A wetland classification method based on remote sensing images and mamba network

Abstract

The application provides a wetland classification method based on remote sensing images and a Mamba network, and realizes collaborative modeling of wetland time sequence change characteristics and spatial structure details by constructing a Mamba classification network fusing a time attention mechanism and a spatial attention mechanism, and the method comprises the following steps: acquiring multi-spectral remote sensing images of a research area and performing pretreatment; constructing training samples and a pixel-level label data set; inputting the training samples into a classification network with Mamba as the main body for training, wherein the classification network comprises a Mamba main body module, a time attention module and a spatial attention module; using the trained network to infer a to-be-classified image, outputting a wetland classification result and generating a wetland mapping result. The application can effectively improve the discrimination ability of fine-grained patches, complex transition zones and adjacent class confusion areas, improve the stability and boundary integrity of the wetland classification result, and has good engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, specifically relating to a wetland classification method based on remote sensing images and the Mamba network. Background Technology

[0002] Wetlands, as vital natural ecosystems, play an irreplaceable role in water resource regulation, biodiversity conservation, ecological environment maintenance, and support for human production and daily life. With global climate change and increasing human activity, wetland ecosystems face degradation, fragmentation, and functional decline, placing higher demands on the detailed investigation, dynamic monitoring, and scientific management of wetland resources. Obtaining high-precision and up-to-date wetland classification information is fundamental to wetland protection and management.

[0003] Remote sensing technology provides an important data source for large-scale wetland surveys. Among them, the Sentinel-2 satellite, with its multispectral observation capabilities, high spatial resolution, and short revisit period, can provide continuous, multi-temporal surface observation data, and has broad application potential in wetland monitoring and classification. However, wetland types exhibit strong heterogeneity, with significant spectral aliasing often existing between different wetland types. Furthermore, the complex morphology of wetland boundaries and large spatial scale variations pose significant challenges to the refined classification of wetlands.

[0004] Fine-grained wetland classification based on remote sensing imagery is a crucial foundation for ecological assessment and water and land resource management. However, wetlands exhibit significant differences in land cover scale, irregular morphology, and complex boundary transitions. Types such as shallow water-tidal flats and wetland vegetation-cultivated land are prone to spectral aliasing and pixel mixing, leading to misclassification in transitional zones and fragmented areas. Furthermore, under multi-temporal observation conditions, fluctuations in water level, phenological changes, and differences in imaging conditions can cause spectral representations of the same land cover to fluctuate, reducing classification stability and reliability. Figure 1 To the point of being responsive.

[0005] Existing wetland classification methods based on Sentinel-2 imagery mostly employ traditional machine learning models or conventional convolutional neural network structures, relying primarily on local spatial features for discrimination. These methods struggle to effectively characterize the temporal changes in wetlands and adequately represent the long-range spatial dependencies and boundary structure information of wetland objects. In multi-temporal remote sensing imagery applications, traditional methods have limited ability to utilize temporal information and are easily affected by seasonal variations, environmental noise, and other factors, thus impacting the stability and generalization ability of the classification results.

[0006] In recent years, deep learning methods have been widely used in the field of remote sensing image interpretation. However, how to achieve long-range temporal feature modeling of multi-temporal remote sensing images while maintaining computational efficiency, and further enhance the expressive ability of wetland boundary and spatial texture features, remains a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to provide a wetland classification method based on remote sensing imagery and the Mamba network. Addressing the challenges of confusion between wetland types, difficulty in identifying boundary transition areas, and unstable characterization under multi-temporal conditions, this invention provides a wetland classification method based on Sentinel-2 multispectral remote sensing imagery, comprising the following steps: Multispectral remote sensing images of the study area are acquired and preprocessed. The preprocessed multispectral images are then labeled to construct a training dataset for wetland classification. This training dataset is input into a Mamba-based classification network for training, which outputs wetland classification results. The classification network includes a feature embedding layer, a Mamba backbone module, a temporal attention module, a spatial attention module, and a classification output layer. Specifically, the feature embedding layer maps the preprocessed multispectral remote sensing images to feature representations, obtaining features for each temporal phase. The Mamba backbone module encodes these temporal features and establishes contextual relationships, resulting in encoded features for each temporal phase. The temporal attention module performs adaptive weighted fusion of these encoded features to obtain fused features. The spatial attention module recalibrates the fused features in both spatial and channel dimensions. The classification output layer outputs pixel-level wetland classification results.

