Change detection over the Aral Sea using relative radiometric normalization based on deep learning

The desertification of the Aral Sea causes various environmental destruction and local community collapse. In order to prepare countermeasures, changed areas caused by desertification should be quickly and accurately detected. However, if the radiometric dissimilarity between bi-temporal satellite images is severe, the probability of false detection increases. Therefore, a relative radiometric normalization (RRN) approach based on deep learning is proposed to accurately detect the changed areas.To this end, a deep learning network is designed to extract pseudo-invariant features (PIFs), which is invariant pixels with similar spectral characteristics. More specifically, training dataset generated based on the center points of objects defined by applying an image segmentation are inputted to the network. After training the deep learning network, the PIFs are extracted by measuring the similarity between deep features. The radiometric dissimilarity is non-linearly normalized by estimating an artificial neural network based on the extracted PIFs. Then, changed areas by desertification are detected in object units by combining the pixel-based change map and the segmented objects. Bi-temporal Landsat-8 images acquired from the Aral Sea in 2013 and 2021 were used as experimental images. The proposed method showed sufficient performance for detecting the overall change in land cover due to desertification.