Deep Learning Technique for Cloud Detection using Satellite Data

Abstract

Cloud detection is a crucial task and has varied ranges of implications in retrieving important parameters using satellite data. Identifying clouds from clear sky hold great importance in many satellite Imagery applications. Many approaches are used for performing cloud detection on satellite data products. Some of the well-known approaches are a threshold-based approach, a machine learning approach, and a few others, but these approaches lack robustness as these approaches require a profuse amount of time in performing feature-selection. Most of the algorithms fail in taking advantage of spatial arrangement and are time intensive. In tasks like image recognition and object detection, deep learning has outperformed compared to other approaches. In this paper, a deep learning model was proposed for performing cloud detection using INSAT 3D satellite data product which overcomes all the above-mentioned limitations. The proposed model architecture consists of encoder and decoder modules, which will perform sampling, feature extraction, and up-sampling. The proposed model takes five features consisting of SWIR, VIS, TIR1, TIR2, and MIR spectral band’s/channel’s data as input and generates cloud mask as output. Generated cloud mask performs better distinction between cloudy and non-cloudy pixels under different surface conditions, mostly over ice and snow. The proposed model will generate a day-time cloud mask as SWIR and VIS spectral bands data are available only during the day-time.