CN-121259613-B - Mangrove forest dense time sequence detection method and device

CN121259613BCN 121259613 BCN121259613 BCN 121259613BCN-121259613-B

Abstract

The invention provides a mangrove forest dense time sequence detection method and device, and relates to the technical field of target identification, wherein the method comprises the steps of obtaining a radar image and an optical remote sensing image of a region to be detected; the method comprises the steps of invoking a pre-built double-branch feature extraction network to conduct feature extraction on a radar image and an optical remote sensing image respectively and independently to obtain a radar feature and an optical remote sensing feature, conducting feature fusion on the radar feature and the optical remote sensing feature to obtain a first fusion feature, calculating to obtain an SSMI feature based on reflectivity data of the radar and the optical remote sensing image, conducting feature fusion on the first fusion feature and the SSMI feature to obtain a second fusion feature, invoking a pre-built target network to process the second fusion feature to obtain a candidate detection result, enabling the target network to be used for simultaneously identifying a change position and a change type of a target object, and conducting space-time consistency correction on the candidate detection result to obtain the detection result for output.

Inventors

YAN JINING
SUN HAONAN
He Haixu
YANG SUZHEN
Jia Renming
ZHANG ZHENG

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260505
Application Date: 20251204

Claims (8)

1. The mangrove forest dense time sequence detection method is characterized by comprising the following steps of: obtaining a radar image and an optical remote sensing image of a region to be detected, wherein the region to be detected comprises mangrove as a main detection object; Invoking a pre-built double-branch feature extraction network to perform feature extraction on the radar image and the optical remote sensing image respectively and independently to obtain radar features and optical remote sensing features, wherein the double-branch feature extraction network is provided with two mutually independent feature extraction branches, and feature extraction methods adopted by different feature extraction branches are different; performing feature fusion on the radar features and the optical remote sensing features to obtain first fusion features; calculating to obtain SSMI characteristics based on the radar and the reflectivity data of the optical remote sensing image; performing feature fusion on the first fusion feature and the SSMI feature to obtain a second fusion feature; Invoking a pre-constructed target network to process the second fusion characteristic to obtain a candidate detection result, wherein the target network is used for identifying the change position and the change type of the target object at the same time; Carrying out space-time consistency correction on the candidate detection result to obtain a detection result for output; the target network is a time sequence semantic segmentation network; and the invoking the pre-constructed target network to process the second fusion feature to obtain a candidate detection result, wherein the invoking comprises the following steps: Downsampling the second fusion feature through multi-order one-dimensional convolution operation, and respectively generating a first feature map; Gradually increasing receptive fields and compressed feature dimensions, and respectively adopting one-dimensional transposition convolutions with different sampling multiples to perform scale recovery and feature preservation aiming at feature graphs which are sampled and output in a designated stage; Splicing the stored multi-scale features in a jump connection mode, and generating the candidate detection result through up-sampling and convolution processing; The loss function in the time sequence semantic segmentation network is generated by fusing a cross entropy loss function and a Dice loss, and the loss value of the time sequence semantic segmentation network is determined by the sum of the cross entropy loss value and the Dice loss value: ; ; Wherein, the In order to cross-entropy loss calculation results, As a result of the computation of the Dice loss, And (3) with For the label value and the predicted value of the i-th time sequence point land cover type, N is the total number of time sequence points, And (3) with And B is the number of samples used for updating the model parameters in one training for the label value and the predicted value of whether the ith time sequence land coverage type is changed.
2. The mangrove dense timing detection method of claim 1, wherein the dual-branch feature extraction network includes a first feature extraction branch and a second feature extraction branch; Constructing the dual-branch feature extraction network, comprising: Preparing the first feature extraction branch for extracting the radar feature based on a network model having long timing processing capabilities, the network model including Mamba model; preparing the second feature extraction branch for extracting the optical remote sensing feature based on a lightweight convolution model, the lightweight convolution model comprising a one-dimensional convolution model.
3. The mangrove dense timing detection method of claim 1, wherein the feature fusion of the radar feature and the optical remote sensing feature to obtain a first fusion feature comprises: and combining a target mechanism to perform feature fusion on the radar features and the optical remote sensing features to obtain the first fusion features, wherein the target mechanism comprises an attention mechanism and a gating fusion mechanism.
