CN-116824373-B - Deep learning-based large-scale rapid remote sensing water body extraction method and system

CN116824373BCN 116824373 BCN116824373 BCN 116824373BCN-116824373-B

Abstract

The invention belongs to the technical field of information technology service, and discloses a large-scale rapid remote sensing water body extraction method and system based on deep learning. The invention has the high efficiency of traditional machine learning and the deep feature extraction of deep learning, and has great innovation in data preparation, network frame optimization and image clustering post-processing. The method comprises the steps of firstly preliminarily determining a pixel clustering range by utilizing traditional machine learning, then further clustering pixels in the range by utilizing a convolutional neural network layer, and finally obtaining a water body extraction result through binarization after clustering. The invention has strong applicability, can be used for various actual requirements including flood drawing, surface water extraction, shoreline change monitoring and the like, and can meet the requirements of rapid, large-scale and high-precision water body extraction.

Inventors

LI PENG
ZHU QUANTAO
LI ZHENHONG
WANG HOUJIE

Assignees

中国海洋大学

Dates

Publication Date: 20260512
Application Date: 20230627

Claims (9)

1. A large-scale rapid remote sensing water body extraction method based on deep learning is characterized by comprising the following steps: Dividing the preprocessed image into training images capable of meeting computer power requirements, defining a clustering range of a deep learning frame around a dividing threshold according to a minimum error dividing method, and clustering pixels in the range by using the deep learning frame, wherein a middle region range corresponding to the dividing threshold is a range between a high-confidence water body and a non-water body, the high-confidence water body is a pixel region with extremely low reflectivity in the SAR image, and the high-confidence non-water body is a pixel region with extremely high reflectivity in the SAR image; when training the training images with extremely uneven water and non-water duty ratio, the two super parameters are utilized to prevent the pixel blocks of the water or non-water from being lost along with the continuous increase of training times; the two super parameters are set based on pixel range [0,255] of Uint8 type data, and specific limiting conditions are defined by the formula: ; Wherein, the And Respectively the first And The pixel value of the super pixel block output by the secondary training is 90,160, and the super parameters are 90,160 respectively; And binarizing the clustered images by using pixel characteristics of the original training images to finally generate a water-non-water classification chart.
2. A large-scale quick remote sensing water body extraction system based on deep learning is characterized by comprising: The system comprises a data preprocessing module, a processing module, a data clipping module, a data processing module and a data processing module, wherein the data preprocessing module is used for inputting original data and then carrying out track correction and thermal noise removal; The method comprises the steps of generating a convolution network clustering module, generating a segmentation label after a network original image is segmented by super pixels, outputting a feature map after the original image is subjected to three-layer convolution layer operation, generating a clustering feature map after the feature map is further subjected to the steps of original image fusion, segmentation label clustering and two super-parameter constraint, generating a loss function by utilizing the clustering feature map and the feature map, and continuously optimizing parameters of each node of the network through backward propagation, wherein the two super-parameter constraint means that an area with an excessively low pixel value or an area with an excessively high pixel value is limited by the super-parameter and disappears in the clustering process, and the specific limiting condition is as follows: , Wherein, the And Respectively the first And The pixel value of the super pixel block output by the secondary training is 90,160, and the super parameters are 90,160 respectively; And the post-clustering binarization module combines the pixel intensity information in the original image with the clustering result output by the network to obtain a clustering result diagram which can best embody the characteristics of the original image.
3. The deep learning-based large-scale rapid remote sensing water extraction system of claim 2, wherein the pre-processing is followed by converting the intensity data into dB data, the formula is: ; wherein, I is the pixel intensity after pretreatment, and dB is the decibel unit data generated by conversion.
4. The deep learning-based large-scale rapid remote sensing water extraction system according to claim 2, wherein the image unit after preprocessing by the preprocessing module is dB, the data type is 32-bit floating point type, the data type is converted into unsigned 8-bit integer type before the convolutional neural network clusters the images, 1% linear stretching is performed on all training images before the data type conversion, and the stretching range of all images is the same.
5. The deep learning-based large-scale rapid remote sensing water extraction system as claimed in claim 2, wherein the early training needs to preserve low-dimensional features by fusing input images, some small-area water in the images slowly disappear as the low-dimensional features as the number of network training increases, so that the final clustering precision is reduced, the feature images output by the original images and the convolution layer are fused to obtain new feature images containing feature information and the original image information, the fusion method is to carry out weighted summation on the original images and the feature images according to a certain proportion, the weight of the original images is gradually reduced as the number of network training increases, the weight of the feature images is gradually increased, and an image fusion formula is as follows: , Wherein the method comprises the steps of To fuse the pixel values of the "new feature image" after the original image, The number of training times is indicated and, Representing the corresponding pixel value of the input image, And the corresponding pixel value of the output characteristic image of the convolution layer is represented.
6. The deep learning-based large-scale rapid remote sensing water extraction system of claim 2, wherein the new feature map generated after fusion also requires segmentation label processing.
7. A computer device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the deep learning based wide range rapid remote sensing water extraction system of claim 2.
8. A computer readable storage medium for storing a computer program which, when executed by a processor, causes the processor to perform the steps of the deep learning based wide range rapid remote sensing water extraction system of claim 2.
9. An information data processing terminal, characterized in that the information data processing terminal is used for realizing the deep learning-based large-scale rapid remote sensing water body extraction system.

