the slice. The naming rule for the slice is "row letter + column number" (e.g., c56.tiff).

[0099] e2) Feature Data Generation: Calculate the elevation and slope data at the corresponding positions, normalize them, and render them into a three-channel RGB Tiff format.

[0100] e3) Dataset Partitioning: Randomly partition the slices into a training set and a validation set at a ratio of 9:1 to ensure balanced data distribution.

[0101] 7) ResUNet++ Model Training and Optimization

[0102] An improved ResUNet++ structure is proposed. Through multi-feature fusion and deep learning techniques, it realizes the high-precision and fully automatic extraction of basins in DEM data. The network consists of an encoding module, a feature fusion module, and a decoding module, and finally generates a binary classification result consistent with the input resolution and size. The network structure is as Figure 2 shown.

[0103] f1) Experimental Environment: The environment for network training uses the Windows 11 operating system, configures the NVIDIA RTX 4070 Ti GPU, develops and trains the model in the Python 3.9 environment using the Pytorch 2.4.1 deep learning framework, and utilizes CUDA 11.8 to fully utilize the GPU performance to accelerate training.

[0104] f2) Loss Function: In network training, the weighted binary cross-entropy (BCE) loss function, which is very effective for binary classification tasks, is adopted. Its formula is:[[]]

[0105]

[0106] where y i is the true label, p i is the probability value predicted by the model, and N is the number of samples. It can measure the difference between the predicted probability and the true label for each pixel or region. It drives the model optimization by maximizing the matching degree between the predicted probability and the true label. In the case of an imbalanced dataset, the BCE loss can further enhance the attention to difficult regions through a weighted strategy.

[0107] f3) Precision Evaluation Index: Select the Dice coefficient to measure the similarity between the predicted segmentation and the true segmentation by the model. The formula is:[[]]

[0108]

[0109] Among them, A is the region predicted by the model, B is the true boundary region, |A| and |B| are the areas of the predicted region and the true region respectively, and |A∩B| is the intersection area of the predicted region and the true region. The value of the Dice coefficient ranges between 0 and 1. The closer it is to 1, the greater the overlap between the model prediction and the true boundary, and the better the performance. In the boundary recognition task, the Dice coefficient can effectively evaluate the segmentation quality of the model, especially suitable for the case where the boundary region has details and irregular shapes.

[0110] f4) Network training adjustment and parameter setting: Use the Stochastic Gradient Descent (SGD) optimizer instead of the original Adam optimizer. Set the initial learning rate to 0.01, the momentum parameter to 0.9, and complete the network training with 200 epochs of training iterations.

[0111] 8) Automatic inference of basin range

[0112] Calculate the elevation and slope for the DEM data of the area to be measured. After slicing according to the same rules, input it into the trained ResUNet++ model. After the prediction results are stitched and processed by image processing, the final output is the basin boundary data.

[0113] In the embodiment of the present invention, a certain basin and its surrounding area are used as the test area. This area is located in the northeastern part of the Qinghai-Tibet Plateau (90°-99° E, 36°-39° N), with an average altitude of about 2,800 meters. The landform types cover mountains, gobi, salt lakes and wind-eroded yardangs. Its complex geological structure (controlled by the Altun fault zone and the Qilian mountain fold belt) and extremely arid climate (annual average precipitation <50 mm) result in strong boundary ambiguity and significant noise interference. This area is rich in mineral resources (such as lithium and potassium resources in the Qarhan Salt Lake) and has ecological fragility (the proportion of desertification area >70%), providing a typical scenario for verifying the technical advantages and engineering applicability of the method of the present invention in complex landform segmentation.

[0114] 1) Use the mean change point method to determine the optimal analysis window size as 9×9 based on the DEM data, calculate the terrain undulation degree therefrom, and then calculate the slope separately. Divide the geomorphic category of each grid according to the geomorphic division standard shown in Table 1, and set the slope threshold for the plain / mountain area to 8°. Mark the plain area as 1 and the mountain area as 0, and then add the three marked data through raster overlay operation. The area with the largest value is the initial plain area, and export this area as raster data.