[0008] Furthermore, the Mamba backbone module is used to encode the features of each time phase and establish contextual associations to obtain the encoded features of each time phase. Specifically, the features of each time phase formed by feature embedding of the preprocessed multispectral remote sensing image are expanded into a sequence according to a predetermined scanning order. By utilizing state-space recursion to characterize long-range dependencies, the global context modeling capability is enhanced while maintaining computational efficiency. The mathematical expression is: , in The hidden state is the state representation of the current position t and its preceding context information during the sequence recursion process. It is the state representation of position t-1 and its preceding context information during the sequence recursion process, where A, B, and C are training parameters. The output is the encoded features corresponding to each sequence position. It is a sequence.

[0009] Furthermore, the temporal attention module is used to adaptively weight and fuse the encoded features of each temporal phase to obtain fused features. Specifically, it performs weighted fusion of T encoded features, with the weights adaptively calculated by the attention mechanism to highlight key temporal phases and suppress noisy temporal phases. The mathematical expression is as follows: , Where q, , Let be the query, key, and value obtained by linear transformation of the encoding features of each time phase, respectively, where d is the feature dimension, and softmax() is the normalization function. The attention weight coefficient corresponding to the i-th temporal phase encoded feature is used to characterize the contribution of that temporal phase to the temporal feature fusion. These are the features after fusion.

[0010] Furthermore, the spatial attention module is used to recalibrate the fused features in both spatial and channel dimensions. The specific algorithm implementation is as follows: , in This indicates that convolutions with different dilation rates are used to expand the receptive field. GAP represents global average pooling, S is the spatial weight map, w is the channel weight vector, sigmoid() is the activation function, and ReLU() is the rectified linear activation function. and These are two sets of linear transformation parameters in the channel attention branch. Features after fusion This is the enhanced feature map after spatial attention and channel attention weighting.

[0011] Furthermore, in the training of the Mamba-based classification network, features from different levels / scales are first aligned to a unified spatial scale, and then weighted convergence is performed using fusion weights. The fusion weights are learned end-to-end by the network or adaptively generated by the attention mechanism, and satisfy the overall normalization constraint.

[0012] Furthermore, when training the Mamba-based classification network, the network is iteratively trained based on training samples, and the network parameters are updated through backpropagation; wherein, the training objective is to minimize the loss function, which includes at least pixel-level classification loss.

[0013] Furthermore, when training the Mamba-based classification network, training samples are selected using land use data and publicly available high-resolution image data, and the wetland classification results are validated using field survey data.

[0014] Furthermore, the multispectral remote sensing imagery includes a multi-temporal image sequence of the same study area. The multi-temporal image sequence is filtered according to a preset time window and registered and aligned according to the same spatial reference for subsequent temporal feature modeling of the temporal attention mechanism.

[0015] Furthermore, the preprocessing includes: masking clouds, shadows, and invalid pixels based on scene classification layers; stitching and cropping multi-temporal images; and resampling and normalizing each band to ensure that each temporal image meets a uniform spatial resolution and band scale.

[0016] Furthermore, the sample annotation includes: dividing the preprocessed multispectral remote sensing image into image blocks according to a preset window, generating a corresponding pixel-level wetland category label for each image block, and constructing a training dataset in the form of image block-label pairs.

[0017] The Mamba classification network constructed in this invention, which integrates temporal and spatial attention, can fully utilize multi-temporal and spatial structural information under multispectral imagery conditions. On the one hand, it enhances the modeling ability of temporal variation features through the temporal attention mechanism, thereby improving the stability of classification results under different observation conditions. On the other hand, it strengthens the expression of boundary and texture details through the spatial attention mechanism (preferably spatial-channel attention), thereby improving the recognition ability of transition zones and complex regions. At the same time, by combining the Mamba backbone with multi-scale feature fusion, it improves the accuracy, boundary quality, and stability of wetland classification results while ensuring computational efficiency. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 4 This is a structural diagram of the classification network.

[0019] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a location map of the wetland classification study area in one embodiment of the present invention; Figure 3 This is a model classification result diagram in one embodiment of the present invention; Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] The following combination Figure 1The specific implementation of this invention will be further described below. This embodiment uses Sentinel-2 multispectral remote sensing imagery as input to construct a Mamba wetland classification network that integrates temporal attention and spatial-channel attention mechanisms. Combined with a multi-scale feature fusion strategy, it achieves fine wetland classification and enhanced boundary details.

