Research On SAR Water Information Extraction Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)has all-weather,all-day imaging capabilities and has significant advantages in disaster detection.In recent years,frequent flooding is one of the most destructive natural disasters,so flood detection has received widespread attention.The extraction of SAR image water region information provides the water region part and the change of water region part in SAR image,which is the key link in SAR image flood detection.Convolutional Neural Network(CNN)can extract deep water information features which are robust and discriminative.Therefore,the research on SAR image water region information extraction algorithm based on convolutional neural network is the frontier development direction of SAR image flood detection.Focusing on the topic of SAR image water region information extraction,this paper proposes a SAR water region information extraction algorithm based on convolutional neural network.Moreover,a SAR image flood detection system for the Huaihe River Basin was researched and developed.The main content of the article is as follows:(1)Aiming at the problem of insufficient training samples in the extraction of water information of SAR images based on CNN,a SAR image water information extraction algorithm based on Multi-depth Convolutional Neural Network(MD-CNN)is proposed.The algorithm first builds the MD-CNN model through adaptive network depth selection based on the proportion of water pixels in the input dual-time SAR image,thereby reducing the algorithm's need for training samples.On this basis,saliency detection is used to further reduce the number of training samples.Finally,the algorithm uses MDCNN to separately extract the distribution of water region in the dual-time SAR images,and compares the differences in the distribution of water region in the two SAR images to achieve the detection of water region changes.Experimental results verify the effectiveness of the algorithm.(2)A SAR image flood detection system for the Huaihe River Basin was researched and developed.The system includes a water region extraction part and a water region change detection part.The water region extraction part is mainly used to extract the water region distribution in the input SAR image.This part converts the water region extraction problem into a binary classification problem and uses the deep learning SAR image classification method to obtain the water region distribution situation.The water region change detection part is used to detect Changes in water region.This part first extracts water region from the input pair of SAR images,and then uses the difference method or ratio method to obtain the water region changes;According to the actual test data of Lu'an part of the Huaihe River Basin in Anhui Province and the East Lake of Huoqiu County in 2017,the result show that the system can accurately extract water region,detect changes and interact well with users.