Rain/no-rain classification from combined radar- Radiometer data using machine learning

Abstract

Satellites with microwave-based sensors are launched to study and estimate global precipitation due to the instrument's sensitivity to precipitation, as well as their global coverage. The Tropical Rainfall Measuring Mission (TRMM) mission jointly launched by NASA of US and JAXA of Japan to study the global rain within the tropics had two microwave-based instruments on board; the Precipitation Radar (PR) and the TRMM Microwave Imager (TMI). Radar, being an active instrument, measures rainfall by detecting the echo accurately over both land and oceans. On the other hand, the passive instrument, radiometer, senses radiation emerging from the land/ocean surface, passing through the raining atmosphere. Due to the absence of contrast over land pixels, radiometers do not perform well over land surfaces in the detection of rainfall. Hence, over land rain screening algorithms are predominantly based on high-frequency channels which are sensitive to the scattering of radiation. However, their performance in screening the pixel for rain is based on threshold values. The present study demonstrates a superior algorithm powered by machine learning (ML) techniques. We apply the artificial neural network (ANN) and the random forest algorithm for screening the pixels over land for rainfall. The algorithms are trained with TMI measured brightness temperatures, while the TRMM PR detected rain flag is used as the truth. The ML models were trained with 2012 JJAS data over the Indian landmass. The performance of the ML models is tested with the monthly data of 2012 and with the 2013 JJAS dataset over the same region to demonstrate the robustness of the model. The imbalance of the dataset (a small fraction of rain cases as compared to no-rain cases) has been addressed using the synthetic minority oversampling technique (SMOTE). A comparison is made between the two ML algorithms against the GPROF algorithm, which is implemented in the operational data product for the TRMM version 7 data product. Results show that the ML-based methods performed better as compared to the GPROF. On the 2013 JJAS dataset, the maximum F1 score obtained was 0.65 for the actual dataset, while the score with oversampled dataset was 0.68. For the same dataset, the GPROF algorithm showed an F1 score of 0.59. Apart from the nine brightness temperatures as input, new derived features were also selected based on two feature selection methods: mutual information and analysis of variance. It was seen that the models' performance on the two feature sets obtained from the two methods were very much similar.