4. The mangrove dense timing detection method of claim 3 wherein the combining the target mechanism to feature fuse the radar features with the optical remote sensing features comprises: Respectively extracting high-dimensional features from the radar features and the optical remote sensing features to obtain feature vectors X and Y; calculating an attention weight matrix W of the feature vectors X and Y through a dot product attention scaling mechanism; Multiplying the attention weight matrix W by the feature vector Y, and superposing the multiplication result and the feature vector X to obtain the first fusion feature.
5. The mangrove dense timing detection method of claim 1, wherein the feature fusing the first fused feature with the SSMI feature to obtain a second fused feature comprises: Registering the first fusion feature with the SSMI feature; Converting the SSMI characteristics after characteristic alignment into target dimensions through mapping processing, wherein the target dimensions are the same as the dimensions of the first fusion characteristics after characteristic alignment; splicing the SSMI features with the same dimension with a first fusion feature; determining an adaptive fusion weight of each time step; Separating the self-adaptive fusion weights to obtain a first weight corresponding to the SSMI characteristics and a second weight corresponding to the first fusion characteristics; And dynamically fusing the first weight, the second weight and the spliced SSMI characteristics of each time step with the first fusion characteristics to obtain the second fusion characteristics.
6. The mangrove dense timing detection method of claim 1, wherein the target network includes at least any one of a timing semantic segmentation network, a dual-branch contrast network, and a timing change detection network based on an attention mechanism.
7. The mangrove forest dense timing detection method of claim 1, wherein the performing space-time consistency correction on the candidate detection results to obtain detection results for output comprises: Modifying the center point type in a window area into a target type by using a first sliding window for the whole time sequence of one pixel point in the candidate detection result, wherein the target type is the land coverage type in the current window area, the occurrence frequency of the land coverage type is the largest, and the frequency meets the land coverage type in a preset numerical range; merging the sequence with the duration of the change of the land cover type smaller than the preset time range into the sequence which is nearest and has the longest duration of the land cover type which is stable; And carrying out mode filtering on the mangrove distribution extraction result in one time phase in the candidate detection results by adopting a second sliding window.
8. A mangrove dense timing detection apparatus, comprising: The acquisition module is used for acquiring a radar image and an optical remote sensing image of an area to be detected, wherein the area to be detected comprises mangrove as a main detection object; The first calling module is used for calling a pre-built double-branch feature extraction network to perform feature extraction on the radar image and the optical remote sensing image respectively and independently to obtain radar features and optical remote sensing features, the double-branch feature extraction network is provided with two mutually independent feature extraction branches, and feature extraction methods adopted by different feature extraction branches are different; The first fusion module is used for carrying out feature fusion on the radar features and the optical remote sensing features to obtain first fusion features; The calculation module is used for calculating and obtaining SSMI characteristics based on the reflectivity values of the radar image and the optical remote sensing image; The second fusion module is used for carrying out feature fusion on the first fusion feature and the SSMI feature to obtain a second fusion feature; The second calling module is used for calling a pre-constructed target network to process the second fusion characteristic to obtain a candidate detection result, and the target network is used for identifying the change position and the change type of the target object at the same time; the correction module is used for carrying out space-time consistency correction on the candidate detection results to obtain detection results for output; the target network is a time sequence semantic segmentation network; and the invoking the pre-constructed target network to process the second fusion feature to obtain a candidate detection result, wherein the invoking comprises the following steps: Downsampling the second fusion feature through multi-order one-dimensional convolution operation, and respectively generating a first feature map; Gradually increasing receptive fields and compressed feature dimensions, and respectively adopting one-dimensional transposition convolutions with different sampling multiples to perform scale recovery and feature preservation aiming at feature graphs which are sampled and output in a designated stage; Splicing the stored multi-scale features in a jump connection mode, and generating the candidate detection result through up-sampling and convolution processing; The loss function in the time sequence semantic segmentation network is generated by fusing a cross entropy loss function and a Dice loss, and the loss value of the time sequence semantic segmentation network is determined by the sum of the cross entropy loss value and the Dice loss value: ; ; Wherein, the In order to cross-entropy loss calculation results, As a result of the computation of the Dice loss, And (3) with For the label value and the predicted value of the i-th time sequence point land cover type, N is the total number of time sequence points, And (3) with And B is the number of samples used for updating the model parameters in one training for the label value and the predicted value of whether the ith time sequence land coverage type is changed.

