Change Detection In Remote Sensing Images Based On Deep Learning

Since the 1960s,the level of remote sensing technology has been continuously developed and improved.Remote sensing images have been widely used in various fields of life.Therefore,a large amount of remote sensing image data has emerged,and the data information in the images has become more abundant,making the targets in remote sensing images Information detection has attracted the attention of many researchers,and change detection is an important research direction in remote sensing image target detection technology.Remote sensing change detection technology is the process of identifying and analyzing the relevant feature change information in the target area in the multi-temporal remote sensing image.Its essence is to obtain the difference information of the feature in the same area at different times in the image to achieve dynamic surface monitoring.How to accurate and effective display of change detection information,providing important information for analyzing and predicting what will happen in this area has become a research hotspot.This paper proposes a compound algorithm that combines deep learning and slow feature analysis and improves optimization.Relevant experimental analysis is carried out on several single traditional remote sensing change detection algorithms,including image difference method,image ratio method,principal component analysis method,change vector analysis method,multiple change detection method and iterative weighted multiple change detection method.The feature extraction accuracy shown is not high and the changing pixel information is difficult to determine,resulting in poor detection results and difficulty in meeting the current detection needs.Therefore,an algorithm combining deep learning and slow feature analysis is proposed,using deep learning algorithms It has strong learning ability,fitting complex mapping relationship,extracting feature information and other advantages;slow feature analysis has good binary classification characteristics in suppressing unchanged areas and highlighting changed areas.To give full play to the unique advantages between the two algorithms,the similarity between the result map detected by the experiment and the change reference map is closer,and the detailed information of the image is fully retained.From the subjective vision and objective evaluation,it is greatly improved compared to the single traditional algorithm.Optimized and improved the iterative function of image data.In the deep learning algorithm,in view of the long training time of the algorithm,this paper finds the relationship between the regularization parameter value and the accuracy of the detection result according to the experimental data,and finds that the regularization parameter exceeds a certain value.In the steady state,when the regularization parameter is this value,the error value corresponding to the output of the error function is set as the threshold,and the condition for determining whether to continue the iteration is avoided to a certain extent.The situation that the detection effect is not significantly improved due to multiple iterations is avoided.The amount of calculation is optimized,and the situation of under-fitting due to insufficient iterations is eliminated,and the efficiency and accuracy of change detection are improved.