Search

CN-121982589-A - Edge-guided high-resolution remote sensing image complex scene water body extraction method

CN121982589ACN 121982589 ACN121982589 ACN 121982589ACN-121982589-A

Abstract

The invention belongs to the technical field of remote sensing geography application, and particularly relates to an edge-guided high-resolution remote sensing image complex scene water body extraction method, which comprises the following steps of S1, inputting high-resolution unmanned aerial vehicle images; S2, extracting multi-scale features of a CNN-converter mixed architecture, S3, enhancing three paths of features of an edge guiding reasoning module, S4, carrying out pixel-level self-adaptive fusion of a pixel sharing feature fusion module, and S5, outputting a complex scene water body segmentation result. The invention effectively solves the difficult problems of extracting the water bodies of the complex scenes such as the semi-dry water bodies, vegetation covered water bodies, narrow river channels, shadow water bodies, turbid muddy water and the like in the high-resolution unmanned aerial vehicle image, remarkably improves the boundary positioning precision and the characteristic recognition capability, and provides an effective technical solution for the fields of water resource monitoring, flood disaster assessment and the like.

Inventors

  • WANG BIAO
  • WU YANLAN
  • YANG HUI
  • XU SHENG
  • QIN JUN

Assignees

  • 安徽大学
  • 安徽省生态环境监测中心(安徽省重污染天气预报预警中心、安徽省机动车排气污染监控中心)

Dates

Publication Date
20260505
Application Date
20260129

Claims (8)

  1. 1. The edge-guided high-resolution remote sensing image complex scene water extraction method is characterized by comprising the following steps of: s1, inputting high-resolution unmanned aerial vehicle images; S2, extracting multi-scale features of a CNN-converter mixed architecture; S3, enhancing the three path characteristics of the edge guiding reasoning module; s4, pixel-level self-adaptive fusion of a pixel sharing characteristic fusion module; s5, outputting a complex scene water body segmentation result.
  2. 2. The edge-guided high-resolution remote sensing image complex scene water extraction method is characterized by comprising the steps of S1, obtaining an RGB image of a high-resolution unmanned aerial vehicle with a spatial resolution of centimeter level, carrying out fine manual labeling on the image to generate a pixel-level water label, and randomly dividing a cut image-label pair into a training set, a verification set and a test set according to a ratio of 8:1:1.
  3. 3. The edge-guided high-resolution remote sensing image complex scene water extraction method of claim 2, wherein the complex scene water types covered by the high-resolution unmanned aerial vehicle RGB image include half-dry, vegetation cover, narrow river, shadow interference and muddy water.
  4. 4. The method for extracting the complex scene water from the edge-guided high-resolution remote sensing image according to claim 3, wherein after the pixel-level water label is generated, the water class is assigned to 1, the background class is assigned to 0, the whole image is subjected to sliding window cutting according to 1024×1024 pixels, the overlapping degree is set to 30% so as to ensure the integrity of boundary area information, and image blocks with more than 80% of areas being invalid information are removed.
  5. 5. The method for extracting the complex scene water from the edge-guided high-resolution remote sensing image is characterized in that in the step S2, the encoder performs multi-scale feature extraction on the input image by adopting a lightweight ResNet-18, four feature maps { F 0 ,F 1 ,F 2 ,F 3 } with different resolutions are generated through four downsampling stages, the number of feature channels is 64, 128, 256 and 512 respectively, the decoder adopts a transducer decoder to restore the spatial resolution step by step, the number of heads is set to be 8, the feedforward network expansion factor is 4, global context information is modeled through a self-attention mechanism, jump connection is set at each level of the decoder, and the feature maps of the corresponding level of the encoder are transmitted to the decoder.
  6. 6. The method for extracting the complex scene water body of the edge-guided high-resolution remote sensing image is characterized in that in the step S3, an edge-guided reasoning module receives a jump connection feature F_skip and a prediction result P_pred, an edge feature F_edge is extracted by using a Laplacian operator, a low-confidence potential water body area is identified through three reasoning path enhancement features, namely a background exclusion path passes through (1-Sigmoid) and a boundary driving path refines boundary positioning by using Laplacian (P_pred), the F_skip and the F_edge are multiplied by a high-frequency feature path to strengthen edge response, and three path outputs are integrated through 3×3 convolution after channel dimension splicing, and feature enhancement is completed through spatial attention optimization and residual connection combining CBAM mechanisms.
  7. 7. The edge-guided high-resolution remote sensing image complex scene water extraction method is characterized in that in the step S4, a pixel sharing feature fusion module respectively processes deep semantic features F_deep of a decoder and jump connection shallow geometric features F_skip by adopting a symmetrical double-branch architecture, both branches adopt grouping convolution to perform feature transformation, the grouping number is set to be 8, preliminary fusion is achieved through element-by-element addition to obtain F_fused, a cascading double-attention mechanism is introduced, namely spatial attention passes through Concat (AvgPool (F_fused) and MaxPool (F_fused), space weight is generated through 7×7 convolution, channel attention adopts an adaptive convolution kernel strategy to generate channel weight, the convolution kernel size k is determined according to channel dimension self-adaption, and finally feature enhancement is achieved through pixel-by-pixel element multiplication and residual connection.
  8. 8. The edge-guided high-resolution remote sensing image complex scene water extraction method is characterized in that in step S5, an image to be extracted is input into a trained network model, after multi-scale features are extracted from the image through an encoder, feature expression and self-adaptive fusion are enhanced through the cooperation of EWFI modules and PSFF modules at each level of a decoder, a binary water segmentation result is finally output by the decoder, a large-format image is cut by adopting a sliding window strategy, 30% overlapping degree is set, and a complete water extraction result is generated through splicing after prediction.

