Himawari-8 rainfall estimation from infrared channels based on machine learning methods

Abstract

Accurate information of rainfall estimate from the remotely sensed instrument is highly needed for many applications, including the disaster early warning system that requires heavy rainfall prediction for mitigation purposes. In this study we implemented machine learning methods to estimate rainfall from high temporal resolution data of Himawari-8 over a large area of Indonesia, where floods and landslides caused by heavy rainfall events have been the most frequent disaster for the last 10 years. All brightness temperature of Himawari-8 infrareds (IR) channels were involved to estimate rainfall at the current (near-real) time (t), 1 hour later (t+1), and 2 hours later (t+2) after the data received by using 9 different machine learning models. The rainfall dataset product from Global Satellite Mapping of Precipitation (GSMaP) were used for data training. Two machine learning processes were taken, first is for separating rain and no-rain area, and second is for determining rainfall rate category. The results showed that Multi Linier Perceptron (MLP) model had the highest accuracy in deriving rain area. While, for retrieving rain rate, the Linear Discriminant Analysis (LDA) model was the most accurate compared to the other models. A good accuracy from the LDA has been obtained for current time estimation (accuracy = 81%) as well as for prediction at one and two hours later (accuracy = 79%, and 78% respectively). A comparison test has also been performed to determine which variables have the most significant contribution to the rainfall retrieved. It showed that 3 IR channels at 6.2, 10.4, and 13.3 μm were the minimum predictors that should be used to obtain a minimum 79% accuracy. Addition of other predictors such as other channels of IR or a combination of brightness temperature difference (BTD) could increase its accuracy, i.e. 1%-2% which is not too significant.