CN-121982535-A - River network extraction method based on river probability learning and self-adaptive stream combustion

CN121982535ACN 121982535 ACN121982535 ACN 121982535ACN-121982535-A

Abstract

The invention discloses a river network extraction method based on river probability learning and self-adaptive stream combustion, which comprises the steps of obtaining remote sensing images and DEM data of a target area, preprocessing, calculating multi-water body spectral indexes and linear geometric features based on the preprocessed data, fusing to form high-dimensional feature vectors, generating a multi-scale river probability map by using a training gradient lifting tree model, fusing to obtain a comprehensive river probability map, taking pixels with probability values meeting a threshold as seed points, extracting a river vector mask by using a directional constraint area growth algorithm, estimating the width of a local river channel according to the probability, dynamically calculating the combustion depth of each pixel based on the convergence accumulation amount, the width and the linear geometric features, and obtaining a hydrologic conditional DEM, and extracting a final river network from the DEM.

Inventors

Zang Yufu
CHU ZHAOCAI
CUI ZHEN

Assignees

南京信息工程大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A river network extraction method based on river probability learning and self-adaptive stream combustion is characterized by comprising the following steps: S1, acquiring remote sensing optical image and digital elevation model data of a target area, sequentially carrying out radiation correction, atmosphere correction and cloud shadow mask processing on the remote sensing optical image, resampling the processed remote sensing optical image and digital elevation model data to uniform spatial resolution, and obtaining a standardized input data pair; S2, calculating various water body spectral indexes of each pixel in the image based on the remote sensing optical image in the standardized input data pair, extracting linear geometric features of each pixel based on a multi-scale Hessian matrix, and fusing a spectral index set formed by the various water body spectral indexes with the linear geometric features to obtain a high-dimensional feature vector representing the spectral properties and the spatial morphology of the pixel; s3, training a river-oriented gradient lifting tree model by using the marked river/non-river sample data set by taking the high-dimensional feature vector as input, optimally learning the gradient lifting tree model by minimizing a loss function with regularization term, outputting a river attribution probability value of each pixel to generate a multi-scale river probability map, and performing cross-scale maximum fusion on the multi-scale river probability map to obtain a comprehensive river probability map; S4, taking pixels with probability values meeting a preset threshold value in the comprehensive river probability map as initial seed points, and adopting a directional constraint region growing algorithm to extract river vectors, wherein the region growing algorithm carries out double constraint through the river attribution probability values of the pixels and the consistency of local main directions of neighborhood pixels and current seed points, and outputs river vector mask data with continuous space and complete morphology; s5, estimating the width of a local river channel by using an automatic river width extraction algorithm based on the river vector mask data, dynamically calculating the combustion depth of each river channel pixel according to the confluence accumulation amount of each river channel pixel, the local river channel width and the maximum response value of linear geometric features by using a space self-adaptive stream combustion algorithm, carrying out terrain correction on the digital elevation model in the standardized input data pair by using the combustion depth to obtain a hydrologic conditional DEM, and extracting a final river network with an accurate space position and a correct topological communication relation from the hydrologic conditional DEM.
2. The method of claim 1, wherein the plurality of water spectrum indexes in step S2 include a normalized differential water index NDWI, an improved normalized differential water index MNDWI, and a multispectral water index MuWI, and the calculation modes of the indexes are as follows: wherein Green represents the Green light wave band pixel value of the remote sensing optical image, Representing the blue band pixel value of the remote sensing optical image, NIR representing the near infrared band pixel value, Representing the short-wave infrared band, SWIR1 representing the short-wave infrared 1 band pixel value, SWIR2 representing the short-wave infrared 2 band pixel value, And (5) representing the normalized difference value between the wave band i and the wave band j in the remote sensing optical image.
