Shallow Convolution Neural Network for Real-Time Change Detection in Agricultural Lands

The remote sensing technology is helpful for collecting information regarding usage of multispectral data from backscattered coefficients and optical sensors through sensors. The main challenge for an unsupervised change detection is distinguishing among a potential change to significant changes from noise. The Shallow Convolution Neural Network (SCNN) is proposed for change detection in agriculture lands. The Onera Satellite Change Detection (OSCD) dataset is used and it is preprocessed by Non-Local Means (NLM) filters which is helpful to denoise hyperspectral or multispectral images. Through preserving land structures, it enables accurate detection of changes in agricultural areas. Then, the CNN is used for extracting features in which primary layer captures low-level features such as textures and edges while deep layers recognize much complex structures and patterns. The SCNN mitigates overfitting issues particularly when training which leads in satellite-based monitoring thereby providing effective change detection. The Overall Accuracy (OA), Kappa, IoU and mIoU are taken to estimate SCNN performance. The SCNN obtains 95.19%, 0.896, 0.842 and 0.875 of OA, kappa, IoU and mIoU for OSCD dataset which is better than Deep Neural Network (DNN).