Description

Edge-guided high-resolution remote sensing image complex scene water body extraction method Technical Field The invention belongs to the technical field of remote sensing geography application, and particularly relates to an edge-guided high-resolution remote sensing image complex scene water body extraction method. Background The surface water body is an important component part of the earth ecological system, and has important significance for ecological environment stabilization and sustainable development of human society. The distribution of the surface water body and the space-time variation thereof are accurately identified and monitored, and the method plays a key role in the fields of flood disaster monitoring, water resource management, environmental protection and the like. The remote sensing technology is an important means for water body identification and monitoring because the remote sensing technology can acquire the surface information of a large area without being limited by ground conditions. Traditional water extraction methods mainly comprise a technology based on a spectrum index and a technology based on image classification. The spectral index method is used for extracting the water body by calculating different wave bands in the remote sensing image and combining threshold segmentation, such as normalized difference water body index (NDWI), improved normalized difference water body index (MNDWI) and the like. The method has higher calculation efficiency and better effect on the medium-low resolution satellite images. However, in high-resolution remote sensing images, the traditional spectrum index method faces significant challenges that on one hand, the traditional spectrum index method depends on a few key wave bands and cannot fully utilize space detail information of the high-resolution images, and on the other hand, in a complex scene, the spectrum characteristics of water bodies and non-water bodies (such as building shadows, wet soil, dark watertight surfaces and the like) are highly similar, so that the water body extraction precision is insufficient. The method based on image classification realizes the distinction between water and non-water through a machine learning or deep learning model, and can obtain higher accuracy in a larger range, but has the limitations of dependence on a large amount of manual annotation data, high computational complexity, low processing speed and the like, and is not suitable for real-time monitoring and emergency response scenes. With the development of deep learning technology, a semantic segmentation model based on a Convolutional Neural Network (CNN) gradually becomes a mainstream method for water extraction. The FCN introduces the convolutional neural network into the semantic segmentation field for the first time, so that pixel-level dense prediction is realized, the U-Net effectively fuses different layers of features through an encoder-decoder architecture and a jump connection mechanism, deepLab series are introduced into a cavity convolutional and cavity space pyramid pooling module, the receptive field is enlarged, and the multi-scale feature extraction capability is enhanced. The CNN-based method has remarkable progress in the remote sensing image water body extraction, but has the limitation that the receptive field is limited due to the locality of convolution operation, the long-range dependence of a water body area is difficult to establish, and the crushing and segmentation are easy to generate when the water body with complex boundaries is processed. In recent years, the transducer architecture brings new breakthrough for semantic segmentation by virtue of the global modeling capability of a self-attention mechanism, but the secondary computation complexity of the transducer architecture leads to the rapid increase of memory consumption when processing high-resolution images, and the subtle characterization capability of local boundary details is insufficient. The rapid development of unmanned aerial vehicle remote sensing technology brings new opportunities for acquiring water body information. Compared with satellite remote sensing, unmanned aerial vehicle remote sensing has the remarkable advantages of high spatial resolution (reaching centimeter level), flexible flight, relatively low cost and the like, and is particularly suitable for fine monitoring and emergency response in a small-range area. However, ultra-high spatial resolution brings rich surface detail information and also brings new technical challenges to water extraction. Under the centimeter-level resolution, the complex scene water body extraction faces two major core problems, namely firstly, the problem of boundary complexity, namely the irregular geometric form of the water body boundary is amplified by the ultrahigh resolution, the boundary positioning precision is insufficient due to the complex conditions of alternating water and land, vegetation shie