Lake water surface extraction method based on multi-source data fusion and deep feature screening

Abstract

The invention relates to the field of lake water surface extraction, in particular to a lake water surface extraction method based on multi-source data fusion and deep feature screening. The method comprises the following steps: preprocessing multi-source remote sensing data, and eliminating differences among data sources; designing 1D-CNN and HRnet-AM feature extraction modules, and extracting spectral features, texture features and shape features from images of different data sources; constructing an adaptive weight quantization module A-WCM, performing fusion analysis on the multivariate features, and quantizing the weight; designing a redundancy elimination module (REM), and eliminating redundant or irrelevant features; carrying out layer-by-layer fusion and recombination on features extracted from each branch by adopting a deep semantic fusion technology; a decision tree optimal growth model OGMT is constructed, and the complexity of the model is effectively reduced through an optimal path selection and feature screening mechanism. According to the method, the recognition efficiency and accuracy of the lake water surface are improved, the lake water surface change can be more reliably captured and analyzed, and powerful support is provided for management and maintenance of a lake ecosystem.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing image processing, lake water surface extraction, and deep learning, and particularly to a lake water surface extraction method for multi-source data fusion and deep feature screening. Background Art

[0002] The use of remote sensing technology for lake water surface extraction has now become a crucial working means. However, in this process, it usually relies on a single image source, such as Landsat-8 satellite images, resource satellite images, high-resolution satellite images, synthetic aperture radar images, etc., to distinguish and extract based on the spectral characteristics, texture features, and shape contours shown by the lake in the remote sensing images. Although remote sensing technology has broad application prospects in lake water surface extraction, each image source has its inherent advantages and limitations. For example, although Landsat-8 satellite images are rich in spectral information, their spatial resolution is relatively low, which to a certain extent limits the accurate recognition of ground object textures. Resource satellites and high-resolution satellites both have high spatial resolution and can clearly capture surface details, but they have fewer bands, relatively scarce spectral information, and are easily restricted by satellite orbits and revisit cycles, and the continuity of data may be affected. In addition, in the lake areas represented by the eastern plains in China, precipitation is abundant, and due to the obvious cloud cover phenomenon, optical remote sensing satellites such as Landsat-8, resources, and high-resolution are difficult to meet the requirements of real-time observation and low cloud cover. Synthetic Aperture Radar (SAR) images have good cloud-penetrating ability and can provide good texture information for ground objects, but due to orbit and airspace limitations, there are problems of discreteness and missing in the observation angle and resolution of SAR images. On the other hand, the presence of speckle noise will also affect the clarity of SAR images. To solve the problems of low lake recognition accuracy and boundary mismatch caused by a single data source, experts and scholars at home and abroad have proposed methods of multi-source data fusion. However, the current research still stays at using traditional artificial methods to extract the data features after data fusion, and fails to fully utilize the advantages of multi-source data fusion, resulting in the extraction accuracy not meeting the expected requirements. Summary of the Invention

[0003] To solve the problem of difficult automatic extraction of lake water surfaces in the prior art, the present invention provides a lake water surface extraction method for multi-source data fusion and deep feature screening, which mainly includes:

[0004] S1: Obtain multi-source remote sensing image data, preprocess the multi-source remote sensing image data, and extract the features of the data; perform cutting processing on the preprocessed remote sensing image data to obtain image blocks, and at the same time construct a training set;

[0005] S2: Construct an MD-DFE neural network model, which includes a 1D-CNN and an HRnet-AM feature extraction module with differential parallelism, an adaptive weight quantization module A-WCM, and a redundant feature elimination module REM. The 1D-CNN and HRnet-AM feature extraction modules are used to extract semantic features of different image blocks; the adaptive weight quantization module A-WCM is used to quantify the weights of different semantic features, and the redundant feature elimination module REM is used to eliminate redundancy from the extracted multi-dimensional features;

[0006] S3: Implement multi-source semantic feature fusion using deep semantic fusion technology;

[0007] S4: Use the training set to train the MD-DFE neural network model, extract the fused semantic features, input the fused semantic features into the optimal growth model of decision tree OGMT, obtain nodes and thresholds through the optimal path selection and feature screening mechanism, construct a decision tree classification model, and use the decision tree classification model to achieve automatic extraction of lake water surfaces.

