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

Abstract

The present application relates to the field of lake water surface extraction, and more particularly to a lake water surface extraction method based on multi-source data fusion and deep feature screening. The method comprises: preprocessing multi-source remote sensing data to eliminate differences between different data sources; designing a 1D-CNN and HRnet-AM feature extraction module to extract spectral features, texture features and shape features from images of different data sources; constructing an adaptive weight quantification module A-WCM to fuse and quantify the weights of multiple features; designing a redundancy elimination module REM to eliminate redundant or irrelevant features; using deep semantic fusion technology to layer by layer fuse and recombine the features extracted by each branch; and constructing a decision tree optimal growth model OGMT to effectively reduce the complexity of the model through optimal path selection and feature screening mechanism. The present application improves the recognition efficiency and accuracy of the lake water surface, and can more reliably capture and analyze the changes of the lake water surface, providing strong support for the management and maintenance of the lake ecosystem.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing, lake surface extraction, and deep learning, and particularly to a method for extracting lake surfaces using multi-source data fusion and deep feature screening. Background Technology

[0002] Remote sensing technology has become a crucial tool for extracting lake surface data. However, this process typically relies on a single image source, such as Landsat-8 satellite imagery, ZY satellite imagery, Gaofen satellite imagery, or synthetic aperture radar imagery, distinguishing and extracting lakes based on their spectral characteristics, texture features, and shape outlines. While remote sensing technology has broad application prospects in lake surface extraction, each image source has its inherent advantages and limitations. For example, although Landsat-8 satellite imagery is rich in spectral information, its relatively low spatial resolution limits the accurate identification of ground textures to some extent. ZY and Gaofen satellites have high spatial resolution, capable of clearly capturing surface details, but they have fewer bands, relatively scarce spectral information, and are easily limited by satellite orbits and revisit cycles, potentially affecting data continuity. Furthermore, the lake areas in my country, represented by the eastern plains, receive abundant rainfall, and due to significant cloud cover, optical remote sensing satellites such as Landsat-8, ZY, and Gaofen struggle to meet the requirements for real-time observation and low cloud cover. Synthetic Aperture Radar (SAR) imagery possesses excellent cloud and fog penetration capabilities, providing valuable texture information for ground features. However, due to limitations imposed by orbit and airspace, SAR images suffer from discrepancies and incompleteness in terms of observation angle and resolution. Furthermore, speckle noise also affects the clarity of SAR images. To address the issues of low accuracy in lake identification and boundary mismatches caused by single data sources, experts and scholars both domestically and internationally have proposed multi-source data fusion methods. However, current research still relies on traditional manual methods to extract features from the fused data, failing to fully utilize the advantages of multi-source data fusion, resulting in extraction accuracy not yet meeting expected requirements. Summary of the Invention

[0003] To address the challenge of automating lake surface extraction in existing technologies, this invention provides a lake surface extraction method based on multi-source data fusion and deep feature filtering, primarily comprising:

[0004] S1: Acquire multi-source remote sensing image data, preprocess the multi-source remote sensing image data, and extract the features of the data; cut the preprocessed remote sensing image data to obtain image patches, and construct a training set at the same time;

[0005] S2: Construct the MD-DFE neural network model, which includes differentiated parallel 1D-CNN and HRnet-AM feature extraction modules, an adaptive weight quantization module (A-WCM), and a redundant feature removal module (REM). The 1D-CNN and HRnet-AM feature extraction modules are used to extract semantic features of different image patches; the adaptive weight quantization module (A-WCM) is used to quantize the weights of different semantic features; and the redundant feature removal module (REM) is used to remove redundancy from the extracted multi-dimensional features.

