An optical remote sensing water body identification method and system based on multi-source radiation disturbance enhancement and content-aware sampling

Abstract

This invention discloses an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling, comprising: acquiring optical remote sensing images and corresponding ground truth masks; extracting training samples using a content-aware intelligent block segmentation strategy to construct a dynamically balanced training dataset; constructing an adversarial radiation perturbation enhancement pipeline including virtual shadow synthesis to augment the training dataset; constructing a deep semantic segmentation model; iteratively training the deep semantic segmentation model using the augmented training dataset combined with a hybrid loss function until convergence is met; using the trained deep semantic segmentation model to infer the optical remote sensing image under test, and reconstructing the water body identification result containing geographic coordinate information using a geospatial weighted stitching algorithm. Based on this, this invention significantly improves the accuracy of water body extraction under complex urban shadows and variable lighting conditions, and directly outputs seamless vectorized results with geographic coordinates.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to an optical remote sensing water body identification method and system based on multi-source radiation perturbation enhancement and content-aware sampling. Background Technology

[0002] With the rapid development of Earth observation technology, optical remote sensing imagery has become the main data source for obtaining information on surface water bodies. Currently, water body extraction methods are mainly divided into two categories: one is the index method based on spectral features, such as the Normalized Difference Water Index (NDWI) and its improved form (MNDWI). This type of method is simple to calculate but difficult to determine the threshold; the other is the deep learning method based on convolutional neural networks (CNN), such as U-Net and DeepLab, which achieve pixel-level segmentation through end-to-end feature learning.

[0003] Although deep learning methods have significantly improved the automation of water body identification, they still face the following serious challenges in practical engineering applications:

[0004] Severe interference from foreign objects with similar spectra: In optical remote sensing images, the shadows cast by tall buildings, cloud shadows, and deep water areas have extremely similar spectral characteristics (all are low reflectivity regions). Existing training methods often lack adversarial learning mechanisms for shadows, causing models to easily misidentify shadows as water bodies in densely populated urban areas or under cloudy weather conditions, resulting in a large number of artifacts.

[0005] Extreme imbalance between positive and negative samples: In large-format remote sensing images, water bodies typically account for only a very small proportion and are sparsely distributed. Traditional random or regular grid cropping strategies result in a large number of pure land (negative samples) or pure water (redundant samples) in the training dataset, lacking "difficult-to-distinguish samples" rich in land-water boundaries, leading to insufficient segmentation accuracy of the model for small rivers and pond edges.

[0006] Large image reconstruction seams and loss of geographic information: Existing deep learning inference processes typically employ simple block prediction and stitching, which easily produces obvious checkerboard effects (stitching seams) at the boundaries of map blocks. Furthermore, most algorithms only focus on pixel coordinates, often losing the original image's projected coordinate system (CRS) and geographic affine transformation parameters (GeoTransform). This prevents the recognition results from being directly spatially overlaid with Geographic Information Systems (GIS) or other vector data for analysis.

[0007] Existing techniques typically attempt to detect and remove shadows during the preprocessing stage, or to suppress shadows using multi-band indices (such as NDWI). However, these methods often fail in complex urban high-rise building areas and are prone to inadvertently affecting real water bodies. Furthermore, conventional data augmentation (such as rotation and cropping) cannot simulate the spectral confusion characteristics of "different objects with the same spectrum," resulting in insufficient model discrimination ability for shadowed areas. Summary of the Invention

[0008] To overcome the problems of severe shadow confusion, uneven sample distribution, and mismatch between geospatial information and prediction results in existing optical remote sensing water body identification, this invention provides an optical remote sensing water body identification method and system based on multi-source radiation perturbation enhancement and content-aware sampling. By constructing an adversarial radiation perturbation enhancement pipeline based on physical feature decoupling, the method forces a deep semantic segmentation model to decouple spectral brightness features and texture context features, significantly improving the water body extraction accuracy under complex urban shadows and variable lighting conditions, and directly outputting seamless vectorized results with geographic coordinates.

[0009] According to one aspect of the present invention, an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling is provided, comprising: step S1: acquiring optical remote sensing images and corresponding ground truth masks, and extracting training samples using a content-aware intelligent block segmentation strategy to construct a dynamically balanced training dataset; step S2: constructing an adversarial radiation perturbation enhancement pipeline including virtual shadow synthesis to augment the training dataset; step S3: constructing a deep semantic segmentation model, and iteratively training the deep semantic segmentation model using the enhanced training dataset combined with a hybrid loss function until the convergence condition is met; step S4: using the trained deep semantic segmentation model to infer the optical remote sensing image under test, and reconstructing the water body identification result containing geographic coordinate information using a geospatial weighted stitching algorithm.

