SOMDNCD: Image Change Detection Based on Self-Organizing Maps and Deep Neural Networks

Image change detection is a research hotspot in many fields of application, such as environmental monitoring, disaster investigation, urban research, and more. How to reduce the influence of speckle noise when conducting change detection in an acquired synthetic aperture radar (SAR) image is a challenging issue. This research shows that reasonably balancing noise suppression with the preservation of the edges of regions is the key to generating a good change map. Therefore, a new image detection method based on a selforganizing map and deep neural network (SOMDNCD) is proposed. First, the method uses a median filter to improve the difference image that is generated by the mean-ratio operator, which reduces the influence of the image point noise on generating difference maps. Compared with the difference map formed by the logarithmic ratio operator, the edge information in the image is excellently retained and the missed detection rate is reduced; second, the network preprocesses the difference map, obtains a preliminary change map, and divides the pixels of the difference map into three types: no change, noise, and change. Finally, a deep neural network is used to train a noise-like training set on the network to reduce the residual noise in the change class and obtain the final change graph. The experimental results show that compared with other current mainstream methods, the proposed SOMDNCD change detection method directly addresses noise and is universal for a variety of data sets. The proposed method exhibits a lower missed detection rate in the SAR image data set and a more ideal false alarm rate than other methods.