Rainfall Estimation Using Convolutional Neural Networks in Various Environmental Backgrounds

Recently, rainfall estimation systems using surveillance cameras and Convolution Neural Networks are one of the exciting topics to discuss. This system is proposed to solve the limited number of rainfall instruments. Unfortunately, the performance of the existing method hasn't got satisfactory results. This study focuses on the surveillance camera background effect in estimating rainfall using various Convolution Neural Networks. The image background was collected from four different environments: a building, a tree, an asphalt road, and a combination of such areas. This paper used two Convolution Neural Networks (CNN) architectures, VGG16 and Resnet32. An actual RGB image was extracted from standard commercial surveillance cameras used as input for the model with a 360× 240 pixels resolution. The rainfall target is collected from an automatic rain gauge and classified into six classes. The total image sample of all backgrounds is 29880, used to train and test the model. Based on the experiment, the roads and trees background have the highest accuracy consistently for both architectures. This result shows the contrast between the rain streak and the image background is needed to increase the model performance.