[0010] Further, in step S1, the extraction of training samples using a content-aware intelligent segmentation strategy to construct a dynamically balanced training dataset includes: synchronously sliding and traversing the optical remote sensing image and its corresponding ground truth mask using a sliding window to extract several image blocks and their corresponding ground truth mask blocks, and calculating the proportion of water pixels in the ground truth mask block corresponding to each image block; marking image blocks with a water pixel proportion within the effective sample threshold range as positive samples and adding them to the training dataset; marking image blocks with a water pixel proportion equal to zero as candidate negative samples, dynamically calculating the number of negative samples to be retained based on the number of positive samples and a preset negative sample balance coefficient, and randomly selecting candidate negative samples from all candidate negative samples based on the number of negative samples to be retained and adding them to the training dataset.

[0011] Further, the data augmentation includes: randomly triggering virtual shadow synthesis with a first preset probability, wherein the virtual shadow synthesis includes: randomly generating several irregular polygonal masks on each image patch in the training dataset, and performing nonlinear brightness attenuation on the pixels within the masks; randomly triggering geometric transformation with a second preset probability, wherein the geometric transformation includes: horizontally flipping, vertically flipping, or rotating 90 degrees on each image patch in the training dataset; randomly triggering nonlinear illumination transformation with a third preset probability, wherein the nonlinear illumination transformation includes: simulating nonlinear spectral responses under different seasons and solar altitude angles on each image patch in the training dataset using random gamma transformation; randomly triggering atmospheric scattering simulation with a fourth preset probability, wherein the atmospheric scattering simulation includes: simulating imaging disturbances during atmospheric transmission on each image patch in the training dataset using Gaussian noise or local contrast enhancement; and randomly triggering brightness and contrast adjustment with a fifth preset probability, wherein the brightness and contrast adjustment includes: randomly adjusting the brightness and contrast of each image patch in the training dataset.

[0012] Further, in step S3, the deep semantic segmentation model is iteratively trained using the enhanced training dataset and a hybrid loss function until the convergence condition is met. This includes: inputting the enhanced training samples into the deep semantic segmentation model for forward propagation to obtain the corresponding prediction results; calculating the error between the prediction results and the ground truth mask using the hybrid loss function, and updating the model parameters of the deep semantic segmentation model through a backpropagation algorithm; repeating the above iterative process until the convergence condition is met. The convergence condition is that the error between the prediction results and the ground truth mask no longer decreases within a preset number of consecutive rounds, or reaches a preset maximum number of training rounds.

[0013] Furthermore, the hybrid loss function described in step S3 is composed of a weighted average of Focal Loss and Dice Loss, and the calculation formula is as follows:

[0014] L_total=λ×L_Focal+(1-λ)×L_Dice,

[0015] Where L_total represents the mixed loss function, L_Focal is the Focal Loss value, L_Dice is the Dice Loss value, and λ is the weight of L_Focal.

[0016] Furthermore, the deep semantic segmentation model described in step S3 uses a U-Net structure as its backbone network. The input layer of the deep semantic segmentation model is expanded to have four channels: red, green, blue, and near-infrared, to utilize the strong absorption characteristics of water in the near-infrared band, enabling the deep semantic segmentation model to extract strong contrast physical features at the water-land interface. The deep semantic segmentation model also embeds a feature decoupling dual attention module and a multi-scale context-aware bottleneck layer. The feature decoupling dual attention module includes a channel attention branch and a spatial attention branch. The channel attention branch adaptively learns the feature weights of each band through global average pooling and a fully connected network to enhance the feature representation of the near-infrared channel. The spatial attention branch is used to guide the network to force attention to regions with rich structural information to generate spatial highlight masks. The multi-scale context-aware bottleneck layer is used to simultaneously capture the global context information of large lakes and the local morphological features of small tributaries.

[0017] Further, step S4 includes: setting an overlapping region during sliding window prediction to extract several image blocks from the optical remote sensing image to be tested that have overlapping regions with adjacent image blocks; inputting several image blocks into a trained deep semantic segmentation model and outputting water body prediction results corresponding to each image block; constructing a weight template matrix with the same size as the image block and high center weight and low edge weight; combining the weight template matrix to perform weighted average fusion of the water body prediction results corresponding to each image block to obtain a large-format water body prediction result; reading the geospatial metadata of the optical remote sensing image to be tested and mapping the geospatial metadata to the large-format water body prediction result to obtain a water body identification result containing geographic coordinate information.