[0022] like Figure 1 As shown, the method in this embodiment generally includes image acquisition and preprocessing, sample construction, network training and inference output, and accuracy evaluation and generalization verification. The specific process is as follows: (1) Sentinel-2 image acquisition and temporal phase selection: Sentinel-2 multispectral images within the coverage area of ​​the study area are acquired, and temporal phase selection is performed based on the imaging date range, cloud cover threshold, and quality indicators to form an image set or multi-temporal sequence for training and inference. Figure 2 This is a location map of the wetland classification study area used in this embodiment; (2) Preprocessing: Cloud / shadow masking, image stitching and cropping, geometric registration, resampling and band unification and normalization are performed on the selected images; when unifying the bands, the bands of different resolutions are resampled at a unified scale and the bands are aligned so that the multi-band and multi-temporal data meet the consistent spatial reference and numerical scale requirements.

[0023] (3) Training Sample and Label Construction: Image-label training sample pairs are created based on preprocessed images. Image samples can be divided into image patches by a fixed window, and the sliding window step size and overlap ratio are set; labels can be labeled at the pixel level to support fine classification. The training samples are further divided into training set, validation set and test set.

[0024] (4) Training configuration and iterative optimization: Input the training samples into the classification network for iterative training, minimize the loss function using supervised learning, and update the network parameters through backpropagation; the learning rate can be adjusted, early stopping and model selection can be performed based on validation set metrics. Figure 3 The image shows the classification results of the model.

[0025] (5) Reasoning and stitching output: Sliding window reasoning is used on the image to be classified to obtain the pixel-level category probability map of each window; average fusion or confidence weighted fusion can be used for overlapping areas, and finally the whole classification result is stitched together.

[0026] (6) Evaluation and verification: Calculate the overall accuracy (OA), Kappa coefficient, F1 and other indicators based on the confusion matrix; optional generalization test under different scenario conditions to verify the stability of the model under different observation conditions.

[0027] The classification network comprises a feature embedding layer, a Mamba backbone module, a temporal attention module, a spatial attention module, and a classification output layer. The feature embedding layer maps preprocessed multispectral remote sensing images to feature representations, obtaining features for each temporal phase. The Mamba backbone module encodes these temporal features and establishes contextual relationships, resulting in encoded features for each temporal phase. The temporal attention module adaptively weights and fuses these encoded features to obtain fused features. The spatial attention module recalibrates the fused features in both spatial and channel dimensions, first reweighting spatial locations to highlight boundaries and structures, and then recalibrating channels to emphasize bands and / or feature channels that are more sensitive to classification. The classification output layer outputs pixel-level wetland classification results.

[0028] like Figure 4 As shown, the input image for the Mamba-based classification network is a preprocessed multispectral remote sensing image. Where H and W are spatial dimensions, and C is the number of spectral channels.

[0029] The input image X is mapped through the feature embedding layer to obtain temporal features, which are then input into the Mamba backbone module. The Mamba backbone module expands the preprocessed multispectral remote sensing image, after feature embedding, into a sequence according to a predetermined scanning order. By utilizing state-space recursion to characterize long-range dependencies, the global context modeling capability is enhanced while maintaining computational efficiency. The mathematical expression is: , in The hidden state is the state representation of the current position t and its preceding context information during the sequence recursion process. It is the state representation of position t-1 and its preceding context information during the sequence recursion process, where A, B, and C are training parameters. The output is the encoded features corresponding to each sequence position. As a sequence, this structure can model long-range dependencies while maintaining linear complexity.

[0030] The time attention module is used to adaptively weight and fuse the encoded features of each time phase to obtain the fused features. Specifically, it performs weighted fusion of T encoded features, with the weights adaptively calculated by the attention mechanism to highlight key time phases and suppress noisy time phases. The mathematical expression is as follows: , Where q, , Let be the query, key, and value obtained by linear transformation of the encoding features of each time phase, respectively, where d is the feature dimension, and softmax() is the normalization function. The attention weight coefficient corresponding to the i-th temporal phase encoded feature is used to characterize the contribution of that temporal phase to the temporal feature fusion. These are the features after fusion.

[0031] The spatial attention module is used to recalibrate the fused features in both spatial and channel dimensions. The specific algorithm implementation is as follows: , in This indicates that convolutions with different dilation rates are used to expand the receptive field. GAP represents global average pooling, S is the spatial weight map, w is the channel weight vector, sigmoid() is the activation function, and ReLU() is the rectified linear activation function. and These are two sets of linear transformation parameters in the channel attention branch. Features after fusion This is the enhanced feature map after spatial attention and channel attention weighting.