[0008] A storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the steps of the lake water surface extraction method for multi-source data fusion and deep feature screening.

[0009] The beneficial effects brought by the technical solution provided by the present invention are as follows: In the aspect of lake water surface extraction, the existing technologies generally face problems such as low recognition accuracy, boundary mismatch, and insufficient robustness caused by single data sources, which seriously affect the accuracy and intelligent level of lake water surface extraction. To overcome these technical bottlenecks, the present invention proposes a solution for multi-source data fusion and deep feature screening. This method uses Landsat-8 series images, Sentinel-1 SAR images, resource satellite images, and high-resolution satellite images, making full use of the advantages of the high spatial resolution of high-resolution images, the spectral information of multi-spectral images, and the texture feature data of SAR images, thereby solving various problems brought by single data sources and achieving high-precision, intelligent, and automatic extraction of lake water surfaces, providing scientific data support for understanding and predicting the future change trends of regional water balance and ecosystem. Description of the Drawings

[0010] The following will further illustrate the present invention in conjunction with the drawings. In the drawings:

[0011] Figure 1 is a flowchart of a lake water surface extraction method for multi-source data fusion and deep feature screening in an embodiment of the present invention;

[0012] Figure 2 is a schematic diagram of the MD-DFE neural network model in an embodiment of the present invention;

[0013] Figure 3 It is a schematic diagram of the 1D-CNN feature extraction network in the embodiment of the present invention;

[0014] Figure 4 It is a schematic diagram of the HRne-AM network structure in the embodiment of the present invention;

[0015] Figure 5 It is a schematic diagram of the attention mechanism module in the embodiment of the present invention;

[0016] Figure 6 It is a schematic diagram of the redundancy elimination module in the embodiment of the present invention;

[0017] Figure 7 It is a schematic diagram of the adaptive weight quantization module in the embodiment of the present invention; Detailed implementation manners

[0018] For a clearer understanding of the technical features, objectives, and effects of the present invention, the specific implementation manners of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Embodiment 1

[0020] Please refer to Figure 1 , Figure 1 which is a flowchart of a lake water surface extraction method for multi-source data fusion and deep feature screening in the embodiment of the present invention. The deep semantic fusion technology is used to realize the fusion of multi-source semantic features. Finally, a sample data set is made to train the MD-DFE (Multi-source data fusion and deep feature extraction) neural network model, and the fused multi-source features are input into the optimal growth model OGMT of the decision tree to realize the automatic extraction of the lake water surface. The method specifically includes:

[0021] S1: Obtain Landsat-8, resources, high-resolution, synthetic aperture radar and other satellite images respectively to collect multi-source remote sensing data, preprocess the multi-source remote sensing data to eliminate the differences between data sources and ensure the spatial alignment between images; and use high-spatial-resolution satellite data such as high-resolution and resources for segmentation to obtain image blocks.

[0022] S2: Construct an MD-DFE (Multi-source data fusion and deep feature extraction) neural network model, design a differentiated parallel 1D-CNN (1Dimensionality-Convolutional Neural Network) and HRnet-AM (High Resolution Net-Attention Mechanism) feature extraction module to extract the deep semantic features of multi-source data, that is, extract spectral features, texture features and shape features from images of different data sources to maximize the retention of valid information from different data sources; determine the weights of different deep semantic features by constructing an adaptive weight quantization module A-WCM (Adaptive Weight Calculation Module) to improve the model's attention to lake information and reduce the interference of background information; at the same time, use a redundant feature elimination module REM (Redundant Elimination Module) based on the correlation matrix to eliminate redundant or irrelevant features from the extracted multi-dimensional features; realize the precise screening and effective fusion of multi-source data features.

[0023] S3: Adopt deep semantic fusion technology to layer-by-layer fuse and reorganize the features extracted by each branch and form a unified feature representation;

[0024] S4: Based on the fused and reorganized features, use the constructed optimal growth model for decision trees OGMT (Optimal Growth Model for Decision Trees) to accurately classify the fused features, thereby realizing the automatic extraction of lakes. Through the optimal path selection and feature screening mechanism, the complexity of the model is effectively reduced.