[0018] According to one aspect of the present invention, an optical remote sensing water body identification system based on multi-source radiation perturbation enhancement and content-aware sampling is provided, comprising: a data preprocessing module for acquiring optical remote sensing images and corresponding ground truth masks, and extracting training samples using a content-aware intelligent block segmentation strategy to construct a dynamically balanced training dataset; an enhancement training module for constructing an adversarial radiation perturbation enhancement pipeline including virtual shadow synthesis to enhance the training dataset; a deep semantic segmentation model for iteratively training the deep semantic segmentation model using the enhanced training dataset combined with a hybrid loss function until a convergence condition is met; and an inference and reconstruction module for inferring from the trained deep semantic segmentation model to the optical remote sensing image under test, and reconstructing the water body identification result containing geographic coordinate information using a geospatial weighted stitching algorithm.

[0019] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions executed by the processor, the processor invoking the program instructions to perform an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling.

[0020] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling.

[0021] The above technical solution addresses the problems of uneven distribution of water samples, severe interference from heterogeneous objects (such as shadows) in the same spectrum, and loss of geographic information in the prediction results in existing technologies. First, this invention utilizes a content-aware intelligent segmentation strategy to construct a dynamically balanced training dataset. Second, during the training phase, an adversarial radiative perturbation enhancement pipeline, incorporating virtual shadow synthesis, is introduced to augment the training dataset. The strong interference mechanism of the adversarial radiative perturbation enhancement pipeline disrupts the simple mapping relationship between brightness and water body labels, forcing the deep semantic segmentation model to ignore shallow spectral brightness features and instead focus on and learn the more fundamental texture and contextual spatial features of the water body, significantly improving the accuracy of water body extraction in complex urban shadows and under varying lighting conditions. Finally, the enhanced training dataset is used in conjunction with a hybrid loss function to train the deep semantic segmentation model, and a geospatial weighted stitching algorithm is used to reconstruct the output to obtain a seamless vectorized result with geographic coordinates.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] (1) Accurate identification of small water bodies: The content-aware sampling strategy ensures that the training dataset contains rich samples of the water-land boundary. Combined with Focal Loss optimization, it effectively improves the ability to extract small rivers and ponds.

[0024] (2) Possesses strong anti-shadow interference and generalization capabilities: Traditional remote sensing water extraction models often rely excessively on the brightness feature of "low reflectivity (black)," leading to misclassification of building shadows as water bodies. This invention artificially creates a large number of difficult-to-distinguish samples of "spectrally similar but semantically different objects" by randomly superimposing high-intensity brightness attenuation masks in non-water or water areas. This enables the model to accurately distinguish water bodies even in the shadow areas of unseen urban high-rise buildings, significantly reducing the false alarm rate. Through this high-intensity virtual feature space perturbation, the model demonstrates significantly better robustness and anti-interference capabilities than conventionally trained models when facing complex urban shadows or variable cloud shadows unseen in the real world.

[0025] (3) Strong engineering implementation capability: The output results fully retain geospatial reference information and are seamlessly spliced. They can be directly used for vector production in industries such as surveying and mapping and water conservancy without post-processing. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 The flowchart illustrates an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling, as provided in this embodiment of the invention.

[0028] Figure 2 A flowchart illustrating the model training process provided in this embodiment of the invention.

[0029] Figure 3 This is a schematic diagram of content-aware intelligent block sampling and dynamic balance of positive and negative samples provided in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of virtual shadow synthesis provided in an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram of geospatial weighted seamless reconstruction provided in an embodiment of the present invention.

[0032] Figure 6 A comparison diagram showing the recognition performance of the method provided in this embodiment of the invention and the traditional U-Net method in complex shadow regions.

[0033] Figure 7 This is a functional module structure diagram of an optical remote sensing water body identification system based on multi-source radiation perturbation enhancement and content-aware sampling, provided in an embodiment of the present invention. Detailed Implementation

[0034] It should be noted that:

[0035] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0038] Please refer to the appendix. Figure 1 and 2 This invention provides an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling, comprising:

[0039] Step S1: Obtain optical remote sensing images and corresponding ground truth masks, extract training samples using a content-aware intelligent segmentation strategy, and construct a dynamically balanced training dataset.

[0040] It should be noted that step S1 aims to extract high-quality training samples from large-format optical remote sensing images to address the imbalance between positive and negative samples caused by the sparseness of water targets. Step S1 specifically includes:

[0041] Step S11, Data preparation: Obtain optical remote sensing images (such as Gaofen-2, Sentinel-2, etc.) containing red, green, blue and near-infrared bands and their corresponding pixel-level water body ground truth masks.

[0042] Step S12: Extract training samples using a content-aware intelligent chunking strategy to construct a dynamically balanced training dataset, specifically including (see Appendix). Figure 3 ):

[0043] 1. Sliding window traversal: Set the size of the sliding window to H×W (128×128 pixels in this embodiment), and the step size to S (128 pixels in this embodiment). Use the sliding window to synchronously slide and traverse the optical remote sensing image and the corresponding ground truth mask to extract a number of image patches and corresponding mask patches.

