Development of a Fully Automated Atmospheric Correction Technique for Applications in Google Earth Engine

Abstract

For any satellite-based study, ground-based reflectance values are required. Nowadays, these ground-based reflectance products are provided by all the major space agencies, commonly designated as level 2 products. However, the availability of the level 2 products takes some time and once in a while, these products are not available. For the development of near real-time monitoring systems, this poses a major problem, and thus it becomes necessary to correct the raw satellite imagery by using atmospheric correction techniques. The Dark Object Subtraction technique (DOS) is one such commonly used technique used previously in many studies. However, it requires the manual selection of the darkest pixels in the imagery, thus making it unsuitable for automation-based systems. This study aims to automate the process of Dark Object Subtraction Sentinel 2A raw satellite imageries within the Google Earth Engine platform. Mean annual LULC maps generated using automated Dark Object Subtraction could replicate the level 2 product quite accurately. These classified imageries for July 2018−July 2019 produced overall classification accuracies of 74.13 and 67.24% using Random Forest Classifier and Support Vector Machines, respectively, compared to 68.96% obtained for both the classification algorithms using level 2 products. In the period July 2019−July 2020, it was obtained as 81.03 and 77.58%, respectively, compared to 79.31% for the same, and for July 2020- July 2021, it was 72.41 and 68.96% against 68.96 and 67.24%. The automated Dark Object Subtraction technique can thus be employed to develop near real-time automated satellite imagery-based systems within the Google Earth Engine platform.