[0025] This embodiment can be implemented in the Datong Lake area of Hunan Province, and the specific content is as follows:

[0026] A. Obtain multi-source remote sensing image data in the mining area, preprocess it and construct a multi-source dataset:

[0027] A1. Acquisition of multi-source remote sensing image data: Download multi-source data according to information such as the region where the lake is located, phenological characteristics (dry season, wet season), and field investigation time. For optical image data, select Landsat-8 multi-spectral remote sensing images with as little cloud or less cloud cover as possible, and satellite data images such as high-resolution and resource satellites; for optical image data, select one Landsat-8-L1C level multi-spectral remote sensing image with the least cloud cover and one high-resolution No. 1 remote sensing image containing four bands of red, green, blue, and near-infrared; for SAR image data, select the interferometric wide-swath mode (IW) dual-polarization ground range images of Sentinel-1 on adjacent dates, and preferably select VV and VH dual-polarization data to enhance the accuracy of ground object classification.

[0028] A2. Preprocessing of multi-source image data: For optical image data, use the "RadiometricCalibration" tool in ENVI software and input relevant calibration parameters such as gain, offset value, and solar altitude angle. First, convert the digital number (DN) values of the image to surface reflectance to eliminate the influence of sensor characteristics and shooting conditions on the image data. Second, use the "FLAASH" module to eliminate the influence of atmospheric scattering and absorption during the image imaging process. Third, perform band resampling through the "Resample" tool. Users can select interpolation methods such as "Nearest Neighbor" or "BilinearInterpolation" to resample the image to a unified 2-meter spatial resolution, which can ensure the consistency of multi-source data at the spatial scale and facilitate subsequent fusion and analysis. Finally, use the "Geometric Correction" module to achieve geometric correction through ground control points (GCPs). Manually select several obvious geographical feature points in the image (such as road intersections, bridges, etc.), and then match them with the corresponding points in the standard geographical coordinate system to ensure the accurate spatial position of the image. For SAR image data, first use the Sen2Cor plugin developed by ESA to perform radiometric calibration on the data to eliminate the interference of the atmosphere on the signal and obtain more accurate surface reflectance characteristics. Second, to keep the data format and resolution of SAR images consistent with those of optical images, use the "Resampling" tool in SNAP (Sentinel Application Platform) software to resample SAR to the same 2-meter resolution as optical images, ensuring the unity of SAR and optical images in spatial resolution. Third, download and apply the latest orbit file through the "Orbit File Correction" tool to automatically correct the orbit of SAR images and ensure their spatial alignment with other remote sensing data. Finally, remove the thermal infrared noise and filter the speckle noise of SAR image data to further improve the image quality and obtain clearer texture features.

[0029] A3. Feature Extraction and Calculation of Multi-Source Remote Sensing Image Data: It mainly includes the extraction and calculation of spectral features and texture features. For spectral features, 9 spectral bands and 8 spectral indices after preprocessing of Landsat-8 multi-spectral remote sensing images are obtained respectively, namely coastal band (B1), blue band (B2), green band (B3), red band (B4), near-infrared band (B5), short-wave infrared band 1 (B6), short-wave infrared band 2 (B7), thermal infrared band (B8), panchromatic band (B8), cirrus band (B9), Normalized Differential Water Index (NDWI), Modified Normal Differential Water Index (MNDWI), Automated Water Extraction Index (AWEI), Modified Automated Water Extraction Index (MAWEI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Moisture Index (NDMI), Normalized Difference Soil Index (NSBI), Enhanced Water Index (EWI). For texture features, first, 2 backscattering coefficients VH and VV of two polarization modes of SAR images and Sentinel-1 Dual-Polarized Water Index (SDWI) are obtained. Then, a 5×5 filter window is selected, and the feature means (Mean), variances (Variance), homogeneities (Homogeneity), contrasts (Contrast), dissimilarities (Dissimilarity), entropies (Entropy), angular second moments (Angular Second Moment), and correlations (Correlation) at each angle are calculated using the Grey Level Co-occurrence Matrix (GLCM) method.

[0030] Table 1: Image Features

[0031]

[0032]

[0033] Multi-scale segmentation of high-resolution images: Using the Ecognition software and its powerful Object-Based Image Analysis (OBIA) function, the boundary and morphological features of ground objects are extracted through multi-scale segmentation technology to make them more compact and closer to the actual lake morphology. First, in the "Multiresolution Segmentation" module, adjust the "Scale Parameter" to set the segmentation scale, and optimize the segmentation results by adjusting the Compactness and Smoothness parameters to make the shape of the object more compact and closer to the actual lake morphology. The algorithm parameters are set as follows: the scale parameter is 45, the shape factor is 0.35, the color factor is 0.6, the compactness factor is 0.5, and the smoothness factor is 0.5.