3. The method according to claim 2, wherein extracting the linear geometric feature of each pixel based on the multi-scale Hessian matrix in step S2 comprises: preprocessing an image corresponding to the multispectral water index MuWI by adopting a bias correction fuzzy C-means algorithm; For the pre-processed MuWI images, a scale space is generated by Gaussian kernel convolution for each scale Computing a Hessian matrix at pixel (x, y) : Wherein, the method comprises the steps of, The MuWI value representing pel (x, y) at scale sigma, 、 Second partial derivatives of the pixel (x, y) in the x, y directions respectively, 、 Is a mixed second partial derivative; For a pair of Decomposing the characteristic value to obtain the characteristic value , Satisfies the following conditions Calculating the linear geometrical response at that scale : Wherein ε is a stability constant; traversing a set of scales Where k is the number of scales, and the pixel is taken at all scales As the linear geometry of the picture element.
4. The method according to claim 1, wherein the training objective function of the gradient-lifted tree model in step S3 is: Wherein, the method comprises the steps of, The training objective function of the gradient lifting tree model is used, and N is the number of training samples; The term is fitted to the data and, A river/non-river genuine label representing the i-th sample, The feature vector for the i-th sample constructed for step S2, As a model output, the output is converted to river probabilities by a sigmoid function: M is the number of decision trees, eta is the learning rate for controlling the contribution of each decision tree, Representing an mth decision tree; Is a structural regularization term used for balancing model accuracy and generalization capability.
5. The method of claim 4, wherein the data fitting term For log-cross entropy loss, the expression is: Wherein, the method comprises the steps of, The ith sample output for the model belongs to the probability value of the river.
6. The method of claim 4, wherein the structural regularization term The expression of (2) is: Wherein, the method comprises the steps of, The number of leaf nodes of the mth decision tree; the method comprises the steps of determining the weight of the jth leaf node of an mth decision tree, wherein gamma is a penalty term used for limiting the number of the leaf nodes, and lambda is a regularization coefficient used for limiting the amplitude of the leaf node weight.
7. The method of claim 1, wherein the cross-scale maximum value fusion of the multi-scale river probability map in step S3 is performed to obtain an expression of a comprehensive river probability map as follows: Wherein, the method comprises the steps of, Representing river attribution probability values corresponding to the pels (x, y) under the scale sigma, The comprehensive river probability value representing the pixel (x, y) is obtained by retaining the strongest river response characteristics of the pixel at each scale.
8. The method according to claim 1, wherein the direction-constrained region growing algorithm growing criteria and algorithm iteration method of step S4 comprises: for any current seed point q and its neighborhood point p, the growth criteria are defined as: Wherein, the method comprises the steps of, A growth judgment value of the neighborhood point p relative to the current seed point q; Is a threshold value that regulates the area growth range; For measuring the direction consistency, the value range is [0,1], the value is the cosine value of the included angle of the remarkable linear direction vector of the point p and the point q, and the remarkable linear direction vector is represented by the characteristic vector corresponding to the maximum characteristic value of the corresponding point Hessian matrix; the iterative mode of the algorithm is that newly-incorporated river pixels are used as seed points for the next round of growth, the growth process is executed iteratively until no new pixels meet the growth criterion, and a completely-connected river mask result is output.
9. The method of claim 1, wherein the spatially adaptive stream combustion algorithm of step S5 has a depth of combustion Obtained by a dynamic calculation model, the expression is: Wherein, the method comprises the steps of, The reference combustion depth is determined by the resolution of the digital elevation model and the regional topography characteristic; A is the flow accumulated value of the current pixel (i, j); Is a dynamic threshold; the channel width corresponding to the current pixel (i, j); The method is the largest linear geometric feature under multiple scales and is used for enhancing the terrain correction effect of small tributaries; And And the balance factor is used for coordinating the influence weight of the river channel width and the linear geometric characteristics on the combustion depth.
10. The method of claim 1, wherein the obtaining of the hydrologic conditional DEM in step S5 is performed by correcting an original digital elevation model based on the combustion depth, where a correction formula is: Wherein, the method comprises the steps of, The elevation value of the pixel (i, j) in the original digital elevation model; the dynamic combustion depth corresponding to the pixel (i, j); And (3) the corrected hydrologic conditional DEM is used for subsequently extracting the river network with the accurate spatial position and topological communication relation.