Description

Deep learning-based large-scale rapid remote sensing water body extraction method and system Technical Field The invention belongs to the technical field of information technology service, and particularly relates to a large-scale rapid remote sensing water body extraction method based on deep learning. Background In the large background of global sea level rise and dramatic coastline changes, coastal zone populations are subject to increasingly severe flood disasters, nearly 23% of the world population being directly exposed to centuries. Thus, rapid and wide-scale monitoring of floods is a piece and its important work, which concerns human property and life safety. Surface water extraction problems, including flood extraction, have evolved in recent years. The original surface water drawing method can obtain the surface water body range very accurately through manual measurement, but has very low efficiency, consumes a large amount of manpower and material resources and has low applicability. The water body extraction method based on remote sensing and traditional machine learning can be divided into supervised classification and unsupervised classification, wherein the unsupervised classification method comprises a plurality of common threshold classification methods, such as methods of maximum inter-class variance, minimum error segmentation and the like, the method has high segmentation efficiency, can rapidly realize large-scale water body extraction by utilizing SAR images, but the water body extraction method relying on a single threshold has larger defects at the same time, has strong and changeable SAR backscattering for areas with complex types of ground objects, and has poorer classification results obtained by utilizing the single threshold in classification details. The supervision classification method comprises a support vector machine, a random forest and the like, and compared with the traditional threshold segmentation method, the precision of the supervision classification method is always improved to a certain extent, but the water extraction efficiency is lower, and the training samples are required to be selected by consuming manpower. The traditional machine learning has very limited capability in dealing with the problem of SAR image speckle noise, which is typically attenuated by filtering methods. In recent years, deep learning has been rapidly developed, and is widely applied to various fields including natural language recognition, target recognition, natural image classification, and the like. The deep learning can extract deep features in the images, and has great advantages compared with the traditional machine learning, so that many researches are applied to classification tasks of multispectral and hyperspectral images, and the classification applications to SAR images are relatively few. The good classification result of the supervised deep learning often depends on massive training data, and training labels are required to be manufactured by consuming human and material resources, so that the method has great disadvantages for pursuing the problem of water extraction such as time-efficient flood drawing. The deep learning method aiming at unsupervised property has great potential in rapid and large-scale flood monitoring, and is an important research direction of flood extraction. Through the analysis, the prior art has the problems and defects that the original surface water drawing method is very low in efficiency, needs to consume a large amount of manpower and material resources and is low in applicability, the water body extraction method based on remote sensing and traditional machine learning is high in segmentation efficiency, but is poor in classification details for areas with complex ground object types, the supervision classification method is improved to a certain extent in accuracy compared with the traditional threshold segmentation method, the water body extraction efficiency is low, needs to consume manpower to select training samples, and the traditional machine learning is very limited in capability of processing SAR image speckle noise, and generally weakens speckle noise through a filtering method. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a large-scale rapid remote sensing water body extraction method based on deep learning. The invention is realized in such a way that a large-scale rapid remote sensing water body extraction method based on deep learning comprises the following steps: the preprocessed image is segmented into training images which can meet the computational power requirements of the computer. In the early stage of convolution clustering training, the pixel characteristics of the original training image are fused, and meanwhile, the important super-pixel block information in the clustering characteristic diagram is reserved as far as possible by combining with the super-pixel segme