[0006] S3: Employs deep semantic fusion technology to achieve multi-source semantic feature fusion;

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

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

[0009] The beneficial effects of the technical solution provided by this invention are as follows: In the extraction of lake surface data, existing technologies generally face problems such as low recognition accuracy, boundary mismatch, and insufficient robustness due to a single data source, which seriously affects the accuracy and intelligence level of lake surface extraction. To overcome these technical bottlenecks, this invention proposes a solution of multi-source data fusion and deep feature screening. This method utilizes Landsat-8 series imagery, Sentinel-1 SAR imagery, resource satellite imagery, and high-resolution satellite imagery, making full use of the advantages of high-resolution imagery's high spatial resolution, multispectral imagery's spectral information, and SAR imagery's texture feature data. This solves the various problems caused by a single data source, achieving high-precision, intelligent, and automated lake surface extraction, providing scientific data support for understanding and predicting future trends in regional water balance and ecosystems. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0011] Figure 1 This is a schematic diagram of the MD-DFE neural network model in an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram of the 1D-CNN feature extraction network in an embodiment of the present invention;

[0013] Figure 3 This is a schematic diagram of the HRne-AM network structure in an embodiment of the present invention;

[0014] Figure 4 This is a schematic diagram of the attention mechanism module in an embodiment of the present invention;

[0015] Figure 5 This is a schematic diagram of the redundancy removal module in an embodiment of the present invention;

[0016] Figure 6 This is a schematic diagram of the adaptive weight quantization module in an embodiment of the present invention; Detailed Implementation

[0017] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1

[0019] This invention discloses a method for extracting lake surface data through multi-source data fusion and deep feature selection. The method employs deep semantic fusion technology to achieve multi-source semantic feature fusion. Finally, a sample dataset is created to train an MD-DFE (Multi-source data fusion and deep feature extraction) neural network model. The fused multi-source features are then input into the Optimal Growth Decision Tree (OGMT) model to achieve automated extraction of lake surface data. Specifically, this method includes:

[0020] S1: Acquire satellite images from Landsat-8, Ziyuan, Gaofen, and Synthetic Aperture Radar respectively to collect multi-source remote sensing data, preprocess the multi-source remote sensing data to eliminate differences between data sources and ensure spatial alignment between images; and use high spatial resolution satellite data such as Gaofen and Ziyuan to segment and obtain image blocks.

[0021] S2: Construct an MD-DFE (Multi-source data fusion and deep feature extraction) neural network model, design differentiated parallel 1D-CNN (1-Dimensionality-Convolutional Neural Network) and HRnet-AM (High Resolution Net-Attention Mechanism) feature extraction modules to extract deep semantic features from multi-source data, namely, extracting spectral features, texture features, and shape features from images from different data sources to maximize the preservation of effective information from different data sources; construct an Adaptive Weight Calculation Module (A-WCM) to determine the weights of different deep semantic features, improve the model's focus on lake information, and reduce interference from background information; at the same time, use a Redundant Elimination Module (REM) based on the correlation matrix to remove redundant or irrelevant features from the extracted multi-dimensional features; achieve accurate screening and effective fusion of features from multi-source data.

[0022] S3: Employs deep semantic fusion technology to fuse and recombine the features extracted from each branch layer by layer, forming a unified feature representation;

[0023] S4: Based on the fused and recombined features, the constructed Optimal Growth Model for Decision Trees (OGMT) is used for accurate classification of the fused features, thereby achieving automated lake extraction. The model complexity is effectively reduced through optimal path selection and feature filtering mechanisms.

[0024] This embodiment can be implemented in the Datong Lake area of ​​Hunan Province, and the specific details are as follows:

[0025] A. Acquire multi-source remote sensing image data of the mining area, preprocess it, and construct a multi-source dataset:

[0026] A1. Acquisition of Multi-Source Remote Sensing Image Data: Download multi-source data based on information such as the lake's location, phenological characteristics (dry season, wet season), and field survey time. For optical image data, select Landsat-8 multispectral remote sensing images with minimal or no cloud cover, as well as Gaofen and Ziyuan satellite imagery. For optical image data, select the Landsat-8-L1C level multispectral remote sensing image with the lowest cloud cover and Gaofen-1 remote sensing images containing red, green, blue, and near-infrared bands. For SAR image data, select Sentinel-1 interferometric wide-swath (IW) dual-polarized ground distance images from adjacent dates, prioritizing VV and VH dual-polarized data to enhance the accuracy of ground cover classification.

