A Deep Learning Approach for Imagery Masking of Spectral Sensors

Currently, some of the implemented atmospheric correction processors for remote sensing spectral sensors, use masking algorithms based on thresholding of spectral indices with sensor Top-Of-Atmosphere (TOA) reflectance. This concept allows the use of a limited amount of spectral bands, which is optimal for multi-spectral sensors (~10-20 bands), but for the case of the high spectral dimensionality of hyper-spectral sensors, spectral thresholding underutilizes the number of available bands. Given this limitation, we propose a masking algorithm which performs spatial and spectral feature extraction based on a 2D convolutional neural network, fitting the model with the available classification maps from the Python-based Atmospheric COrrection (PACO) processor. The training samples are selected based on their uncertainty to belong to a given class, The validation is performed using two independent human expert labelled datasets. The resulting classification maps show an improvement from the original ones of PACO.