[0115] 2) Convert the data obtained in the previous step into a binary image, and use mathematical morphology operations and image processing techniques for hole filling and edge burr smoothing. The specific operations are as follows:

[0116] Since there are many long, narrow mountain valleys between mountains or scattered small hills in the plains during the process of dividing mountains and plains, an area threshold of 6 km is considered for hole filling and noise filtering. Considering that the DEM data used has a resolution of 12.5 m, the corresponding pixel number threshold is 6000000 / 12.5 / 12.5 = 38400. 2 Use a Python program to fill in the patches on the binary image with fewer than 38400 pixels, and then perform multiple mathematical morphology operations on the binary image after hole filling to achieve the effect of removing edge burrs and smoothing the boundary. Export the processed initial plain boundary as vector line data and vector surface data for subsequent processing. The initial plain areas before and after processing are as

[0117] shown. Figure 3

[0118] 3) Based on the initial plain line data obtained in the previous step, create buffer zones with a width of 30 km on both the inside and outside, and use the clipping operation to obtain the 12.5 m resolution DEM data within the buffer zones. According to the DEM data, calculate 8 characteristic factors within the buffer zone, namely: elevation, slope, slope change rate, total accumulated curvature, terrain undulation degree, elevation coefficient of variation, surface roughness, and surface cutting depth.

[0119] Set the number of characteristic factor combinations to 5, and calculate the Theil entropy values for all possible combinations, that is, there are a total of combinations. Finally, the Theil entropy value of the combination (elevation, slope, slope change rate, total accumulated curvature, terrain undulation degree) is the highest at 21.44. Therefore, these five characteristic factors are selected as the characteristic data.

[0120] To eliminate the influence of dimensions, it is necessary to normalize the characteristic factors, and then perform band synthesis on the normalized characteristic factors to obtain five-channel characteristic data, as Figure 4 shown.

[0121] 4) Use the ISODATA classifier to perform clustering analysis on the characteristic data. First, preset the ISODATA clustering analysis parameters: the expected number of clusters is 2, the minimum sample threshold for a class is 20, the inter-class distance threshold is 5, and the maximum number of iterations is 10. The number of classes in the result after classification by the ISODTA classifier is generally more than the expected number of clusters. Therefore, use a Python script to automatically eliminate other classes, and finally obtain the classification result as Figure 5 shown.

[0122] 5) Mosaic the classification result obtained in the previous step with the initial plain area, assign the plain area a value of 1 and the mountain area a value of 0, and convert the raster basin data into vector line data. Finally, obtain the identification result map of a certain basin, as Figure 6 ​as shown

[0123] 6) Use a Python program to define a window with a size of 256×256 and a step size of 128 to traverse the binary classification image. If both pixel values 0 and 1 exist in the window, crop the area and save it in Tiff format. The cropped slices are named in the way of row number + column number.

[0124] Calculate the slope data of a certain area based on the DEM data, add the elevation and slope data to the ArcGIS software, convert the two types of data into an RGB three-channel format, and export it as a Tiff format.

[0125] Use a Python program to extract the geographic range of the data slices and save it in Shapefile format. Then, crop the elevation and slope data in RGB format according to the slice range vector data. Finally, data slices of each characteristic factor are obtained.

[0126] Use Python to randomly divide the characteristic factor data in a ratio of 9:1 to obtain the training set data and the validation set data respectively.

[0127] 7) Use the obtained training set data and validation set data to train the ResUNet++ network. The network uses the SGD optimizer and the learning rate is set to 0.01.

[0128] 8) To verify the generalization ability of the method of the present invention in a cross-regional scenario, a certain lake basin is selected as an independent test area in the embodiment. The landform type of this area is mainly a humid lake basin system, which is significantly different from the arid inland basin of the training area (a certain basin). Its dynamic lake shoreline, high vegetation coverage and ecological sensitivity characteristics pose higher requirements on the transfer learning ability and multi-source data fusion accuracy of the model.

[0129] Calculate the slope data based on the 12.5m resolution DEM data of a certain lake basin area, crop it into the required data slices according to the rules described above, input it into the neural network for testing. The effect after splicing the neural network output slices is as shown in Figure 7 (a) in, and then mosaic the neural network inference result with the internal initial plain area, as shown in Figure 7 (b) in, and then perform multiple morphological processing and hole filling operations on the boundary area to eliminate edge burrs and holes, as shown in Figure 7 (c) in, and finally convert the raster file into vector line data to finally obtain the identification result map of a certain lake basin, as shown in Figure 7 (d) in.

[0130] The test results show that the proposed method can still achieve high-precision extraction of geomorphic boundaries in a certain lake basin that was not involved in the training, which verifies its spatial adaptability and technical universality for complex geographical environments.