Description

River network extraction method based on river probability learning and self-adaptive stream combustion Technical Field The invention belongs to the technical field of intelligent processing of remote sensing information, and particularly relates to a river network extraction method based on river probability learning and self-adaptive stream combustion. Background Land river networks are important fundamental elements of hydrologic cycle, ecosystem maintenance and water resource management. With the rapid development of remote sensing and geographic information technology, the extraction of river networks by using optical remote sensing images and Digital Elevation Models (DEMs) has become a core means in the fields of river basin analysis, flood simulation, environmental monitoring and the like. At present, the river network data requirements of high precision, consistent topology and being capable of adapting to complex terrains and ground object environments in the industry are urgent increasingly, and the remote sensing data processing technology and the hydrologic model are pushed to continuously evolve towards multi-source fusion, intellectualization and automation directions. The existing river network extraction technology mainly surrounds two types of data sources, namely a method based on optical remote sensing images, wherein water body information is enhanced through spectrum indexes (such as NDWI and MNDWI), or geometrical characteristics such as edge detection and region growth are combined to extract linear water body contours, a machine learning model (such as a random forest and a multi-layer perceptron) is further introduced to fuse spectrum and texture characteristics to realize pixel-level water body classification, and a method based on DEM topographic analysis, mainly adopts a confluence simulation method such as a D8 single-flow algorithm, calculates water flow direction and accumulation according to ground surface elevation, and automatically generates a river network skeleton with hydrologic connectivity. In order to combine the advantages of the two, subsequent researches develop a multi-source fusion method, such as a remote sensing stream combustion (RSSB) technology, and the spectrum index result is used as weight information to be burnt into a DEM so as to correct the terrain and generate a river network with position accuracy and topology consistency. Although the existing method has advanced to a certain extent, the method has obvious limitations that most fusion methods depend on single spectrum indexes at the characteristic level, and cannot fully combine geometric structure information such as river morphology, scale and the like, so that the error extraction and fracture are easy to occur in shadow coverage, urban interference or complicated morphology (such as a braided roadway) areas, and secondly, the terrain correction link usually adopts fixed combustion depth or experience parameters, lacks self-adaptive adjustment capability for local river width, confluence intensity and other landform hydrologic characteristics, and is easy to generate position deviation or pseudo river phenomenon in flat areas and micro-landform areas, and the technical flow from characteristic extraction to terrain correction is still in a segmented and fractured state and does not form a synergetically optimized integrated intelligent processing frame. Disclosure of Invention The invention aims to provide a river network extraction method based on river probability learning and self-adaptive stream combustion, which solves the problems of weak characteristic distinguishing capability, terrain correction rigidification and flow splitting of the existing river network extraction method by integrating a multi-characteristic collaborative expression and self-adaptive terrain correction integrated technology, thereby realizing intelligent extraction of a river network with high precision and consistent topology under a complex environment. The river network extraction method comprises the following steps: S1, acquiring remote sensing optical image and digital elevation model data of a target area, sequentially carrying out radiation correction, atmosphere correction and cloud shadow mask processing on the remote sensing optical image, resampling the processed remote sensing optical image and digital elevation model data to uniform spatial resolution, and obtaining a standardized input data pair; S2, calculating various water body spectral indexes of each pixel in the image based on the remote sensing optical image in the standardized input data pair, extracting linear geometric features of each pixel based on a multi-scale Hessian matrix, and fusing a spectral index set formed by the various water body spectral indexes with the linear geometric features to obtain a high-dimensional feature vector representing the spectral properties and the spatial morphology of the pixel; s3, training a river-ori