[0044] 2. Content-aware screening: Calculate the proportion of water pixels R_water in the mask patch corresponding to each image patch, R_water = (total number of water pixels in the window) / (H×W), where H is the height of the sliding window and W is the width of the sliding window.

[0045] 3. Positive and negative sample classification and dynamic balance:

[0046] ① Positive sample (Positive Patch) determination: Set the effective sample threshold interval [T_min, T_max]. Mark the image patches with R_water ∈ [T_min, T_max] as positive samples and add them to the training dataset, filtering out noise patches with extremely low water content (R_water < T_min) and redundant patches of pure water (R_water > T_max). In this embodiment, set T_min = 0.01 and T_max = 0.90. The setting of this effective sample threshold interval can automatically eliminate noise samples with extremely low water content (less than 1%) and redundant samples of pure water (more than 90%), forcing the deep semantic segmentation model to focus on learning the edge features at the water-land boundary. It should be noted that the lower limit T_min and the upper limit T_max of the threshold interval can be set according to actual needs and are not limited here.

[0047] ② Negative Patch Control: Image patches with R_water = 0 are marked as candidate negative samples. To prevent the model from making biased predictions due to excessive background, a negative sample balancing coefficient α is introduced (0.45 in this embodiment). The current number of positive samples N_pos is counted. Based on the number of positive samples N_pos and the preset negative sample balancing coefficient α, the number of negative samples N_neg to be retained is dynamically calculated, i.e., N_neg ≈ N_pos × α. Then, N_neg candidate negative samples are randomly selected from all candidate negative samples and added to the training dataset. It should be noted that positive samples refer to image patches containing water bodies or water-land boundaries, used to train the deep semantic segmentation model to learn the features and edges of water bodies; negative samples refer to pure background image patches that do not contain water bodies (such as pure urban buildings or pure forests), used to act as difficult-to-distinguish backgrounds and suppress false positives generated by the deep semantic segmentation model. After the positive and negative samples are extracted in step S1, they together form a dynamically balanced training dataset. In the subsequent steps S2 (data augmentation) and S3 (model training), the training dataset input to the deep semantic segmentation model in each iteration is actually a set of positive and negative samples extracted in a fixed ratio, and their role runs through the entire training process.

[0048] Step S2: Construct an adversarial radiation perturbation enhancement pipeline that includes virtual shadow synthesis to augment the training dataset.

[0049] It should be noted that step S2 is performed online during the data loading phase of model training. Its aim is to improve the generalization ability of the deep semantic segmentation model, particularly its resistance to shadow interference, by constructing an adversarial radiation perturbation enhancement pipeline based on physical feature decoupling. Specifically, physical feature decoupling addresses the issue that traditional remote sensing water extraction models often over-rely on the brightness feature of "low reflectivity (black)," leading to misclassification of building shadows as water bodies. This invention artificially creates a large number of difficult-to-distinguish samples of "spectrally similar but semantically different objects" by randomly superimposing high-intensity brightness attenuation masks in non-water or water body areas. Forced semantic learning disrupts the simple mapping relationship between brightness and water body labels, forcing the deep semantic segmentation model to ignore shallow spectral brightness features and instead focus on and learn the more fundamental texture features (such as ripples) and contextual spatial features (such as riverbanks and water-land topology). Enhanced generalization ability: Through this high-intensity virtual feature space perturbation, the deep semantic segmentation model demonstrates significantly better robustness and anti-interference ability than conventionally trained models when facing complex urban shadows or variable cloud shadows that have never been seen in the real world.

[0050] Furthermore, the enhancement pipeline is essentially an online data augmentation process for the training dataset constructed in step S1. It runs in real-time in memory and does not require the generation of physical files on disk. The enhancement pipeline consists of two parts: geometric transformation and radiometric transformation. Radiometric transformation includes virtual shadow synthesis, nonlinear illumination transformation, atmospheric scattering simulation, and brightness and contrast adjustment. During model training, when data is read, for each image patch in the training dataset, the following enhancement steps are triggered randomly and executed sequentially with independent preset probabilities:

[0051] Virtual shadow synthesis is randomly triggered with a first preset probability (please refer to the appendix for details). Figure 4 The method involves randomly generating several irregular polygonal masks on various image patches in the training dataset, and applying nonlinear brightness attenuation to the pixels within the masks. This preserves the texture features of the ground features while altering their spectral features, thereby constructing adversarial examples that visually resemble water bodies but are semantically labeled as non-water bodies. This forces the model to decouple spectral brightness features from texture context features. In this embodiment, 1-2 irregular polygonal masks are randomly generated on the image, with the mask's edge blurriness (shadow_dimension) set to 5 and the application probability to 0.3. This operation simulates the shadows cast by clouds or urban high-rise buildings on the surface of water, forcing the network to learn to distinguish between "low-reflectivity water bodies" and "low-reflectivity artificial shadows" during training.

