Change Detection For Remote Sensing Images Based On Neighboring Information And Deep Learning

Remote sensing image change detection refers to the process of detecting surface changes by analyzing several images acquired over the same region at different times using change detection algorithms.Traditional pixel-based methods use each pixel as a basic unit for analysis.When applied to medium-/high-resolution remote sensing images,they cannot take care of the neighboring information around each pixel.While the traditional object-based change detection algorithms can utilize the context information around the pixel,in this regard,they are more suitable for processing high-resolution remote sensing images.However,when performing change detection on segmented irregular targets,some traditional methods rely too much on hand-designed feature extraction operators and feature extraction is insufficient,which limits the accuracy of such methods.Deep neural networks have powerful non-linear feature extraction and mapping capability.They can transform raw input data into abstract feature space,enabling the algorithm to obtain a more accurate change detection result by comparing and analyzing the abstract features of the raw data.To solve the problems existing in traditional pixel-based and target-based change detection algorithms,this paper mainly focuses on change detection for SAR images based on fully connected conditional random fields and multi-spectral images based on superpixel segmentation and deep learning.In the pixel-based SAR image change detection algorithm,a post-processing method based on fully connected conditional random field is proposed to optimize the initial change detection results.Using the difference image and the initial binary map,a fully connected conditional random field is established with each pixel as a node.The Gibbs energy of the fully connected conditional random field is minimized to obtain the optimized labels of pixels to get a more accurate change map.For the object-based change detection algorithm,a change detection method based on superpixel segmentation and deep neural networks is proposed.In order to obtain superpixel pairs with consistent segment boundaries from two remote sensing images,we use a simple but effective strategy to directly segment the difference image.Then we can get superpixel pairs by segmenting two images with the superpixel boundary mask.Because each segmented superpixel has a different shape and size,the superpixel feature encoding network is designed to avoid manually feature engineering.It can directly take irregular superpixel pairs as the input vectors and extract abstract features from them.Finally,these abstract features are used for final change detection,which alleviates the bottleneck of manually feature engineering and improves the accuracy of change detection results simultaneously.Comparing with several classic change detection algorithms,qualitative and quantitative experimental results on multiple real SAR and multi-spectral datasets have proved the effectiveness of the proposed algorithms.They can solve the change detection problem of high-resolution remote sensing images more efficiently.