[0027] A2. Preprocessing of Multi-Source Image Data: For optical image data, the "Radiometric Calibration" tool in ENVI software is used, with relevant calibration parameters such as gain, offset, and solar altitude angle input. First, the digital values ​​(DN values) of the image are converted to surface reflectance to eliminate the influence of sensor characteristics and shooting conditions on the image data. Second, the "FLAASH" module is used to eliminate the influence of atmospheric scattering and absorption during image imaging. Third, band resampling is achieved using the "Resample" tool. Users can choose interpolation methods such as "Nearest Neighbor" or "Bilinear Interpolation" to resample the image to a uniform 2-meter spatial resolution. This ensures the consistency of multi-source data in spatial scale, facilitating subsequent fusion and analysis. Finally, geometric correction is achieved using the "GeometricCorrection" module through Ground Control Points (GCPs). Several obvious geographical features in the image (such as road intersections, bridges, etc.) are manually selected and matched with corresponding points in the standard geographic coordinate system to ensure the accuracy of the image's spatial location. For SAR image data, firstly, the Sen2Cor plugin developed by ESA was used to perform radiometric correction on the data, eliminating atmospheric interference and obtaining more accurate surface reflectance features. Secondly, to maintain consistency with the data format and resolution of optical imagery, the "Resampling" tool in SNAP (Sentinel Application Platform) software was used to resample the SAR data to a 2-meter resolution consistent with the optical imagery, ensuring spatial resolution uniformity between the SAR and optical images. Thirdly, the latest orbit file was downloaded and applied using the "Orbit FileCorrection" tool to automatically correct the orbit of the SAR imagery, ensuring spatial alignment with other remote sensing data. Finally, thermal infrared noise removal and speckle noise filtering were performed on the SAR impact data to further improve image quality and obtain clearer texture features.

[0028] A3. Feature Extraction and Calculation of Multi-Source Remote Sensing Image Data: This mainly includes the extraction and calculation of spectral and texture features. For spectral features, nine spectral bands and eight spectral indices were obtained from the preprocessed Landsat-8 multispectral remote sensing images. These are: Coastal Band (B1), Blue Band (B2), Green Band (B3), Red Band (B4), Near Infrared Band (B5), Shortwave Infrared Band 1 (B6), Shortwave Infrared Band 2 (B7), Thermal Infrared Band (B8), Panchromatic Band (B8), Cirrus Band (B9), Normalized Differential Water Index (NDWI), Modified Normalized Differential Water Index (NDWI), Automatic Water Extraction Index (AWEI), Modified Automatic Water Extraction Index (MAWEI), Normalized Difference Vegetation Index (NDVI), Normalized Differential Humidity Index (NDMI), Normalized Differential Soil Index (NSBI), and Enhanced Water Index (EWI). For texture features, firstly, two backscattering coefficients, VH and VV, for the two polarization modes of the SAR image are obtained, and the Sentinel-1 Dual-Polarozed Water Index (SDWI) is used. Then, a 5×5 filter window is selected, and the GreyLevel Co-occurrence Matrix (GLCM) method is used to calculate the mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation of features at each angle.

[0029] Table 1: Image Features

[0030]

[0031] A4. Multi-scale segmentation of high-resolution imagery: Utilizing the powerful Object-Based Image Analysis (OBIA) function of Ecognition software, multi-scale segmentation technology is employed to extract the boundary and morphological features of ground objects, making them more compact and closer to the actual lake morphology. First, in the "MultiresolutionSegmentation" module, the "Scale Parameter" is adjusted to set the segmentation scale. Then, the compactness and smoothness parameters are adjusted to optimize the segmentation results, making the object shapes more compact and closer to the actual lake morphology. The algorithm parameters are set as follows: scale parameter 45, shape factor 0.35, color factor 0.6, compactness factor 0.5, and smoothness factor 0.5.

