Spatio-temporal analysis of land use/land cover change detection in small regions using self-supervised lightweight deep learning

Abstract

Change detection (CD) has sparked a lot of scientific interest in recent decades as one of the core concerns in Earth observation. The enhancement of the CD source data with the availability of multitemporal images with varying resolutions provides ample change indicators due to the rapid improvement of satellite sensors. However, precisely detecting real changed locations continues to be a complicated task. CD from remote sensing images (RSI) becomes challenging when the labeled data for supervised learning is unavailable. This article proposes a novel CD framework using a self-supervised learning (SSL) approach to overcome these limitations. First, the superpixel segmentation method of simple linear iterative clustering (SLIC) using a structural similarity index is incorporated to produce a difference image (DI). The change features are extracted to represent the difference information using spatial features between the corresponding superpixels. Second, a parallel clustering algorithm called fuzzy C-means (FCM) separates the DI into three clusters of changed, unchanged, and intermediate classes. The image patches of changed, unchanged and intermediate classes are constructed as training and testing samples. A lightweight deep convolutional neural network (LWDCNN) is trained with the training samples to detect the semantic difference and classify the testing samples into the changed or unchanged class. Finally, merging intermediate and change class labels generates a robust and high-contrast CD map. Numerical experiments were performed on two small regions like the Alappuzha, Kerala, India, and Paris building dataset to demonstrate the usefulness of the proposed approach, achieving an overall accuracy of 98.28% and 96.43% for determining changes effectively. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.