Multi-source remote sensing fusion model and method for coastal wetland vegetation species level classification

Abstract

The invention belongs to the technical field of multi-source remote sensing data processing and fusion, discloses a multi-source remote sensing fusion model and method for coastal wetland vegetation species level classification, and aims to solve the problems of insufficient resolution, low species identification precision and poor multi-source data complementarity in the prior art. According to the scheme, the method comprises the following steps: preprocessing sentinel No.1 SAR data and sentinel No.2 optical data, and unifying the data to 10-meter resolution; constructing an SAR feature set and an optical spectrum phenological feature set; a CMVN double-branch model is adopted, after preprocessing is conducted through a head module, features are fused through six SOCM cross-modal fusion modules, and after field label training is combined, six types of vegetation species-level results such as spartina alterniflora are output. The overall precision is 0.916, the anti-interference performance is high, and support is provided for dynamic monitoring and ecological evaluation of the wetland.

Description

Technical Field

[0001] This invention belongs to the field of multi-source remote sensing data processing and fusion technology, specifically relating to a multi-source remote sensing fusion model and method for species-level classification of coastal wetland vegetation. Background Technology

[0002] Coastal wetlands are critical ecosystems, and monitoring their vegetation dynamics is crucial for biodiversity conservation and coastal zone management. Existing wetland vegetation classification technologies have several limitations: traditional optical remote sensing imagery is susceptible to cloud cover and atmospheric effects, limiting data availability; single SAR images struggle to accurately distinguish spectrally similar vegetation; existing classification products are mostly ecosystem-level classifications, failing to achieve precise species-level identification, and suffer from low spatial resolution (mostly 30 meters) and insufficient temporal continuity, making it difficult to capture short-term dynamic changes in wetland vegetation. Furthermore, existing multi-source data fusion methods fail to effectively integrate SAR structural water information with spectral phenological information from optical images, resulting in classification accuracy insufficient for species-level monitoring needs, thus hindering the development of coastal wetland ecological research and precise management. Summary of the Invention

[0003] To overcome the above-mentioned technical problems, this invention provides a multi-source remote sensing fusion model and method for species-level classification of coastal wetland vegetation.

[0004] The present invention adopts the following technical solution: A multi-source remote sensing fusion model for species-level classification of coastal wetland vegetation is proposed. It employs a dual-branch architecture, processing synthetic aperture radar (SAR) feature sets and optical spectral phenological feature sets respectively. Each branch contains a head module, which consists of a 3×3 two-dimensional convolution, two-dimensional batch normalization, and a ReLU6 activation function. The core of the model is a cross-modal fusion module (SOCM), used to achieve interaction and fusion of the two types of features. The fused features are then compressed in dimension through global average pooling and mapped to the coastal wetland vegetation species classification results via a fully connected layer.

[0005] Preferably, the processing steps of the cross-modal fusion module include: dividing the input features into three equal parts according to the channel dimension, and processing them respectively through window attention, shifted window attention, and long-distance attention; mapping the output features of the SAR branch to queries and the output features of the optical branch to keys and values; calculating the intermodal correlation weights through scaling dot product attention, and using these weights to weight the optical features, and then superimposing them with the SAR features to obtain the cross-modal fusion features.

[0006] Preferably, the cross-modal fusion module consists of 6 modules connected in series; the classification results output by the fully connected layer correspond to six major coastal wetland vegetation types: Spartina alterniflora, Phragmites australis, Suaeda salsa, Curcuma longa, mangroves, and Tamarix chinensis, and the spatial resolution of the classification results is 10 meters.

[0007] This invention also discloses a multi-source remote sensing fusion method for species-level classification of coastal wetland vegetation, comprising the following steps: S1: Perform targeted preprocessing on Sentinel-1 SAR data and Sentinel-2 optical data, and unify them to a spatial resolution of 10 meters; S2: Based on the preprocessed data, construct SAR feature sets and optical spectral phenological feature sets respectively; S3: Combine the SAR feature set and the optical spectral phenological feature set as input data, and combine the field vegetation type labels to divide the training set and the validation set in a 7:3 ratio to train the multi-source remote sensing fusion model as described in any one of claims 1-3. S4: Apply the trained model to multi-source remote sensing data of the target area, and output species-level classification results of coastal wetland vegetation through feature extraction, cross-modal fusion and classification calculation.