[0032] A5. Dataset Construction: Nine bands and eight spectral features were extracted from multispectral imagery, resulting in 17 variables; two backscattering coefficients and nine texture features were extracted from SAR imagery, resulting in 11 variables; high-resolution imagery and resource satellite imagery were segmented at multiple scales to obtain image objects. For each segmented object, the mean of each variable extracted from the multispectral and SAR images was calculated, resulting in two column vectors. Then, the minimum bounding rectangle of each segmented object was created from the high-resolution imagery's multi-scale segmentation results, generating high-resolution image patches for deep feature extraction.

[0033] B. Construction of the MD-DFE neural network model: (e.g.) Figure 1 As shown, the specific steps are as follows.

[0034] B1. Construct an MD-DFE neural network model, that is, construct a hierarchical parallel convolutional neural network for deep semantic feature extraction from multi-source datasets. The extraction process is as follows: (1) Using pixels as the object unit, 1D-CNN is used to process the two feature column vectors extracted 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 (e.g. Figure 2 ). Among them, convolutional layers 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 generates a number of feature maps equal to the number of convolutional kernels. All feature maps of each initial input vector are merged into a one-dimensional column vector through a flattening layer. Finally, after randomly dropping some neurons in the fully connected layer through the random deactivation regularization module (Dropout), the features are re-extracted, and the deep semantic features obtained from the feature column vector are output. (2) For the processing of high-resolution image blocks, it is necessary to expand the receptive field in order to obtain multi-scale information. Therefore, this invention designs the following Figure 3 The HRnet-AM feature extraction module is shown. The main body of HRnet-AM consists of three basic blocks: the Bottleneck module, which enhances feature representation through convolutional operations and batch normalization; the BasicBlock, a simpler network structure typically used to build deeper neural networks; and the HighResolutionModule, which is dedicated to maintaining and fusing feature maps of different resolutions. In addition to these three basic blocks, this invention introduces an attention mechanism to improve the overall focus of the model and enhance contextual understanding (e.g., ...). Figure 4 Specifically, the structure that ensures the preservation and utilization of high-resolution information includes four stages and four parallel convolutional stream branches. The high-resolution remote sensing image convolutional stream is the initial stage. High-resolution convolutional streams are gradually added to low-resolution convolutional streams, and the multi-resolution convolutional streams are connected to form new stages. High-resolution and low-resolution features are always preserved in each stage, and cross-resolution information is repeatedly exchanged through a multi-scale fusion module to avoid information loss caused by sampling.

[0035] B2. Design as follows Figure 5 The feature redundancy removal module is shown. Considering the potential for repetition in spectral and texture features extracted from multi-source image data, this invention designs a feature redundancy removal module to improve the model's computational speed. For the extracted multi-source features, irrelevant or redundant features are first removed by compression, and then expanded to obtain the fused multi-source data features required for classification.

[0036] B3. Design as follows Figure 6 The adaptive weight quantization module shown above multiplies the weight values ​​with the deep semantic feature column vectors extracted by hierarchical parallel convolution, and then uses the Concat method to fuse the weighted feature column vectors, thereby increasing the feature dimensions describing the image, while keeping the amount of information under each feature dimension unchanged.

[0037] B4. Design a deep feature fusion module for multi-source features. The multi-source deep semantic features of the object extracted according to the above steps are input into the deep feature fusion module for multi-source feature fusion. First, a novel 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 using the three feature column vectors as units. Finally, the weight values ​​of the importance of the multi-source data features are calculated through two fully connected layers.

[0038] C. Train the MD-DFE neural network model using the training and validation sets, extract and fuse deep features, and use the OGMT classifier for lake extraction.