[0034] A5. Dataset construction: The multispectral image extracts 9 bands and 8 spectral features, a total of 17 variables; the SAR image extracts 2 backscattering coefficients and 9 texture features, a total of 11 variables; the high-resolution image and the resource satellite image obtain image objects through multi-scale segmentation. For each segmented object obtained after segmentation, calculate the mean value of each variable extracted from the multispectral image and the SAR image within each object to obtain two column vectors. Then, create the minimum bounding rectangle of each segmented object for the result of the multi-scale segmentation of the high-resolution image, thereby generating high-resolution image blocks for the extraction of deep features.

[0035] B. Construction of the MD-DFE neural network model: As Figure 2 shown, the specific steps are as follows.

[0036] B1. Build the MD-DFE neural network model, that is, build a hierarchical parallel convolutional neural network for extracting deep semantic features from multi-source datasets. The extraction process is as follows: (1) Taking pixels as object units, use 1D-CNN to process the two feature column vectors extracted from Landsat-8 and Sentinel-1 images respectively. This network is designed with 4 layers, including 2 convolutional layers, 1 flattening layer, and 1 fully connected layer (as Figure 3 ). Among them, convolutional layer 1 and 2 contain 4 and 8 convolutional kernels with a size of 5 and a stride of 1 respectively, and the activation function is LeakyReLU. Each convolutional layer will generate feature maps with the same number as the number of convolutional kernels. All the feature maps of each initial input vector will be merged into 1 one-dimensional column vector through the flattening layer. Finally, after randomly discarding some neurons through the Dropout regularization module in the fully connected layer, the features are re-extracted, and the deep semantic features obtained from the feature column vectors are output. (2) For the processing of high-resolution image blocks, in order to obtain multi-scale information, it is necessary to expand the receptive field. Therefore, the present invention designs asFigure 4 The HRnet-AM feature extraction module shown in the figure. The main body of HRnet-AM consists of three basic blocks. The Bottleneck module is responsible for enhancing the expressive power of features through convolutional operations and batch normalization. The BasicBlock is a simpler network structure, usually used to build deeper neural networks. The HighResolutionModule is dedicated to maintaining and fusing feature maps of different resolutions. Outside the three basic blocks, the present invention introduces an attention mechanism module, Attention mechanism, to improve the overall focus of the model and enhance the context understanding ability (such as Figure 5 ). Specifically, the structure for ensuring the retention and utilization of high-resolution information includes 4 stages and 4 parallel convolutional stream branches. Taking the convolutional stream of high-resolution remote sensing images as the initial stage, gradually adding the high-resolution convolutional stream to the low-resolution convolutional stream, and connecting the multi-resolution convolutional streams to form a new stage. High-resolution and low-resolution features are always retained in each stage, and cross-resolution information is repeatedly exchanged through the multi-scale fusion module to avoid information loss caused by sampling.

[0037] B2. Design the feature redundancy elimination module as shown in Figure 6 the figure. Considering that there may be duplicate problems in the spectral, texture and other features extracted from multi-source image data, in order to improve the operation speed of the model, the present invention designs a feature redundancy elimination module. For the extracted multi-source features, first compress and eliminate irrelevant or redundant features, and then expand to obtain the fused multi-source data features required for classification.

[0038] B3. Design the adaptive weight quantization module as shown in Figure 7 the figure. Multiply the weight values by the deep semantic feature column vectors extracted by hierarchical parallel convolution, and use the Concat method to fuse the weighted feature column vectors, thereby increasing the feature dimension for describing the image, but the amount of information under each dimension feature remains unchanged.

[0039] B4. Design the multi-source feature deep fusion module. Input the object multi-source deep semantic features extracted in the above steps into the deep feature fusion module for multi-source feature fusion. First, a new adaptive weight quantization module is designed to obtain the feature weights for each dimension. Specifically, the deep semantic features extracted from multi-source data are rearranged, and then global average pooling is used to average the rearranged deep semantic features in units of three feature column vectors, and then the weight values of the importance of multi-source data features are calculated through two fully connected layers.