[0008] Preferably, the preprocessing of Sentinel-1 SAR data in step S1 includes: sequentially performing orbit correction, thermal noise removal, amplitude calibration, speckle noise filtering, and terrain correction, and then converting it to decibel format before using the mean function for annual synthesis.

[0009] Preferably, the preprocessing of the Sentinel-2 optical data in step S1 includes: selecting the two phases of the vigorous growth period and the withering period, extracting the four bands of blue light, green light, red light and near-infrared light and performing resampling and band fusion, and calculating the normalized vegetation index based on the red light and near-infrared bands.

[0010] Preferably, the construction of the SAR feature set in step S2 includes: selecting VV and VH single-polarization data, and simultaneously constructing three types of derived features: polarization backscattering normalized difference index, sum of dual-polarization echo intensity, and dual-polarization difference index.

[0011] Preferably, the construction of the optical spectral phenological feature set in step S2 includes: fusing eight optical bands from two time phases with two normalized vegetation indices to form a spectral phenological feature vector.

[0012] Preferably, when dividing the training set and validation set in step S3, it is necessary to ensure that the samples of each vegetation category are evenly distributed; the model training process needs to complete feature preprocessing, cross-modal fusion and parameter optimization for classification calculation.

[0013] Preferably, the species-level classification results output in step S4 correspond to six types of vegetation: Spartina alterniflora, Phragmites australis, Suaeda salsa, Curcuma longa, mangroves, and Tamarix chinensis, and the overall accuracy of the classification results reaches 0.916, with a Kappa coefficient of 0.898.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a coastal wetland vegetation classification model and method that can integrate multi-source remote sensing data, achieve high-precision species-level classification, and has strong anti-interference capabilities. It solves the problems of insufficient resolution, low species identification accuracy, and underutilization of the complementarity of multi-source data in existing classification technologies, and provides technical support for dynamic monitoring and ecological assessment of coastal wetlands.

[0015] High classification accuracy: Through multi-source data fusion and attention mechanism optimization, the overall classification accuracy reaches 0.916 and the Kappa coefficient is 0.898, which can stably distinguish wetland vegetation species with similar spectra or scattered distribution. Achieve precise species-level classification: Breaking through the limitations of existing product ecosystem-level classification, it can accurately identify 6 types of core wetland vegetation, providing support for research on species distribution patterns and the expansion of invasive species; Strong complementarity of multi-source data: It effectively integrates the all-weather monitoring advantages of SAR with the spectral phenological information of optical images, has outstanding resistance to cloud cover and tidal interference, and is suitable for complex coastal environments; High spatial resolution: It supports 10-meter spatial resolution classification, which can accurately depict the distribution of small-scale wetland patches and narrow intertidal vegetation, meeting the needs of refined monitoring; Adapted for long-term monitoring: The model structure is adapted to the processing of annual continuous remote sensing data, which can support the tracking of long-term dynamic changes in coastal wetland vegetation, providing technical support for ecological restoration and coastal zone management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the dual-branch fusion and classification link of the CMVN model of this invention. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. Unless otherwise specified, the raw materials and equipment used can be purchased from the market or are commonly used in the art. The methods in the embodiments, unless otherwise specified, are conventional methods in the art. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The multi-source remote sensing fusion model and method for species-level classification of coastal wetland vegetation are as follows: First, targeted preprocessing was performed on two types of core remote sensing data. For the Level-1 ground distance detection data from Sentinel-1 synthetic aperture radar, trajectory correction, thermal noise removal, amplitude calibration, speckle noise filtering, and terrain correction were carried out sequentially. After converting the data to decibel format, annual synthesis was performed using the mean function to reduce the interference of tidal dynamics and seasonal fluctuations of salt marshes. For the Sentinel-2 optical imagery, the peak and off-peak periods were selected in combination with the phenological characteristics of coastal vegetation. The four bands of blue, green, red, and near-infrared light were extracted and resampled and fused. Then, the normalized vegetation index was calculated based on the red and near-infrared bands to highlight the differences in vegetation growth status. Finally, the two types of preprocessed data were unified to a spatial resolution of 10 meters to ensure the consistency and comparability of multi-source data.