[0039] C1. Creating a sample dataset for lake classification. First, remote sensing image data of the lake area is labeled, with labels indicating water bodies and non-water bodies, resulting in binary classification images for both. To prevent memory overflow, the images and their corresponding labels are cropped using a sliding method, with the cropped size being 256×256 pixels. Data augmentation is then performed on the cropped original and labeled images, including 180° rotation, horizontal flip, and vertical flip to obtain sufficient samples and avoid overfitting during training. Then, the sample data is randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0040] C2. Train the constructed hierarchical parallel neural network model. First, add fully connected layers to the hierarchical parallel convolutional neural network to output classification information. Then, train the network using the training set constructed in step C1, and use the validation set to verify the model accuracy after each iteration. This invention uses the cross-entropy loss function during model training, and the optimizer is Adam. The calculation method of the cross-entropy loss function is as follows:

[0041]

[0042] in Indicates the true category, Indicates the model's predicted category. The number of image objects in the validation set. This is the model loss value, which represents the magnitude of the model's error.

[0043] C3. Set the number of iterations (epoch). In this embodiment, epoch = 200, and perform multiple iterations of training. Each iteration uses gradient descent to update the model parameters and reduce the model's loss value. Simultaneously, the model parameters (the model parameters refer to the weights of connections between layers in the neural network) are optimized and updated. The network parameter settings are shown in Table 2, and the server configuration is shown in Table 3.

[0044] C4. Constructing the Optimal Growth Model (OGMT) of the Decision Tree. Through iterative analysis of the training dataset consisting of test and target variables, a binary tree-like decision tree structure is constructed. The Gini coefficient (Gini Index) is used as the criterion for selecting the most relevant feature and the splitting threshold. Specifically, first, the root node is initialized, treating all samples as a single node. The Gini coefficient of each feature with respect to the current dataset is calculated, and the feature with the smallest Gini coefficient is selected as the node splitting criterion. Second, based on the selected optimal splitting point, the current feature is split into two child nodes, with each child node representing one value of that feature. After each split, a new root node and child nodes are generated and added to the decision tree. Third, the above steps are repeated for the new nodes until a stopping condition is met (stopping condition: splitting stops when samples in each child node belong to the same class or a preset threshold is reached). Finally, the OGMT classifier is obtained.

[0045] C5. Input the fused deep feature sample data of the image objects into the OGMT classifier, train the OGMT classifier with the training set, obtain the nodes and thresholds, construct the decision tree classification model, and extract the lake surface through the trained OGMT model.

[0046] Table 2 Network Parameter Settings

[0047]

[0048] Table 3 Server Configuration

[0049]

[0050] The experiment used four classification evaluation indicators—cartographic accuracy, user accuracy, overall accuracy, and Kappa coefficient—to test the model accuracy. After multiple iterations of training, the model with the highest accuracy was selected for fine-grained classification of mining area land use.

[0051] D. Apply the trained model to automatically extract lakes: Specific steps include:

[0052] D1. For the acquired remote sensing image objects, calculate their spectral and texture features and extract them as column vectors. Calculate the spectral and texture feature vectors for all unclassified image objects, construct a multi-source dataset, and then input it into a trained hierarchical parallel convolutional neural network model to obtain the fused deep features of each image object. Finally, input it into the OGMT classifier for classification to achieve automated extraction of lake water surfaces.

[0053] D2. Evaluate the accuracy of lake extraction results using test and validation sets. Construct a confusion matrix and evaluate accuracy using metrics such as precision (P), recall (R), false positive rate (MRate), and mean intersection-union ratio (MIoU). Apply the trained model to extract lake surface data.

[0054] The meanings of each indicator are as follows:

[0055] Precision (P): Also known as accuracy, it represents the proportion of correctly classified water samples out of the total number of samples predicted to be water. A higher P value indicates higher extraction precision.

[0056] Recall (R): also known as the completeness of a search, it represents the proportion of correctly classified lake samples out of all real lake samples. The higher the value, the fewer the lakes are missed.

