Back-propagation in the neural network of visibility estimation model based on Himawari_8 Satellite during forest fire smoke periods on Sumatera and Borneo Island, Indonesia

Abstract

Abstract Smoke from forest and land fires may significantly impair horizontal visibility, which affects a wide range of aspects of life, including human health and transportation. Satellite and its remote sensing technology can monitor a target area spatially. Visibility, one of the proxies for smoke quantifiers, has been proposed as the product of a satellite-based model that can benefit human life. This study used back-propagation in neural network (BPNN), a machine learning technology, to develop a visibility estimation model based on The Himawari-8 satellite using several combinations of BPNN tuning. It also compared the estimated visibility estimation with METAR data, as well as root mean square error (RMSE) and R2 correlation to check its accuracy. In this case, visibility was classified into three, namely class 1 visibility (below 1,600 m), class 2 (between 1,600 and 3,000 m), and class 3 (more than 3,000 m). The results showed that the highest accuracy of the visibility estimation model was obtained from the combination of input bands no. 2,4,5,11, 13, 14,15, with R2 correlation of 0.703.