CN-122024064-A - Mangrove extraction method and device based on multi-source data progressive correction

CN122024064ACN 122024064 ACN122024064 ACN 122024064ACN-122024064-A

Abstract

The invention discloses a mangrove extraction method and device based on multi-source data progressive correction, and belongs to the technical field of remote sensing monitoring. The method comprises the steps of obtaining a Sentinel-2 multispectral image, a GF-2 multispectral image, a field unmanned aerial vehicle remote sensing image and DEM data, carrying out preliminary extraction on mangrove based on the spectral index of the Sentinel-2 multispectral image and carrying out error correction based on the field unmanned aerial vehicle remote sensing image, carrying out correction on the boundary of a primarily extracted mangrove area based on the GF-2 multispectral image by adopting a machine learning method and carrying out error correction based on the field unmanned aerial vehicle remote sensing image, constructing a terrain correction index based on the DEM data and the GF-2 multispectral image, and removing a potential ponding area from the corrected mangrove area to obtain a final mangrove distribution area. The invention adopts a progressive frame of spectrum extraction, geometric correction and topography optimization, solves the problem of low mangrove extraction precision caused by spectrum confusion from the topography hydrology factor, and improves the mangrove extraction precision.

Inventors

REN ZHIHUA
SHEN ZHEN
MA YI
LI YI
MA WANDONG
WANG QIWEI
SHI YUANLI
ZHANG XINSHENG
Tai Wenfei
CHEN XUHUI
LI JING

Assignees

生态环境部卫星环境应用中心

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (9)