[0032] In the process of training and inference of classification network, in order to take into account the expression of wetland patches at different spatial scales, features from different levels / scales can be aligned to a unified spatial scale first, and then weighted convergence can be performed using learnable fusion weights. The fusion weights can be learned end-to-end by the network or adaptively generated by the attention mechanism, and satisfy the overall normalization constraint to avoid the dominance of single-scale features.

[0033] Upsampling alignment can be achieved using bilinear interpolation, nearest neighbor, or transposed convolution; fusion weights can be set as learnable parameters or adaptively generated by the attention branch based on the input. This module is used to improve the expressive power of wetland patches and transition regions at different scales and enhance the spatial consistency of classification results.

[0034] Loss function (description): The training objective is to minimize pixel-level classification loss; preferably, cross-entropy loss is used as the basic classification term, and a boundary consistency constraint term and a Dice class loss term to alleviate class imbalance can be further superimposed according to the boundary fineness and class distribution. The weight coefficients of each loss term can be set as hyperparameters and selected on the validation set to balance overall accuracy and boundary integrity.

[0035] The input data consists of preprocessed Sentinel-2 image patches and their corresponding labels. Each image patch contains multiple spectral channels and can be organized into a multi-temporal sequence. The Mamba backbone is used to encode the input features and establish a wide range of contextual relationships, enhancing the ability to express the large-scale background and long-range spatial dependencies of wetlands through sequential modeling. The temporal attention module, oriented towards multi-temporal input, adaptively weights and fuses features from different temporal phases to highlight temporal information relevant to wetland classification and suppress temporal noise significantly affected by water level fluctuations, residual thin clouds, or phenological fluctuations, thereby improving temporal consistency and classification stability. The spatial attention module is used to strengthen the expression of spatial structural information, highlighting transition zone boundaries and texture differences. Preferably, the spatial attention module employs a spatial-channel joint recalibration method, enhancing boundary and fragmented patch responses in the spatial dimension and assigning higher weights to spectral channels more conducive to wetland identification in the channel dimension, thereby improving the fine classification ability of complex boundary regions and regions confused with adjacent categories.

[0036] Figure 3 , Figure 4 To verify the effectiveness of each module setting, an ablation experiment was conducted under the same training and inference settings to compare the classification performance of different module combinations. The results are shown in Table 1.

[0037] As shown in Table 1, based on baseline A0, introducing temporal attention (A1) improves the utilization of multi-temporal information; introducing spatial-channel attention (A2) significantly enhances boundary and texture details; and introducing multi-scale fusion (A3) helps to identify wetland patches at different scales. The complete model A4, which integrates the above modules, achieves the best overall performance (OA=91.4, Kappa=0.88, mF1=0.86), representing improvements of 2.8, 0.04, and 0.05 respectively compared to baseline A0. This indicates that the network structure of this invention can significantly improve the accuracy and stability of wetland classification results.

[0038] The above are merely preferred embodiments of the present invention. For the preprocessing flow, temporal organization method, attention module structure, loss function composition, and post-processing strategy, equivalent substitutions or simple modifications can be made without departing from the spirit of the present invention, and all such modifications should fall within the protection scope of the present invention.

[0039] Table 1 shows the results of the ablation experiments for each module: , As can be seen from the ablation comparison in Table 1, the baseline model with Mamba as its backbone already has strong global modeling capabilities. With the addition of temporal attention, the model can make fuller use of complementary information from multi-temporal observations and reduce interference from anomalous temporal phases. With the addition of spatial-channel attention, the classification of transition zones and boundary regions is more stable, and the detail continuity of the classification map is better. Further introduction of multi-scale fusion can enhance the ability to express small patches, slender strip wetlands, and areas with significant scale changes, thereby improving both overall accuracy and fine-grained category recognition.

[0040] In practice, the preprocessing workflow (cloud / shadow masking, registration and alignment, band unification), sample slicing strategy (window size, step size, and overlap ratio), and loss term weights can be adjusted according to the land cover composition and sample distribution of the study area. Simultaneously, overlapping fusion using sliding window inference and simple post-processing (such as connected component filtering, hole filling, and boundary smoothing) can further reduce salt-and-pepper noise and improve the integrity and control of wetland patch boundaries. Figure 1 To the point of being responsive.

[0041] Furthermore, to obtain results that can be used for mapping and management, the classification raster results can be post-processed as necessary (such as small patch denoising, connected component filtering and boundary smoothing) and vectorized results and area statistics can be output to support wetland resource surveys, dynamic monitoring and ecological management applications.