[0040] C. Use the training set and the validation set to train the MD-DFE neural network model, extract the fused deep features, and use the OGMT classifier for lake extraction.

[0041] C1. Make a sample data set for lake classification. First, annotate the remote sensing image data of the lake area. The annotation objects are water bodies and non-water bodies respectively, and a binary classification label image of water bodies and non-water bodies is obtained. To prevent memory overflow, the image and the corresponding label image are subjected to sliding cropping, and the size after cropping is 256×256 pixels. Data augmentation is performed on the cropped original and label images, including 180° rotation, horizontal flipping, and vertical flipping operations to obtain sufficient samples and avoid overfitting during the training process. Then, the sample data is randomly divided into a training set, a validation set, and a test set according to 6:2:2.

[0042] C2. Train the constructed hierarchical parallel neural network model. First, add a fully connected layer to the hierarchical parallel convolutional neural network to output classification category information, and then use the training set constructed in step C1 to train the network, and use the validation set to test the model accuracy after each iterative training. In the model training process of the present invention, a cross-entropy loss function is adopted, and the optimizer is selected as Adam. The calculation method of the cross-entropy loss function is:

[0043]

[0044] where y represents the true category, represents the model prediction category, S is the number of image objects in the validation set, and L is the model loss value, which represents the error size of the model.

[0045] C3. Set the number of iterations (iteration) epoch. In this embodiment, the number of epochs = 200, and multiple iterative trainings are performed. In each iteration, the gradient descent method is used to update the model parameters and reduce the loss value of the model. At the same time, the parameters of the model are optimized and updated (the parameters of the model refer to the weight values of the connections between layers in the neural network). The settings of the network parameters are shown in Table 2, and the configuration of the server is shown in Table 3.

[0046] C4. Construct the Optimal Growth Model of Decision Tree (OGMT). Through the cyclic analysis of the training data set composed of test variables and target variables, a decision tree structure in the form of a binary tree is constructed. The Gini Index is used as the criterion for selecting the most attribute features and division thresholds. Specifically, first, initialize the root node, regarding all samples as a node, which serves as the root node of the entire tree: calculate the Gini Index of each feature for the current data set, and select the feature with the smallest Gini Index as the criterion for node splitting. Secondly, split the current feature into two child nodes according to the selected best cut-off point, and each child node represents a value of the feature. After each division, new root nodes and child nodes are generated and added to the decision tree. Thirdly, repeat the above steps for the new nodes until the stopping condition is reached (stopping condition: stop splitting when the samples in each child node belong to the same category or reach the preset threshold). Finally, the OGMT classifier is obtained.

[0047] C5. Input the fused deep feature sample data of image objects into the OGMT classifier, train the OGMT classifier with the training set, obtain the nodes and thresholds, construct a decision tree classification model, and realize the extraction of lake water surfaces through the trained OGMT model.

[0048] Table 2 Network Parameter Settings

[0049]

[0050]

[0051] Table 3 Server Configuration

[0052]

[0053] The experiment uses four classification evaluation indicators, namely cartographic accuracy, user accuracy, overall accuracy, and Kappa coefficient, to test the model accuracy. After multiple iterative trainings, the model with the highest accuracy is selected for fine classification of mining area land use.

[0054] D. Apply the trained model for automatic lake extraction: The specific steps are as follows:

[0055] D1. For the obtained remote sensing image objects, calculate their spectral and texture features respectively and extract them as column vectors. Calculate the spectral and texture feature vectors of all unclassified image objects, construct a multi-source data set, then input it into the trained hierarchical parallel convolutional neural network model to obtain the fused deep features of each image object, and finally input it into the OGMT classifier for classification to realize the automatic extraction of lake water surfaces.

[0056] D2. Use the test set and validation set to evaluate the accuracy of the lake extraction results. Construct a confusion matrix and conduct accuracy evaluation through indicators such as Precision (P), Recall (R), Misclassification Rate (MRate), and Mean Intersection over Union (MIoU). Apply the trained model to perform lake water surface extraction work.

[0057] The meanings of each indicator are as follows:

[0058] Precision (P): Also known as the precision rate, it represents the proportion of correctly classified water body samples to the number of samples predicted as water bodies. The higher its value, the higher the extraction accuracy.

[0059] Recall (R): Also known as the recall rate, it represents the proportion of correctly classified lake samples to all real lake samples. The higher its value, the fewer missed classifications.