1. The mangrove extraction method based on the step-by-step correction of multi-source data is characterized by comprising the following steps: S1, acquiring multi-source data of a research area in the same time period; the multi-source data comprise a middle-resolution Sentinel-2 multi-spectrum image, a high-resolution GF-2 multi-spectrum image, a field unmanned aerial vehicle remote sensing image and DEM data; s2, preprocessing the Sentinel-2 multispectral image, the GF-2 multispectral image and the field unmanned aerial vehicle remote sensing image; S3, primarily extracting mangrove based on the spectrum index of the preprocessed Sentinel-2 multispectral image, and performing error correction based on the remote sensing image of the unmanned aerial vehicle to obtain a primarily extracted mangrove area; S4, correcting the boundary of the initially extracted mangrove area by adopting a machine learning method based on the preprocessed GF-2 multispectral image, and correcting errors based on the remote sensing image of the unmanned aerial vehicle to obtain a corrected mangrove area; S5, constructing a terrain correction index based on the DEM data and the preprocessed GF-2 multispectral image, and removing a potential ponding region from the corrected mangrove region according to the terrain correction index to obtain a final mangrove distribution region; the calculation formula of the terrain correction index is as follows: for terrain modification index, S, RE and ESD are grade, relative elevation and elevation standard deviation, respectively, obtained from the DEM data, For the normalized water index calculated according to the pretreated GF-2 multispectral image, a, b, c, d is the weight of the gradient, the relative elevation, the elevation standard deviation and the normalized water index respectively.
2. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 1, wherein S3 includes: S31, cutting the preprocessed Sentinel-2 multispectral image by using a boundary vector of a research area to obtain resolution image data in the research area; s32, calculating a spectrum index of each pixel in the research area based on the resolution image data in the research area; the spectrum index comprises one or more of a red edge normalized vegetation index, a red edge position index and a water stress index, and the calculation formula is as follows: RE_NDVI is a red edge normalized vegetation index, REP is a red edge position index, MSI is a water stress index, and B4, B5, B7 and B8A, B11 respectively represent remote sensing reflectances of 4 th, 5 th, 7 th, 8A and 11 th wave bands of the pretreated Sentinel-2 multispectral image; s33, comparing each spectrum index with each set index threshold value to obtain a mangrove spectrum extraction result.
3. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 2, wherein S33 includes: s331, comparing a red edge normalized vegetation index of each pixel in a research area with a set vegetation index threshold, and dividing pixels with the red edge normalized vegetation index being greater than the vegetation index threshold into mangroves to obtain a first spectrum extraction result; S332, comparing a red edge position index of each pixel in a research area with a set position index threshold value, and dividing the pixels with the red edge position indexes larger than the position index threshold value into mangroves to obtain a second spectrum extraction result; S333, comparing the water stress index of each pixel in the research area with a set water stress index threshold value, and dividing pixels with water stress indexes smaller than the water stress index threshold value into mangroves to obtain a third spectrum extraction result; S334, the first spectrum extraction result, the second spectrum extraction result and the third spectrum extraction result are combined to obtain the mangrove spectrum extraction result.
4. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 3, wherein S3 further includes: s34, obtaining real points of the mangrove forest according to the remote sensing image of the field unmanned aerial vehicle, and taking the real points as verification sample points; s35, comparing the verification sample points with the mangrove spectrum extraction result to obtain a first error classification set; wherein the first misclassification set includes misclassification points that classify non-mangrove forest as mangrove forest and misclassification points that classify mangrove forest as non-mangrove forest; S36, when the ratio of the misclassification points in the first misclassification set exceeds a set proportion threshold, adjusting the index threshold, and re-extracting to obtain a mangrove spectrum extraction result; When the index threshold is adjusted, the values of the vegetation index threshold and the position index threshold are improved, and the value of the water stress index threshold is reduced; S37, constructing a buffer area outwards by taking the missed partition point in the first error classification set as a center to obtain a missed partition distribution area; S38, redetermining an index threshold according to the spectral characteristics of the missed distribution region, and carrying out mangrove extraction on the missed distribution region by using the redetermined index threshold to obtain an extraction result of the missed distribution region; s39, taking a union set of the mangrove spectrum extraction result and the missed area extraction result to obtain the primary mangrove area.
5. The mangrove extraction method based on progressive correction of multi-source data as set forth in any one of claims 1-4, wherein S4 includes: s41, constructing buffer areas with set distances to the inner side and the outer side of the boundary based on the boundary of the primary extraction mangrove area to obtain the primary extraction mangrove boundary area; S42, cutting the preprocessed GF-2 multispectral image based on the boundary vector of the primary extracted mangrove boundary area to obtain a boundary area to be corrected; S43, training a random forest-based machine learning model through a selected training set comprising positive samples and negative samples, and extracting mangrove forests in the boundary area to be corrected through the trained machine learning model to obtain a mangrove forest extraction result in the boundary area; S44, comparing the verification sample points with the mangrove extraction results of the boundary area to obtain a second error classification set; Wherein the second set of false classifications includes false classification points that classify non-mangrove forest as mangrove forest and missed classification points that classify mangrove forest as non-mangrove forest; S45, performing fine adjustment training on the machine learning model trained by the training set by taking the false classification point and the missing classification point of the second false classification set as a negative sample and a positive sample respectively, and extracting mangrove in the boundary area to be corrected through the machine learning model subjected to fine adjustment training to obtain a mangrove boundary correction result; and S46, replacing the primary extracted mangrove border area in the primary extracted mangrove area by using the mangrove border correction result to obtain the corrected mangrove area.
6. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 5, wherein S5 includes: S51, calculating the gradient, the relative elevation and the elevation standard deviation of each pixel in the research area by using geographic information system software based on DEM data; S52, calculating a normalized water index of each pixel in the research area according to the preprocessed GF-2 multispectral image; the NDWI is a normalized water index, and Green and NIR respectively represent the remote sensing reflectivities of Green wave bands and near infrared wave bands of the pretreated GF-2 multispectral image; S53, constructing a terrain correction index of each pixel in the research area by the following formula: S54, comparing the topography correction index of each pixel in the research area with a determined segmentation threshold value, and dividing the pixels with topography correction indexes larger than the segmentation threshold value into potential ponding areas; And S55, removing the potential ponding area from the corrected mangrove area to obtain a final mangrove distribution area.
7. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 6, wherein the weights for the slope, relative elevation, standard deviation of elevation, and normalized water index are determined by: Determining subjective weights of gradient, relative elevation, elevation standard deviation and normalized water index by adopting an analytic hierarchy process; Determining objective weights of gradient, relative elevation, elevation standard deviation and normalized water index by adopting CRITIC method; And respectively carrying out linear weighting on the subjective weight and the objective weight to obtain weights of gradient, relative elevation, elevation standard deviation and normalized water index.
8. The mangrove extraction method based on progressive correction of multi-source data as set forth in claim 6, wherein the segmentation threshold is determined by: Performing unsupervised clustering fitting on the data set of the terrain correction index by adopting a Gaussian mixture model to obtain two Gaussian distributions; and taking the value of the terrain correction index corresponding to the intersection point of the probability density functions of the two Gaussian distributions as the segmentation threshold.
9. Mangrove extraction device based on multisource data correction step by step, characterized in that the device includes: The data acquisition module is used for acquiring multi-source data of the same time period in the research area; the multi-source data comprise a middle-resolution Sentinel-2 multi-spectrum image, a high-resolution GF-2 multi-spectrum image, a field unmanned aerial vehicle remote sensing image and DEM data; the preprocessing module is used for preprocessing the Sentinel-2 multispectral image, the GF-2 multispectral image and the remote sensing image of the field unmanned aerial vehicle; The primary extraction module is used for carrying out primary extraction on the mangrove based on the spectrum index of the preprocessed Sentinel-2 multispectral image, and carrying out error correction based on the remote sensing image of the unmanned aerial vehicle in the field to obtain a primary extraction mangrove area; The boundary correction module is used for correcting the boundary of the initially extracted mangrove area by adopting a machine learning method based on the preprocessed GF-2 multispectral image, and carrying out error correction based on the remote sensing image of the field unmanned aerial vehicle to obtain a corrected mangrove area; the terrain correction module is used for constructing a terrain correction index based on the DEM data and the preprocessed GF-2 multispectral image, and removing the potential ponding region from the corrected mangrove region according to the terrain correction index to obtain a final mangrove distribution region; the calculation formula of the terrain correction index is as follows: for terrain modification index, S, RE and ESD are grade, relative elevation and elevation standard deviation, respectively, obtained from the DEM data, For the normalized water index calculated according to the pretreated GF-2 multispectral image, a, b, c, d is the weight of the gradient, the relative elevation, the elevation standard deviation and the normalized water index respectively.