[0019] In the feature construction stage, two types of core feature sets were constructed based on the preprocessed data: For synthetic aperture radar data, VV and VH single-polarization data were selected, and three types of derived features were constructed simultaneously: polarization backscattering normalized difference index, sum of dual-polarization echo intensity, and dual-polarization difference index. These features accurately captured information on vegetation structure, total water content, and surface roughness. For optical data, eight optical bands from two time phases and two normalized vegetation indices were fused to form a spectral phenological feature vector, thereby highlighting the differences in spectral response of vegetation at different growth stages and providing differentiated feature support for subsequent fusion.

[0020] The multi-source fusion classification model (CMVN) adopts a two-branch architecture design. like Figure 1 As shown, the four modules included are explained in detail below: Data preprocessing module: The upper part inputs preprocessed Sentinel-1 data, extracts VV and VH polarization and derived SAR difference index, SAR normalized vegetation index, and SAR sum index to capture vegetation structure and moisture characteristics; the lower part inputs dual-temporal Sentinel-2 data, extracts blue, green, red, and near-infrared four bands and dual-temporal normalized vegetation index to capture vegetation spectral and phenological differences, providing basic features for subsequent fusion.

[0021] Feature preprocessing module: Both branches use the head module (3×3 two-dimensional convolution + two-dimensional batch normalization + ReLU6 activation function) to achieve feature standardization and non-linear enhancement - convolution extracts local features, batch normalization accelerates training, and ReLU6 alleviates gradient vanishing, ensuring that the two types of features can be fused in the same dimension.

[0022] The core module for cross-modal fusion is composed of six cascaded synthetic aperture radar-optical cross-modal fusion modules. The input features are divided into three groups according to the channel and processed by window attention, shift window attention, and long-range attention, respectively. Then, synthetic aperture radar features are used as the guide for querying, and optical features are used as the guide for key and value. The two types of features are dynamically weighted and fused through attention weight calculation to enhance the distinguishability of spectrally similar vegetation.

[0023] Classification output module: The fused features are compressed in dimensionality through global average pooling, and then mapped to the classification results of 6 types of vegetation through a fully connected layer. Finally, a 10-meter resolution distribution map of coastal wetland vegetation species is output.

[0024] The dual-branch architecture has one branch processing synthetic aperture radar feature sets and the other branch processing optical spectral phenological feature sets. Both branches perform feature preprocessing through a head module, which consists of 3×3 two-dimensional convolution, two-dimensional batch normalization, and ReLU6 activation function. This module can extract local features, achieve feature standardization, and alleviate the gradient vanishing problem through ReLU6, ensuring that the two types of features have a fusion basis in the same dimension. The core of the model lies in the cross-modal fusion module (SOCM), through which each branch achieves feature interaction and fusion: First, the input features are divided into three equal parts according to the channel dimension, and then processed by window attention, shifted window attention, and long-distance attention respectively. Window attention captures local details, shifted window attention solves the problem of lost window boundary information, and long-distance attention captures global dependencies. Then, the output features of the synthetic aperture radar branch are mapped to queries, and the output features of the optical branch are mapped to keys and values. The correlation weight between modes is calculated by scaling dot product attention. After the optical features are weighted using this weight, they are superimposed with the synthetic aperture radar features to obtain cross-modal fused features. This preserves the sensitivity of synthetic aperture radar to vegetation structure, while adaptively absorbing spectral and phenological information from optical images to enhance feature discriminative power.

[0025] During the classification execution phase, the constructed synthetic aperture radar feature set and optical spectral phenological feature set are combined as input data. Combined with the vegetation type labels collected in the field, the data is divided into training and validation sets in a 7:3 ratio to ensure a balanced distribution of samples from each vegetation category. This data is used to train the multi-source fusion classification model. After training, the model is applied to multi-source remote sensing data of the target area. Through feature extraction, cross-modal fusion, and classification calculation, the final output is the species-level classification results of six major coastal wetland vegetation types: Spartina alterniflora, Phragmites australis, Suaeda salsa, Curcuma longa, mangroves, and Tamarix chinensis.

