Rainfall sensing through image processing for improving spatial-temporal rainfall maps in urban areas

Conventional rain gauging tools have spatial and temporal limitations, which presented new opportunities for rainfall sensing through images as a complementary data source to estimate the rainfall intensity with broad spatial coverage and high resolution. Therefore, this study aims to develop a rainfall classification and estimation model based on real‐world rain images captured by surveillance and smartphone cameras. The results showed that the developed CNN rainfall classification model had an overall accuracy of 89.62% and 93.6% for direct and transfer learning approaches, respectively. The CNN regression model (Model 2) developed from the surveillance camera dataset has achieved the highest performance accuracy with NSE of 98.93%. Meanwhile, approach 1 of the CNN regression model applied to the smartphone image dataset shows the best-performing results with NSE of 79.35%.