Cloud Detection for PERUSAT-1 Imagery Using Spectral and Texture Descriptors, ANN, and Panchromatic Fusion

Abstract

The cloud detection process is a prerequisite for many remote sensing applications in order to use only those cloud-free parts of satellite images and reduce errors of further automatic detection algorithms. In this paper, we present a method to detect clouds in high-resolution images of 2.8 m per pixel approximately. The process is performed over those pixels that exceed a defined threshold of blue normalized difference vegetation index to reduce the execution time. From each pixel, a set of texture descriptors and reflectance descriptors are processed in an Artificial Neural Network. The texture descriptors are extracted using the Gray-Level Co-occurrence Matrix. Each detection result passes through a false-positive discard procedure on the blue component of the panchromatic fusion based on image processing techniques such as Region growing, Hough transform, among others. The results show a minimum Kappa coefficient of 0.80 and an average of 0.94 over a set of 25 images from the Peruvian satellite PERUSAT-1, operational since December 2016.