Rainfall detection on a tropical island using a communication satellite downlink signal and artificial intelligence models

Abstract

The development of rainfall estimates based on satellite links in several high latitude regions shows good results. However, its development for tropical regions is still very less promising, leaving the question of which method is worth using. In this work, we investigate artificial intelligence (AI) models by utilizing the Ku-band satellite downlink signal for rainfall detection on a tropical island. Long-term evaluation of all the models is performed using one year of rainfall intensity data. The findings of this study demonstrate that one-dimensional convolution neural network (1DCNN), long short-term memory (LSTM), and random forest (RF) models can effectively identify rainfall occurrences on tropical islands, with the area under the receiver operating characteristic curve (ROC) value of 0.84, 0.84, and 0.83, respectively. These three models consistently provide better accuracy in detecting the region’s diverse rainfall patterns throughout the year. The 1DCNN has better performance compared to the other models, achieving accuracy values of 0.76 and 0.82 during the periods of lowest and highest rainfall, respectively. Considering the presence of numerous potential mislabeled rain samples in the dataset, this result is excellent. In addition, the utilization of higher resolution rainfall intensity data can be a great potential for real-time rainfall monitoring in the tropical region.