Description

Mangrove extraction method and device based on multi-source data progressive correction Technical Field The invention belongs to the technical field of remote sensing monitoring, and particularly relates to a mangrove extraction method and device based on multi-source data gradual correction. Background Mangrove is a precious ecological system distributed in the intertidal zones of tropical and subtropical coasts, and has important ecological functions of wind prevention, wave elimination, water quality purification, carbon fixation, carbon storage, biological diversity maintenance and the like. The spatial distribution and dynamic change of the mangrove forest can be mastered timely and accurately, and the mangrove forest protection, repair and management work can be developed. The accurate sample data provided by the traditional method for monitoring the mangrove in the field lays a foundation for research, but due to the specificity of the mangrove distribution area, the field sampling is time-consuming and labor-consuming, and the spatial continuity of the obtained data is poor. Remote sensing technology capable of remote detection is an important irreplaceable means for mangrove monitoring due to its macroscopic and rapid advantages. Along with the increasing enrichment of remote sensing data sources and the continuous progress of analysis methods, mangrove information extraction technology presents an overall trend of evolution from relying on a single data source and a simple algorithm to fusion of multi-source data and a complex model. However, in this development, the severe spectrum confusion phenomenon (i.e. "foreign body homography") between mangrove and adjacent stationary water-bearing regions in flat intertidal zones is always a core challenge and difficulty for high-precision extraction. To overcome this difficulty, the prior art has experienced development paths that each, while improving accuracy, also expose their inherent limitations. The single data source method based on the spectral characteristics is strong in universality but limited in precision and detail. The method is the basis of mangrove remote sensing monitoring, and mainly relies on spectral information of resolution optical remote sensing images (such as Landsat series images and Sentinel-2) in a single type, and extraction is performed by calculating spectral indexes (such as normalized vegetation index NDVI, normalized water index NDWI and the like) and setting an empirical threshold. The method has the advantages of strong universality and high calculation efficiency, and is suitable for large-scale rapid census. However, the inherent limitation is that the mixed pixel effect of the medium resolution image causes the boundary blurring and detail loss of the extraction result, especially in the spectrum confusion zone where mangrove and water are distributed in a staggered way, the problem of misclassification is prominent, and the extraction precision is restricted. Even if the wave band information of a plurality of images is used, the fine distinction of the land-water junction of the flat area is difficult to be well solved. With the rapid development of remote sensing technology, the spatial resolution of the image is gradually improved. The method uses high spatial resolution images (such as GF-2 and WorldView series), adopts a machine learning algorithm, comprehensively digs spectral features and spatial texture features of the images, remarkably optimizes geometric outlines of mangrove plaques, and realizes finer identification than medium resolution images. However, the application of high resolution images is often accompanied by the construction of Gao Weite's collection, which is very prone to introducing feature redundancy. Feature redundancy does not contribute to jump of model performance, but can dilute contribution degree of spectrum features really critical for distinguishing mangrove forest from water body, and influence generalization capability of the model. In addition, the high-resolution images are mostly multispectral data, the spectral wave bands are wider, the wave band number is smaller, and therefore, the accurate extraction of mangrove is difficult to be carried out by fully utilizing the spectral information. More importantly, even though the spatial resolution of the image is improved to sub-meter level, the problem of spectrum confusion between mangrove and stationary water accumulation areas caused by foreign matter homography is not solved well. The limitation of a single data source makes multi-source data fusion a major approach to mangrove research. The current multi-source data fusion strategy is a common idea of combining DEM data with optical images. However, the existing fusion method still stays at the 'feature superposition' level, namely, the DEM and the spectral features are input into a classifier together, and the information provided by the DEM data is not fully utilized. D