[0052] Geometric transformations are randomly triggered with a second preset probability: each image patch in the training dataset is horizontally flipped, vertically flipped, or rotated 90 degrees with a preset probability (e.g., 0.5) to increase spatial diversity.

[0053] Nonlinear illumination transformation is randomly triggered with a third preset probability: the nonlinear spectral response under different seasons and solar altitude angles is simulated using random gamma transformation and added to each image patch in the training dataset; the gamma value is randomly selected in the range of 80 to 120 (i.e. 0.8-1.2).

[0054] Atmospheric scattering simulation is randomly triggered with a fourth preset probability. The atmospheric scattering simulation uses Gaussian noise or local contrast enhancement to simulate the imaging disturbance during atmospheric transmission and adds it to each image patch in the training dataset; the variance range is set to 5.0 to 20.0.

[0055] And / or, randomly trigger brightness and contrast adjustment with a fifth preset probability. Brightness and contrast adjustment: randomly adjust the brightness and contrast of each image block in the training dataset, with the variation range controlled within ±20% to adapt to different imaging conditions.

[0056] It should be noted that each time an optical remote sensing image enters the enhancement pipeline, it will be randomly combined into different enhancement forms (e.g., adding shadows this time, adding noise next time). Therefore, the enhanced training dataset ultimately seen by the deep semantic segmentation model not only contains the original positive and negative samples extracted by S1, but also a massive number of adversarial samples with different interference features derived from them in the physical feature space, greatly enriching the diversity of the data. It should also be noted that the first, second, third, fourth, and fifth preset probabilities can all be set according to actual needs, but the first preset probability must be greater than zero. That is to say, virtual shadow synthesis must be triggered, which is the key to improving the anti-shadow interference capability of this invention.

[0057] Step S3: Construct a deep semantic segmentation model. Use the enhanced training dataset and a hybrid loss function to iteratively train the deep semantic segmentation model until the convergence condition is met.

[0058] In step S3, a deep semantic segmentation model is constructed. The enhanced training dataset is used in conjunction with a hybrid loss function to iteratively train the deep semantic segmentation model. Specifically, the enhanced training samples are input into the deep semantic segmentation model for forward propagation to obtain the corresponding prediction results. The hybrid loss function is used to calculate the error between the prediction results and the ground truth mask, and the model parameters of the deep semantic segmentation model are updated using a backpropagation algorithm. This iterative process is repeated until the convergence condition is met. The convergence condition is: the hybrid loss function value on the dataset (i.e., the error between the prediction results and the ground truth mask) no longer decreases within a preset number of consecutive rounds, or reaches a preset maximum number of training rounds. Understandably, the preset number of consecutive rounds and the preset maximum number of training rounds can be set according to actual needs and are not limited here.

[0059] 1. Improved Deep Semantic Segmentation Model Structure: The deep semantic segmentation model uses the U-Net structure as its backbone network. To synergistically complement the aforementioned adversarial radiation perturbation enhancement strategy and thoroughly solve the problems of complex shadow interference and cross-scale water body identification, this invention makes the following three substantial customized improvements to the network structure:

[0060] Improvement 1: Multi-source physical feature input-level adaptation. The number of input layer channels in the network is expanded from the traditional 3 channels to 4 channels (corresponding to red, green, blue, and near-infrared bands). This design utilizes the strong absorption physical characteristics of water in the near-infrared band (extremely low reflectivity) to maximize the spectral contrast difference between the land and water boundaries in the very early stages of feature extraction, thereby extracting strong contrast physical features at the land-water interface and significantly improving the segmentation accuracy of water body edges.

[0061] Improvement 2: Embedded Feature Decoupling Dual Attention Module (FD-DAM). An innovative channel-spatial dual attention mechanism is introduced at the skip connection between the encoder and decoder. ① Channel Attention Branch: Adaptively learns feature weights for each band through global average pooling and a fully connected network. When encountering shadow adversarial examples, the network actively suppresses the vulnerable visible light channel weights and enhances the feature representation of the near-infrared channel; ② Spatial Attention Branch: Generates a spatial highlight mask, guiding the network to focus its attention on regions with rich structural information (such as water ripples and riverbank topography). This dual mechanism perfectly responds to and supports the decoupling learning objective of "suppressing brightness features and enhancing texture features" in step S2 at the network structure level.

