Burn scar detection using cloudy MODIS images via low-rank and sparsity-based models

Abstract

In this paper, we propose a general framework to detect burned area using multiple partially occluded Moderate Resolution Imaging Spectroradiometer (MODIS) images. By treating each MODIS image as the superimposition of 3 layers: background, burned area, and cloud, we first apply a low-rank and sparse matrix decomposition technique known as Robust Principal Component Analysis (RPCA) to separate cloud from the other two components. Secondly, a joint sparsity-based representation for classification algorithm is employed to determine the burned area. We further robustify the algorithm by proposing a novel combined model that can separate and suppress the cloud information while simultaneously detecting the burn scar. Preliminary results using actual MODIS images clearly demonstrated the efficacy of the proposed technique. The probability of correct detection of burned area can reach more than 80% with less than 10% false alarm rate, even in an extremely cloudy day.