[0057] False positive rate (MRate): The ratio of all misclassified samples to the total number of samples.

[0058] Mean Intersection over Union (MIoU): The average of the intersection and union ratios of the true values ​​and the predicted values ​​obtained by the algorithm for all categories.

[0059] The key points of this invention are:

[0060] 1) This invention comprehensively utilizes multispectral imagery, SAR imagery, and high-resolution and resource imagery to construct a multi-source dataset containing spectral and texture feature column vectors and high-resolution remote sensing image patches, reducing the limitations of image data acquired by different sensors in terms of geometry, spectral, and spatial resolution. Addressing the challenge of accurately extracting lake area and boundaries from single images, multi-source data can provide rich spectral, texture features, and detailed information, improving extraction accuracy.

[0061] 2) This invention designs 1D-CNN and HRnet-AM feature extraction modules, capable of extracting spectral, texture, and shape features from different data sources respectively. Features from each data source are processed through an independent module, thereby maximizing the preservation of effective information. This modular design not only enhances the recognition accuracy of lake boundaries and morphological features but also improves the ability to extract features from complex water bodies, laying a high-quality data foundation for feature fusion and classification.

[0062] 3) This invention constructs an Adaptive Weighted Quantization (A-WCM) module, which effectively highlights lake-related features and reduces interference from background information by fusing and weighting multi-source features. Furthermore, to optimize feature dimensions, a Redundancy Removal (REM) module is specifically designed to filter and remove redundant or irrelevant features. This process not only reduces model complexity but also significantly improves operational efficiency, enabling the model to achieve feature fusion and selection more quickly when processing large-scale remote sensing data.

[0063] 4) This invention employs deep semantic fusion technology, which integrates and recombines features from various data sources layer by layer, fully integrating them in the semantic space to form a unified feature representation. Simultaneously, an optimal growth decision tree (OGMT) model is designed, adaptively selecting the best classification path based on feature differences and weights, effectively reducing model complexity while improving classification efficiency, achieving rapid, accurate, and automated extraction of lake information. This model not only boasts high accuracy in lake boundary identification but also significantly shortens processing time, making it suitable for large-scale lake extraction and monitoring tasks.

[0064] This invention can replace traditional methods such as visual interpretation and machine learning algorithms to achieve automated lake extraction, taking into account both efficiency and accuracy. It can be widely applied to lake extraction work and is of great significance for regional water balance projection and the construction of ecological environment monitoring systems.

[0065] Example 2

[0066] A storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the lake surface extraction method of multi-source data fusion and deep feature screening.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting lake surface data through multi-source data fusion and deep feature screening, characterized in that, include: S1. Acquire multi-source remote sensing image data, including Landsat image data, Sentinel image data, Gaofen and Resource Satellite image data, and preprocess the above image data to construct a training set; S2. Construct the MD-DFE neural network model, which includes a differentiated parallel 1D-CNN and HRnet-AM feature extraction module, an adaptive weight quantization module, a redundant feature removal module, and a multi-source feature deep fusion module. The multi-source feature extraction methods include: extracting spectral features from preprocessed Landsat image data using 1D-CNN; extracting texture features from preprocessed Sentinel image data using 1D-CNN; generating image patches from preprocessed high-resolution and resource satellite image data through multi-scale segmentation; and extracting semantic features from the image patches using HRnet-AM. An adaptive weight quantization module is used to quantize the weights of the multi-source features; a redundant feature removal module is used to remove redundancy from the extracted multi-source features; and a multi-source feature fusion module is used to multiply the multi-source feature weights with the feature column vectors and perform multi-source feature fusion using the Concat algorithm. S3. Train the MD-DFE neural network model using the training set, extract the fused features, input the fused features into the Optimal Growth Model of Decision Tree (OGMT), obtain nodes and thresholds through the optimal path selection and feature filtering mechanism, construct a decision tree classification model, and use the decision tree classification model to realize the automatic extraction of lake surface.

2. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, characterized in that, Based on the lake's location, phenological characteristics, and field survey time, multi-source remote sensing image data was downloaded. Sentinel image data is VV+VH dual-polarization data to enhance the accuracy of land cover classification.

3. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, characterized in that, In S1, the preprocessing is as follows: For optical image data, the DN values ​​of the images are first converted into surface reflectance to eliminate the influence of atmospheric scattering and absorption on image imaging during the imaging process. Then, the images are resampled to a uniform spatial resolution to ensure the consistency of multi-source data at spatial scale, which facilitates subsequent fusion and analysis. Finally, geometric correction is achieved through GCPs. The optical image data consists of Landsat-8 multispectral remote sensing images and Gaofen and Resource Satellite image data. For Sentinel image data, radiometric correction is first performed, then it is resampled to the same resolution as the optical image. The latest orbit file is downloaded and applied, and orbit correction of the Sentinel image is automatically performed. Finally, thermal infrared noise removal and speckle noise filtering are performed on the Sentinel image data to improve image quality and obtain clearer texture features.

4. The lake surface extraction method based on multi-source data fusion and deep feature screening as described in claim 1, characterized in that, In S2, for spectral features, multiple spectral bands and spectral indices of the Landsat-8 multispectral remote sensing image after preprocessing are obtained respectively; for texture features, firstly, two backscattering coefficients VH and VV of the Sentinel image with two polarization modes are obtained, and then a 5×5 filter window is selected, and the feature indexes at each angle are calculated using the gray-level co-occurrence matrix method.

5. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, characterized in that, Landsat-8 multispectral imagery yielded 9 bands and 8 spectral features, totaling 17 variables; Sentinel imagery yielded 2 backscattering coefficients and 9 texture features, totaling 11 variables; high-resolution imagery and resource satellite imagery were segmented at multiple scales to obtain image objects; based on each segmented object, the mean of each variable extracted from the multispectral imagery and Sentinel imagery within each object was calculated, resulting in two column vectors; using the object-oriented image analysis function of Ecognition software, the boundary and morphological features of ground features were extracted through multi-scale segmentation technology, and then the minimum bounding rectangle of each segmented object was created from the results of multi-scale segmentation of high-resolution imagery and resource satellite imagery, thereby generating high-resolution image blocks for semantic feature extraction.

6. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, 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 datasets. The extraction process is as follows: (1) Using pixels as the object unit, 1D-CNN is used to process the two feature column vectors extracted 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, in the fully connected layer, after randomly discarding some neurons through the random inactivation regularization module, the features are re-extracted and the semantic features obtained from the feature column vector are output; (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 method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, characterized in that, In S2, the adaptive weight quantization module calculates the product of the weight values ​​and the semantic feature column vectors extracted by the hierarchical parallel convolution, and then uses 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 feature dimension remains unchanged.

8. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 6, characterized in that, In S3, the training process is as follows: First, a fully connected layer is added to the hierarchical parallel convolutional neural network to output classification information. Then, the MD-DFE neural network model is trained using the constructed training set. The model accuracy after each iteration is checked using the validation set. The number of iterations is set, and multiple iterations are performed. In each iteration, the model parameters are updated using the gradient descent method to reduce the model's loss value until the number of iterations is reached, resulting in a trained MD-DFE neural network model.

9. The method for extracting lake surface data through multi-source data fusion and deep feature screening as described in claim 1, characterized in that, In S4, the process of constructing OGMT is as follows: by iteratively analyzing the training dataset consisting of test variables and target variables, a binary tree-like decision tree structure is formed. The Gini coefficient is used as the standard for selecting the most attributed feature and the splitting threshold. That is, the feature with the smallest Gini coefficient is selected as the standard for node splitting. Splitting stops when the samples in each child node belong to the same category or reach the preset threshold, 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 the processor, it implements the steps of the lake surface extraction method of multi-source data fusion and deep feature screening as described in any one of claims 2-9.

Images