Land Change Detection in Sentinel-2 Images Using IR-MAD And Deep Neural Network

This paper presents a method for land change detection in bi-temporal satellite images using deep neural network (DNN) which is seen as the simplest deep learning (DL) architecture. Unlike the conventional change detection methods, the DL-based approaches do not require a large amount of existing knowledge about the study area. Consequently, no human intervention is necessary during the training step leading to the improvement of the change detection automation process. The proposed method uses a DNN algorithm whose inputs are binary training data (change, no change) and the iteratively reweighted multivariate alteration detection (IR-MAD) components. The IR-MAD channels calculated from the bi-temporal satellite images already contain change information, which might guide the DNN to automatically extract more robust features and then to produce more accurate change detection mapping. The performance of the implemented algorithm is verified by using co-registered bi-temporal multispectral images acquired in 2017 and 2022 by Sentinel-2 satellite over the North Eastern part of Algiers city (Algeria). The obtained results are promising and show that the proposed DNN change detection method is more accurate compared to usual conventional methods in terms of statistical precision (Overall accuracy, Kappa coefficient, spectral signature).