HIGH-DIMENSIONAL MULTIRESOLUTION SATELLITE IMAGE CLASSIFICATION: AN APPROACH BLENDING THE ADVANTAGES OF CONVEX OPTIMIZATION AND DEEP LEARNING

To protect valuable mangrove ecosystems, efficient and accurate mangrove area mapping becomes essential, for which high-dimensional multiresolution satellite image classification is the critical technique. The previous index-based methods only consider spectral information, and perform classification pixel-by-pixel ignoring the spatial continuity nature of the mangrove distribution. We introduce convex optimization (CO) into deep learning (DL) to achieve outstanding classification performance, without relying on big data or math-heavy regularization. Based on a rough mangrove multispectral signature estimated by mangrove vegetation index (MVI), but ruling out its key disadvantage of pixel-independent estimation in MVI via DL, our method introduces a deep regularizer employing pixel-dependence into a CO framework. The proposed classification method, termed MSMCA, is applied to mangrove mapping, showing state-of-the-art classification performance.