Description

Mangrove forest dense time sequence detection method and device Technical Field The embodiment of the invention relates to the technical field of image data detection, in particular to a mangrove dense time sequence detection method and device. Background The existing mangrove forest dynamic monitoring technology has three problems that (1) the existing mangrove forest dynamic monitoring technology excessively depends on optical remote sensing data, but the availability of optical images is affected by cloud fog, so that the time resolution is low, the spectral response of the mangrove forest is highly similar to that of coastal vegetation, and the mangrove forest is difficult to accurately identify by simply depending on the optical data. (2) The time resolution is too low to meet the requirement of mangrove forest high-frequency monitoring in a specific area. (3) The tidal inundation problem seriously affects the spectral reflectivity of mangroves in remote sensing images, and is difficult to effectively distinguish mangroves from coastal vegetation and high-flush woods from water bodies. The existing mangrove distribution extraction and change detection methods are various. For example, for distributed extraction, existing methods can generally be divided into two categories. The first is sample-based supervised classification, which uses input data from the original band or vegetation index from the remote sensing image, and employs a conventional machine learning or deep learning classifier to map mangrove forests. However, both machine learning and deep learning methods require a large number of samples to train the classifier, and therefore both the quality of the samples and the selection of the classifier can have a significant impact on the classification results. The second category is an unsupervised classification based on exponential threshold segmentation, which is based on exponential threshold segmentation of the exponential amplitude to identify the target object. Because each index is formed by combining different spectrum bands, the method can express the characteristics of the target ground object more than a single band, but different thresholds are required to be set in different areas, and the mobility is weak. For mangrove change detection, existing methods such as CCDC, BFAST, etc. cannot provide detailed information of mangrove changes such as land cover types before and after the change. Some scholars propose a novel mangrove index (SPECTRAL AND SAR mangrove index, SSMI), through the band spectral sensitivity analysis of optical images and radar images, by utilizing the unique characteristics of mangrove in terms of optical characteristics (green and humidity) and SAR backscattering coefficients, the optical and SAR images are effectively combined, so that complementary spectral and spatial structure information is realized, and the mangrove characteristics are accurately captured. However, the method has the core targets of drawing of a static mangrove forest, incapability of capturing time sequence changes, incapability of outputting a single time point result, incapability of positioning changing time and types, extremely low time resolution, incapability of meeting high-frequency monitoring requirements, rough characteristic fusion mode, incapability of fully playing multi-source data complementarity, simple model framework and poor generalization capability and complex scene suitability. Still other scholars have proposed a mangrove condition continuous detection method based on time series Landsat images. The method first uses a tidal wetland variation identification and characterization (DECODE) algorithm, which is an existing dense time series model, to detect disturbances in tidal wetlands that are subject to tidal fluctuations. The algorithm is well suited for detecting tidal wetland disturbances, but does not provide satisfactory post-disturbance monitoring results due to the large variance of post-disturbance Landsat observations. In order to better monitor post-disturbance conditions, a new time series fitting method, namely, DECODER (DECODE and recovery), is proposed for the recovery phase. Furthermore, for time periods divided by disturbance events, a random forest classifier is constructed which includes time-spectral variables derived from a time-series model to characterize mangrove conditions. However, the method has the defect that the dependence on the optical image is too high, and effective monitoring of mangrove forest can not be realized in a cloudy rainy season or in weather. Disclosure of Invention In order to solve the technical problems, an embodiment of the present invention provides a mangrove dense timing detection method, including: obtaining a radar image and an optical remote sensing image of a region to be detected, wherein the region to be detected comprises mangrove as a main detection object; Invoking a pre-built double-branch feature extrac