Domain Adaptive based Convolutional Neural Network for Change Detection using Time Series Satellite Imagery

The Change Detection (CD) is the process of observing the transition of the earth’s surface which is crucial in urban areas. Existing methods showed various developments in CD using Deep Learning (DL) models, however, they are limited to learn the image features only in one domain. To overcome this, the proposed research developed a Domain Adaptive (DA) based Convolutional Neural Network (CNN) for CD (DA-CNN-CDNet). The time series satellite images from the Sentinel-2 are collected and pre-processed by applying pixel-level and spatial-level transforms for effective training of the model. The features are extracted using Pseudo-Siamese with non-weights and fed to the CNN. The CNN is trained in DA environment to apply the knowledge that CNN has learnt in one domain to one or more other domains which enhances the CD. The experimental results of the proposed DA-CNN-CDNet showed better CD with 99.76% of accuracy, 99.18% of precision, 98.62% of recall, and 97.98% of f1-score when compared to existing methods Cross Layer Convolutional Neural Network (CLNet) and Local Similarity Siamese Network (LSS-Net).