[0062] Improvement 3: Embedding a Multi-Scale Context-Aware Bottleneck Layer (MS-CAB). In the deepest layer of the network (the bottleneck layer with the number of feature channels reduced and fixed to 512), the traditional single convolution is abandoned, and a parallel dilated convolutional spatial pyramid pooling module with different dilation rates (e.g., 1, 3, 5) is introduced. This improvement significantly expands the receptive field of deep feature maps without increasing the number of network parameters, enabling the network to simultaneously capture global contextual information of large lakes and local morphological features of small tributaries.

[0063] 2. Hybrid Loss Function: The total loss function L_total of the semantic segmentation model is defined as the weighted sum of Focal Loss and Dice Loss, variants of the binary cross-entropy loss: L_total = λ × L_Focal + (1-λ) × L_Dice. Where L_total represents the hybrid loss function, L_Focal is the loss value of Focal Loss, L_Dice is the loss value of Dice Loss, and λ is the weight of L_Focal. In this embodiment, λ = 0.5.

[0064] 3. Focal Loss: Parameters alpha=0.25, gamma=2.0. This option increases the training weights of hard-to-classify samples (such as small tributaries or shadow edges) and decreases the weights of easy-to-classify samples (such as large lakes) by focusing the parameters, thus solving the class imbalance problem.

[0065] 4. Dice Loss: Directly optimizes the geometric overlap between the predicted result and the true value, providing better constraints on the accuracy of the segmentation boundary.

[0066] 5. Optimization Strategy: The Adam optimizer is used with an initial learning rate of 0.00005, and a cosine annealing strategy is employed to dynamically adjust the learning rate. The number of training epochs is set to 100. During training, the model parameters with the highest IoU on the validation set are saved.

[0067] Step S4: Use the trained deep semantic segmentation model to infer the optical remote sensing image under test, and use the geospatial weighted stitching algorithm to reconstruct the water body identification result containing geographic coordinate information.

[0068] It should be noted that step S4 aims to perform inference prediction on the optical remote sensing image under test, ensuring seamless stitching and correct geographic coordinates. Specifically, firstly, the optical remote sensing image under test is segmented into several image blocks, and then these image blocks are input into the deep semantic segmentation model. The output is the water body prediction result for each image block (i.e., isolated 128x128 pixel prediction probability matrices without geographic coordinates). The geospatial weighted stitching algorithm is performed entirely outside the deep semantic segmentation model; it is a post-processing step in the system. The geospatial weighted stitching algorithm is used to smoothly stitch these prediction probability matrices together according to the weight template matrix, effectively eliminating the checkerboard effect caused by traditional hard stitching. The specific steps of step S4 are as follows:

[0069] 1. Overlapping Sliding Window Prediction: During sliding window prediction, an overlapping region is defined to extract several image blocks that overlap with adjacent image blocks. These image blocks are then input into a trained deep semantic segmentation model, which outputs the water body prediction results for each image block. In this embodiment, the image block size during prediction is set to 128x128, and the width of the overlap region between adjacent image blocks is 32 pixels. Understandably, the image block size and the width of the overlap region can be set according to actual needs and are not limited here.

[0070] 2. Spatial Weight Template Construction: Construct a weight template matrix W with the same size as the image patch (the weight in the center region of the matrix is ​​1, and it linearly decays to 0 towards the edges). Combine the weight template matrix W with the water body prediction results corresponding to each image patch and perform a weighted average fusion to obtain the water body prediction results for the entire large-format image (please refer to the appendix). Figure 5 For any pixel (x, y) within the overlapping region, its final predicted value P_final is obtained by a weighted average of multiple water body prediction results:

[0071] P_final = (∑ P_i × W_i) / (∑ W_i) ,

[0072] Where P_final is the final predicted value of any pixel (x,y) in the overlapping region, P_i is the predicted probability value of the i-th image block covering the pixel for the pixel (x,y), and W_i represents the spatial weight value of the pixel (x,y) in the weight template corresponding to the i-th image block covering the pixel.

[0073] 3. Geospatial Metadata Mapping: Since large-scale water body prediction results themselves do not possess geographic coordinate information, geospatial metadata needs to be mapped onto these results to obtain water body identification results containing geographic coordinate information (please refer to the appendix). Figure 6 Left: Original image; Middle: Traditional U-Net recognition result; Right: Overlay comparison of the result obtained by this invention based on Adversarial Data Augmentation and the traditional U-Net recognition result. The blue area represents the water body accurately extracted by this invention; the red patches represent shadow areas that were misidentified as water by the traditional U-Net (which this invention has successfully removed). Comparative experimental results show that in a typical urban high-rise building dense area test scenario, compared to the traditional U-Net model, the improved U-Net model proposed in this invention maintains the integrity of the water body while successfully reducing the number of false positives caused by shadows by more than 90%, significantly reducing the false alarm rate.