Claims

1. A wetland classification method based on remote sensing imagery and Mamba network, characterized in that, The method includes: Acquire multispectral remote sensing images of the study area and preprocess the remote sensing images; The preprocessed multispectral remote sensing images are labeled to construct a training dataset for wetland classification. The training dataset is input into a Mamba-based classification network for training, outputting wetland classification results. The classification network includes a feature embedding layer, a Mamba backbone module, a temporal attention module, a spatial attention module, and a classification output layer. Specifically, the feature embedding layer maps preprocessed multispectral remote sensing images into feature representations, obtaining features for each temporal phase; the Mamba backbone module encodes these temporal features and establishes contextual relationships, obtaining encoded features for each temporal phase; the temporal attention module adaptively weights and fuses these encoded features, obtaining fused features; the spatial attention module recalibrates the fused features in both spatial and channel dimensions, obtaining enhanced feature maps; and the classification output layer outputs pixel-level wetland classification results. The Mamba backbone module is used to encode features of each time phase and establish contextual associations to obtain encoded features of each time phase. Specifically, it unfolds the preprocessed multispectral remote sensing images into a sequence according to a predetermined scanning order, after feature embedding of the features of each time phase. By utilizing state-space recursion to characterize long-range dependencies, the global context modeling capability is enhanced while maintaining computational efficiency. The mathematical expression is: ， in The hidden state is the state representation of the current position t and its preceding context information during the sequence recursion process. It is the state representation of position t-1 and its preceding context information during the sequence recursion process, where A, B, and C are training parameters. The output is the encoded features corresponding to each sequence position. For sequence; The spatial attention module is used to recalibrate the fused features in both spatial and channel dimensions. The specific algorithm implementation is as follows: ， in This indicates that convolutions with different dilation rates are used to expand the receptive field. GAP represents global average pooling, S is the spatial weight map, w is the channel weight vector, sigmoid() is the activation function, and ReLU() is the rectified linear activation function. and These are two sets of linear transformation parameters in the channel attention branch. Features after fusion This is the enhanced feature map after spatial attention and channel attention weighting.

2. The method according to claim 1, characterized in that, The time attention module is used to adaptively weight and fuse the encoded features of each time phase to obtain the fused features. Specifically, it performs weighted fusion of T encoded features, with the weights adaptively calculated by the attention mechanism to highlight key time phases and suppress noisy time phases. The mathematical expression is as follows: ， Where q, , Let be the query, key, and value obtained by linear transformation of the encoding features of each time phase, respectively, where d is the feature dimension, and softmax() is the normalization function. The attention weight coefficient corresponding to the i-th temporal phase encoded feature is used to characterize the contribution of that temporal phase to the temporal feature fusion. These are the features after fusion.

3. The method according to claim 1, characterized in that, In the training of the Mamba-based classification network, features from different levels / scales are first aligned to a unified spatial scale, and then weighted convergence is performed using fusion weights. The fusion weights are learned end-to-end by the network or adaptively generated by the attention mechanism, and satisfy the overall normalization constraint.

4. The method according to claim 1, characterized in that, When training the Mamba-based classification network, iterative training is performed on the classification network based on training samples, and the network parameters are updated through backpropagation; wherein, the training objective is to minimize the loss function, and the loss function includes at least pixel-level classification loss.

5. The method according to claim 4, characterized in that, When training the Mamba-based classification network, training samples are selected from land use data and publicly available high-resolution image data, and the wetland classification results are verified using field survey data.

6. The wetland classification method according to any one of claims 1-5, characterized in that, The multispectral remote sensing images include a multi-temporal image sequence of the same study area. The multi-temporal image sequence is filtered according to a preset time window and registered and aligned according to the same spatial reference for subsequent temporal feature modeling of the temporal attention mechanism.

7. The wetland classification method according to any one of claims 1-5, characterized in that, The preprocessing includes: masking clouds, shadows and invalid pixels based on scene classification layers; stitching and cropping multi-temporal images; and resampling and normalizing each band to ensure that each temporal image meets the same spatial resolution and band scale.

8. The wetland classification method according to any one of claims 1-5, characterized in that, The sample annotation includes: dividing the preprocessed multispectral remote sensing image into image blocks according to a preset window, generating a corresponding pixel-level wetland category label for each image block, and constructing a training dataset in the form of image block-label pairs.

Images