[0131] On the other hand, the embodiment of the present invention also provides an automatic identification system for basins in DEM data based on a neural network, including:

[0132] a processor and a memory, where the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an automatic identification method for basins in DEM data based on a neural network as described in the above technical solution.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic identification method for basins in DEM data based on a neural network, characterized in that, It includes the following steps: 1) Based on DEM data, extract the initial plain vector surface data; 2) Perform mathematical morphological operations on the initial plain vector surface data to generate optimized plain boundary vector data; 3) Construct a buffer along the plain boundary, calculate multiple terrain factors within the buffer, screen the feature combination with the highest information content based on the Xue's entropy value method, and generate multi-channel feature data through normalization and band synthesis; 4) Perform dynamic clustering on the multi-channel feature data, divide the plain and mountainous areas, and mosaic and fuse with the initial plain data to output binary basin data; 5) Crop the boundary region slices based on the binary data to generate elevation and slope datasets in RGB three-channel format; 6) Construct and train a neural network; 7) Preprocess the DEM data of the area to be measured and input it into the trained network, and obtain the basin range result through inference and stitching.

2. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: In step 1), based on the DEM data, the mean change point method is used to adaptively calculate the optimal window size for terrain undulation analysis, and the terrain undulation is calculated with this window; the slope data is calculated synchronously; combining the three types of data of elevation, slope and terrain undulation, each grid point is reclassified according to the preset landform classification standard to extract the initial plain vector surface data.

3. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: The mathematical morphological operations described in step 2) include: area threshold setting, hole filling, and edge smoothing.

4. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: The multiple terrain factors described in step 3) include 8 terrain factors: elevation, slope, slope change rate, total cumulative curvature, terrain undulation, elevation coefficient of variation, surface roughness and surface cutting depth. The Xue's entropy value method is used to evaluate the information content of the 8 factors, and the combination of the 5 items with the highest Xue's entropy value is selected. The formula of the Xue's entropy value method is: where n represents the dimension, and S h is the entropy of the n-dimensional data, and M S is the variance.

5. The automatic identification method of basins in DEM data based on a neural network according to claim 1, characterized in that: In step 4), the ISODATA algorithm is used to perform dynamic clustering on the multi-channel feature data, and the parameters of the ISODATA algorithm are set: expected number of clusters, inter-class distance threshold, minimum number of samples, and maximum number of iterations.

6. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: The neural network is an improved ResUNet++ network, which is composed of an encoding module, a feature fusion module and a decoding module as a whole, and finally generates a binary classification result with the same resolution and size as the input; Among them, the encoding module consists of two independent encoders, one for processing DEM data and the other for processing slope data. Each encoder contains an input layer and multiple downsampling layers; in the encoder for processing DEM data, the feature map after each downsampling is directly transmitted to the decoding module through a skip connection; The processing process of the feature fusion module is: splice the encoded features from the two encoders, perform feature mapping through a residual module, and then the features are processed by an atrous spatial pyramid pooling ASPP module and then passed into the decoding module; The decoding module adopts a step-by-step upsampling strategy to gradually restore the spatial resolution of the feature map, obtain the decoded high-level features, and finally the decoded high-level features enter the output integration layer; The number of upsampling layers is the same as that of downsampling layers; the processing process of the output integration layer is as follows: First, through the Atrous Spatial Pyramid Pooling module ASPP. Subsequently, the features processed by ASPP are compressed in channels through a 1×1 convolution. Finally, normalization is performed through the Sigmoid activation function to make the output result applicable to the binary classification task.

7. The automatic identification method of basins in DEM data based on neural network according to claim 6, characterized in that: The input layer of each encoder is responsible for preliminary feature extraction of the input data and consists of two 3×3 convolutional layers, batch normalization, and ReLU activation functions. After the input layer, the encoder performs deep feature learning through four downsampling layers. Each downsampling layer consists of multiple residual convolutional blocks. Each residual block contains two layers of 3×3 convolution, batch normalization, and ReLU activation functions. The input layer and the downsampling layers both use residual connections to alleviate the problem of gradient disappearance, and an SE module is set between each layer.

8. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: The encoding module is provided with four upsampling layers. Before each upsampling layer, the input features are weighted through an attention mechanism to enhance the attention to important regions. Subsequently, the spatial resolution of the feature map is restored through an upsampling operation, and it is concatenated with the corresponding encoded features in the channel dimension. The concatenated features are sequentially processed through two batch normalizations, ReLU activation, and 3×3 convolutional layers to obtain enhanced features. Finally, through a feature addition operation, the enhanced features are fused with the features after channel concatenation.

9. The automatic identification method of basins in DEM data based on neural network according to claim 1, characterized in that: The loss function of the neural network is Binary Cross Entropy BCE, and the formula is: where y i is the true label, p i is the probability value predicted by the neural network, and N is the number of samples.

10. An automatic identification system for basins in DEM data based on a neural network, characterized in that, Including: A processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an automatic identification method for basins in DEM data based on a neural network as described in any one of claims 1-9.

Images