[0060] Misclassification Rate (MRate): It represents the ratio of all misclassified samples to the total samples.

[0061] Mean Intersection over Union (MIoU): It refers to the mean of the ratios of the intersections and unions of the true values and the predicted values obtained by the algorithm for all classes.

[0062] The key points of the present invention are:

[0063] 1) The present invention comprehensively utilizes multi-spectral images, SAR images, and high-resolution and resource images to construct a multi-source data set containing spectral and texture feature column vectors and high-resolution remote sensing image blocks, reducing the limitations in aspects such as geometry, spectrum, and spatial resolution of the image data obtained by different sensors. Aiming at the pain points that it is difficult to accurately extract the lake area, boundary, etc. from a single image, multi-source data can provide rich spectral, texture features, and detailed information, improving the extraction accuracy.

[0064] 2) The present invention designs 1D-CNN and HRnet-AM feature extraction modules, which can respectively extract spectral, texture, and shape features from different data sources. The features of each data source are processed through independent modules, thus maximizing the retention of effective information. This modular design not only enhances the recognition accuracy of lake boundaries and morphological features but also improves the extraction ability for complex water body features, laying a high-quality data foundation for feature fusion and classification.

[0065] 3) The present invention constructs an Adaptive Weight Quantization Module (A-WCM). By performing fusion analysis and weight quantization on multi-source features, it effectively highlights lake-related features and reduces the interference of background information. In addition, to optimize the feature dimension, a Redundancy Elimination Module (REM) is specially set up to screen and eliminate redundant or irrelevant features. This process not only reduces the complexity of the model but also significantly improves the operation efficiency, enabling the model to achieve feature fusion and selection more quickly when processing large-scale remote sensing data.

[0066] 4) The present invention adopts a deep semantic fusion technology to layer-by-layer fuse and reorganize the features of each data source, enabling them to be fully integrated in the semantic space to form a unified feature representation. At the same time, an Optimal Growth Model of Decision Tree (OGMT) is designed. According to the differences and weights of the features, it adaptively selects the best classification path, effectively reducing the complexity of the model while improving the classification efficiency, and realizing the rapid, accurate, and automated extraction of lake information. This model not only has high precision in lake boundary recognition but also greatly shortens the processing time, and is applicable to large-scale lake extraction and monitoring tasks.

[0067] The present invention can replace traditional means such as visual interpretation and machine learning algorithms to achieve the automated extraction of lakes, taking into account both the efficiency and accuracy of lake extraction, and can be widely applied to lake extraction work, which is of great significance for regional water balance deduction and the construction of ecological environment monitoring systems.

[0068] Embodiment 2

[0069] A storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, it realizes the steps of the lake water surface extraction method for multi-source data fusion and deep feature screening described above.

[0070] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A lake surface extraction method based on multi-source data fusion and deep feature screening, characterized in that: include: S1: Acquire multi-source remote sensing image data, and pre-process the multi-source remote sensing image data to extract spectral features and texture features of the data; perform cutting processing on the pre-processed remote sensing image data to obtain image blocks, and construct a training set at the same time; S2: Construct an MD-DFE neural network model, which includes differentiated parallel 1D-CNN and HRnet-AM feature extraction modules, adaptive weight quantization module A-WCM, and redundant feature elimination module REM. The 1D-CNN and HRnet-AM feature extraction modules are used to extract semantic features of different image blocks; the adaptive weight quantization module A-WCM is used to quantify the weights of different semantic features; and the redundant feature elimination module REM is used to eliminate the redundancy of the extracted multi-dimensional features; S3: Adopt multi-source feature fusion module to realize multi-source semantic feature fusion based on vector product and Concat algorithm; S4: Use the training set to train the MD-DFE neural network model, extract the fused semantic features, input the fused semantic features into the decision tree optimal growth model OGMT, obtain nodes and thresholds through the optimal path selection and feature screening mechanism, build a decision tree classification model, and use the decision tree classification model to realize the automatic extraction of lake water surface.

2. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S1, multi-source remote sensing image data include Landsat image data, Sentinel image data, Gaofen and resource satellite image data. Multi-source data, optical image data, and SAR image data are downloaded according to the lake location, phenological characteristics, and field survey time. The optical image data are Landsat-8 multispectral remote sensing images and high-resolution and resource satellite data images. The SAR image data are VV+VH dual-polarization data to enhance the accuracy of ground object classification.

3. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S1, the preprocessing is: For optical image data, the DN value of the image is first converted into surface reflectivity to eliminate the influence of atmospheric scattering and absorption on the image during the imaging process; then, the image is resampled to a uniform spatial resolution to ensure the consistency of multi-source data in spatial scale, which is convenient for subsequent fusion and analysis; finally, geometric correction is achieved through GCPs; For SAR image data, radiation correction is first performed, then resampled to a resolution consistent with the optical image, the latest orbit file is downloaded and applied, and the SAR image is automatically corrected for orbit. Finally, the SAR image data is subjected to thermal infrared noise removal and speckle noise filtering to improve image quality and obtain clearer texture features.

4. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S2, the semantic feature extraction and calculation of multi-source remote sensing image data mainly include the extraction and calculation of spectral features and texture features. For spectral features, multiple spectral bands and spectral indices are obtained after preprocessing of Landsat-8 multispectral remote sensing images. For texture features, the two backscattering coefficients VH and VV of the two polarization modes of SAR images are first obtained, and then a 5×5 filter window is selected, and the gray level co-occurrence matrix method is used to calculate the feature indicators at each angle.

5. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S1, the process of obtaining image blocks is as follows: 9 bands and 8 spectral features are extracted from multispectral images, with a total of 17 variables; 2 backscattering coefficients and 9 texture features are extracted from SAR images, with a total of 11 variables; image objects are obtained by multi-scale segmentation of high-resolution images and resource satellite images; based on each segmented object obtained after segmentation, the mean of each variable extracted from the multispectral image and SAR image in each object is calculated to obtain two column vectors; using the powerful object-oriented image analysis function of Ecognition software, the boundary and morphological features of the objects are extracted through multi-scale segmentation technology to make them more compact and close to the actual lake morphology, and then the minimum outer bounding rectangle of each segmented object is created for the result of multi-scale segmentation of the high-resolution image, thereby generating high-resolution image blocks for the extraction of semantic features.

6. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S2, the MD-DFE neural network model is a hierarchical parallel convolutional neural network, which includes 1D-CNN and HRnet-AM feature extraction modules. These two modules are used to extract deep semantic features from multi-source data sets. The extraction process is as follows: (1) Taking pixels as object units, 1D-CNN is used to process and extract two feature column vectors from Landsat-8 and Sentinel-1 images respectively. The network is designed with 4 layers, including 2 convolutional layers, 1 flattening layer and 1 fully connected layer. All feature maps of each initial input vector will be merged into a one-dimensional column vector through the flattening layer. Finally, the random dropout regularization module is used in the fully connected layer to randomly discard some neurons, and then the features are re-extracted to output the semantic features obtained from the feature column vector. (2) For the processing of high-resolution image blocks, the receptive field needs to be expanded in order to obtain multi-scale information.

7. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1 is characterized in that: In S2, the calculation process of the adaptive weight quantization module is to multiply the weight value and the semantic feature column vector extracted by layered parallel convolution, and use the Concat method to fuse the weighted feature column vectors, thereby increasing the feature dimension describing the image block, but the amount of information under each dimensional feature remains unchanged.

8. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 6 is characterized in that: In S3, the training process is to first add a fully connected layer to the hierarchical parallel convolutional neural network to output classification category information, then use the constructed training set to train the MD-DFE neural network model, use the validation set to verify the model accuracy after each iterative training, set the number of iterations, and perform multiple iterative training. In each iteration, the gradient descent method is used to update the model parameters to reduce the loss value of the model until the number of iterations is reached to obtain a trained MD-DFE neural network model.

9. The lake surface extraction method of multi-source data fusion and deep feature screening as claimed in claim 1, characterized in that: In S4, the process of constructing OGMT is to construct a decision tree structure in the form of a binary tree through a cyclic analysis of the training data set composed of the test variables and the target variables, and use the Gini coefficient as the standard for selecting the most attribute features and the division threshold, that is, the feature with the smallest Gini coefficient is selected as the standard for node splitting. When the samples in each child node belong to the same category or reach the preset threshold, the splitting is stopped, and finally OGMT is obtained.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the lake surface extraction method of multi-source data fusion and deep feature screening described in any one of claims 2 to 9 are implemented.

Images