Remote sensing imagery shadow detection via physical constraint

Shadow detection is a challenging task for the community, which is crucial for interpreting remote sensing imagery. Despite promising accomplishments from existing works, sparse and suspicious shadow distributions, imperceptible and irregular shadow contours challenge them. To this end, we propose a multi-task decoupled learning paradigm (MDNet) to explore the collaboration between shadow region prediction and shadow boundary revision for a win-win situation. Specifically, we design a quadruple prior operator inspired by optical imagery formation models. As a feature selector, the quadruple prior is able to reflect the contamination degree of shadows, thus enhancing the response of quality-degraded regions. Subsequently, we propose a contour iteration scheme with flexible steps. We incorporate corner point and boundary priors to cue movement direction and step size of junctions, thus making dynamic shadow boundary correction possible. Extensive experiments on multiple representative remote sensing imagery shadow detection benchmarks demonstrate the superiority of MDNet in terms of qualitative and quantitative metrics.