[0074] It should be noted that when large-scale water body prediction results are saved as entity files, they have two forms: one is a probability map file containing continuous probability values, and the other is a binary mask file that has undergone threshold binarization (either 0 or 1). These are image files themselves, not file headers. Based on this, this invention utilizes an external program to read the geospatial metadata of the original optical remote sensing image to be measured. The geospatial metadata includes geographic coordinate system information (CRS, such as WGS84 UTM projection) and affine transformation matrix (GeoTransform, including resolution, top-left corner coordinates, and rotation parameters). This geospatial metadata is directly written into the file headers (i.e., metadata tags) of the probability map and binary mask image files generated from the large-scale water body prediction results, and the compression method is set to LZW. This gives ordinary image files geospatial awareness capabilities, ensuring that the output results can be accurately overlaid directly with the original image or other GIS vector data without manual registration.

[0075] Based on the same technical concept as the foregoing embodiments, the present invention also provides an optical remote sensing water body identification system based on multi-source radiation perturbation enhancement and content-aware sampling, such as... Figure 7 As shown, the system includes a data preprocessing module, an augmentation training module, and an inference reconstruction module. The data preprocessing module acquires optical remote sensing images and corresponding ground truth masks, and extracts training samples using a content-aware intelligent segmentation strategy to construct a dynamically balanced training dataset. The augmentation training module constructs an adversarial radiation perturbation augmentation pipeline that includes virtual shadow synthesis to augment the training dataset. A deep semantic segmentation model is constructed, and the augmented training dataset combined with a hybrid loss function is used to iteratively train the deep semantic segmentation model until the convergence condition is met. The inference reconstruction module uses the trained deep semantic segmentation model to infer the optical remote sensing images under test, and uses a geospatial weighted stitching algorithm to reconstruct the water body identification results containing geographic coordinate information.

[0076] Based on the same technical concept as the foregoing embodiments, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling.

[0077] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium that stores computer instructions that cause the computer to execute an optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling.

[0078] In summary, addressing the problems of uneven water sample distribution, severe interference from heterogeneous objects (such as shadows) and loss of geographic information in prediction results in existing technologies, this invention first utilizes a content-aware intelligent segmentation strategy to construct a dynamically balanced training dataset. Secondly, during the training phase, an adversarial radiative perturbation enhancement pipeline, incorporating virtual shadow synthesis, is introduced to augment the training dataset. The strong interference mechanism of this pipeline disrupts the simple mapping relationship between brightness and water body labels, forcing the deep semantic segmentation model to ignore shallow spectral brightness features and instead focus on and learn the more fundamental texture and contextual spatial features of the water body, significantly improving the accuracy of water body extraction in complex urban shadows and under varying lighting conditions. Finally, the enhanced training dataset is used in conjunction with a hybrid loss function to train the deep semantic segmentation model, and a geospatial weighted stitching algorithm is used to reconstruct the output, resulting in a seamless vectorized output with geographic coordinates.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for optical remote sensing water body identification based on multi-source radiation perturbation enhancement and content-aware sampling, characterized in that, include: Step S1: Acquire optical remote sensing images and corresponding ground truth masks, and extract training samples using a content-aware intelligent segmentation strategy to construct a dynamically balanced training dataset. Step S2: Construct an adversarial radiation perturbation enhancement pipeline that includes virtual shadow synthesis to augment the training dataset; Step S3: Construct a deep semantic segmentation model, and use the enhanced training dataset combined with a hybrid loss function to iteratively train the deep semantic segmentation model until the convergence condition is met; Step S4: Use the trained deep semantic segmentation model to infer the optical remote sensing image under test, and use the geospatial weighted stitching algorithm to reconstruct the water body identification result containing geographic coordinate information.

2. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, In step S1, training samples are extracted using a content-aware intelligent chunking strategy to construct a dynamically balanced training dataset, including: The optical remote sensing image and its corresponding ground truth mask are synchronously traversed using a sliding window to extract several image blocks and their corresponding ground truth mask blocks, and the proportion of water pixels in the ground truth mask block corresponding to each image block is calculated. Image blocks whose water pixel ratio falls within the effective sample threshold range are marked as positive samples and added to the training dataset; Image blocks with zero water pixel ratio are marked as candidate negative samples. The number of negative samples to be retained is dynamically calculated based on the number of positive samples and the preset negative sample balance coefficient. Candidate negative samples are randomly selected from all candidate negative samples and added to the training dataset according to the number of negative samples to be retained.

3. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, The data enhancements include: Virtual shadow synthesis is randomly triggered with a first preset probability. The virtual shadow synthesis includes: randomly generating several irregular polygonal masks on each image block in the training dataset, and performing nonlinear brightness attenuation on the pixels within the masks. A geometric transformation is randomly triggered with a second preset probability. The geometric transformation includes: horizontally flipping, vertically flipping, or rotating each image block in the training dataset by 90 degrees. A nonlinear illumination transformation is randomly triggered with a third preset probability. The nonlinear illumination transformation includes: simulating the nonlinear spectral response under different seasons and solar altitude angles on each image block in the training dataset using random gamma transformation. Atmospheric scattering simulation is randomly triggered with a fourth preset probability. The atmospheric scattering simulation includes: simulating imaging disturbances during atmospheric transmission by using Gaussian noise or local contrast enhancement on each image patch in the training dataset. And / or, randomly trigger brightness and contrast adjustment with a fifth preset probability, wherein the brightness and contrast adjustment includes: randomly adjusting the brightness and contrast of each image block in the training dataset.

4. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, In step S3, the deep semantic segmentation model is iteratively trained using the enhanced training dataset and a hybrid loss function until the convergence condition is met, including: The enhanced training samples are input into the deep semantic segmentation model for forward propagation to obtain the corresponding prediction results; The error between the predicted result and the ground truth mask is calculated using a hybrid loss function, and the model parameters of the deep semantic segmentation model are updated using a backpropagation algorithm. Repeat the above iterative process until the convergence condition is met; the convergence condition is: the error between the prediction result and the ground truth mask block no longer decreases within a preset number of consecutive rounds, or reaches a preset maximum number of training rounds.

5. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, The hybrid loss function described in step S3 is composed of a weighted average of Focal Loss and Dice Loss, and the calculation formula is as follows: L_total=λ×L_Focal+(1-λ)×L_Dice, Where L_total represents the mixed loss function, L_Focal is the Focal Loss value, L_Dice is the Dice Loss value, and λ is the weight of L_Focal.

6. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, The deep semantic segmentation model described in step S3 uses the U-Net structure as its backbone network. The input layer of the deep semantic segmentation model is expanded to have four channels: red, green, blue, and near-infrared. This is to take advantage of the strong absorption characteristics of water in the near-infrared band so that the deep semantic segmentation model can extract the strong contrast physical features at the water-land interface. The deep semantic segmentation model also embeds a feature-decoupled dual attention module and a multi-scale context-aware bottleneck layer. The feature-decoupled dual attention module includes a channel attention branch and a spatial attention branch. The channel attention branch adaptively learns the feature weights of each band through global average pooling and a fully connected network to enhance the feature representation of the near-infrared channel. The spatial attention branch is used to guide the network to force attention to regions with rich structural information in order to generate spatial highlight masks. The multi-scale context-aware bottleneck layer is used to simultaneously capture the global context information of large lakes and the local morphological features of small tributaries.

7. The optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in claim 1, characterized in that, Step S4 includes: When predicting with a sliding window, an overlapping area is set to extract several image blocks from the optical remote sensing image to be tested that have overlapping areas with adjacent image blocks. Several image patches are input into a trained deep semantic segmentation model, and the water body prediction results corresponding to each image patch are output. Construct a weight template matrix with the same size as the image patch, but with high weight at the center and low weight at the edges; The water body prediction results corresponding to each image block are weighted and averaged by combining the weight template matrix to obtain the large-format water body prediction results. The geospatial metadata of the optical remote sensing image to be tested is read, and the geospatial metadata is mapped to the large-format water body prediction result to obtain the water body identification result containing geographic coordinate information.

8. An optical remote sensing water body identification system based on multi-source radiation perturbation enhancement and content-aware sampling, characterized in that, include: The data preprocessing module is used to acquire optical remote sensing images and corresponding ground truth masks, and to extract training samples using a content-aware intelligent block segmentation strategy to construct a dynamically balanced training dataset. An enhanced training module is used to construct an adversarial radiation perturbation enhancement pipeline that includes virtual shadow synthesis to augment the training dataset; A deep semantic segmentation model is constructed, and the deep semantic segmentation model is iteratively trained using an enhanced training dataset and a hybrid loss function until the convergence condition is met. The inference and reconstruction module is used to infer the optical remote sensing image under test using a trained deep semantic segmentation model, and to reconstruct the water body identification result containing geographic coordinate information using a geospatial weighted stitching algorithm.

9. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the optical remote sensing water body identification method based on multi-source radiation perturbation enhancement and content-aware sampling as described in any one of claims 1 to 7.

Images