[0026] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-source remote sensing fusion model for species-level classification of coastal wetland vegetation, characterized in that, A dual-branch architecture is adopted to process the synthetic aperture radar (SAR) feature set and the optical spectral phenological feature set respectively. Each branch contains a head module, which consists of a 3×3 two-dimensional convolution, two-dimensional batch normalization, and a ReLU6 activation function. The core of the model is the cross-modal fusion module SOCM, which is used to realize the interaction and fusion of the two types of features. After the fused features are compressed in dimension by global average pooling, they are mapped to the coastal wetland vegetation species classification results through a fully connected layer.

2. The multi-source remote sensing fusion model according to claim 1, characterized in that, The processing steps of the cross-modal fusion module include: dividing the input features into three equal parts according to the channel dimension, and processing them respectively through window attention, shifted window attention, and long-distance attention; mapping the output features of the SAR branch to queries and the output features of the optical branch to keys and values; calculating the intermodal correlation weights through scaling dot product attention, and using these weights to weight the optical features, and then superimposing them with the SAR features to obtain the cross-modal fusion features.

3. The multi-source remote sensing fusion model according to claim 1, characterized in that, The cross-modal fusion module consists of 6 connected modules; the classification results output by the fully connected layer correspond to the six major coastal wetland vegetation types: Spartina alterniflora, Phragmites australis, Suaeda salsa, Curcuma longa, mangroves, and Tamarix chinensis, with a spatial resolution of 10 meters.

4. A multi-source remote sensing fusion method for species-level classification of coastal wetland vegetation, characterized in that, Includes the following steps: S1: Perform targeted preprocessing on Sentinel-1 SAR data and Sentinel-2 optical data, and unify them to a spatial resolution of 10 meters; S2: Based on the preprocessed data, construct SAR feature sets and optical spectral phenological feature sets respectively; S3: Combine the SAR feature set and the optical spectral phenological feature set as input data, and combine the field vegetation type labels to divide the training set and the validation set in a 7:3 ratio to train the multi-source remote sensing fusion model as described in any one of claims 1-3. S4: Apply the trained model to multi-source remote sensing data of the target area, and output species-level classification results of coastal wetland vegetation through feature extraction, cross-modal fusion and classification calculation.

5. The multi-source remote sensing fusion method according to claim 4, characterized in that, The preprocessing of Sentinel-1 SAR data in step S1 includes: sequentially performing orbit correction, thermal noise removal, amplitude calibration, speckle noise filtering, and terrain correction, and then converting it to decibel format before using the mean function for annual synthesis.

6. The multi-source remote sensing fusion method according to claim 4, characterized in that, The preprocessing of Sentinel-2 optical data in step S1 includes: selecting the two phases of the vigorous growth period and the withering period, extracting the four bands of blue light, green light, red light and near-infrared light and performing resampling and band fusion, and calculating the normalized vegetation index based on the red light and near-infrared bands.

7. The multi-source remote sensing fusion method according to claim 4, characterized in that, The construction of the SAR feature set in step S2 includes: selecting VV and VH single-polarization data, and simultaneously constructing three types of derived features: polarization backscattering normalized difference index, sum of dual-polarization echo intensity, and dual-polarization difference index.

8. The multi-source remote sensing fusion method according to claim 4, characterized in that, The construction of the optical spectral phenological feature set in step S2 includes: fusing eight optical bands from two time phases with two normalized vegetation indices to form a spectral phenological feature vector.

9. The multi-source remote sensing fusion method according to claim 4, characterized in that, When dividing the training set and validation set in step S3, it is necessary to ensure that the samples of each vegetation category are evenly distributed; the model training process needs to complete feature preprocessing, cross-modal fusion and parameter optimization for classification calculation.

10. The multi-source remote sensing fusion method according to claim 4, characterized in that, The species-level classification results output in step S4 correspond to six vegetation types: Spartina alterniflora, Phragmites australis, Suaeda salsa, Curcuma longa, mangroves, and Tamarix chinensis. The overall accuracy of the classification results is 0.916, and the Kappa coefficient is 0.898.

Images