A CNN-based Differential Image Processing Approach for Rainfall Classification
Roberta Avanzato, Francesco Beritelli
Abstract
With the aim of preventing hydro-geological risks and overcoming the problems of current rain gauges, this paper proposes a low-complexity and cost-effective video rain gauge. In particular, in this paper the authors propose a new approach to rainfall classification based on image processing and video matching process employing convolutional neural networks (CNN). The system consists of a plastic shaker, a video camera and a low cost, low power signal processing unit. The use of differential images allows for greater robustness, guaranteeing full background subtraction. As regards precision, speed and ability to adapt to variations in precipitation intensity, the proposed method achieves good performance. In particular, the results obtained from seven classes, ranging from "No rain" to "Cloudburst", applying the Discrete Cosine Transform (DCT) to the differential images on 16x16 sub-blocks show an average accuracy of 75% considering, also, the adjacent missclassification. Furthermore, the analysis of precision and sensitivity parameters yields excellent results. The proposed method is very innovative, in fact, the few studies found in the state of the art use only two classification classes (No rain and Rain), while our method contains seven classification classes and overall delivers very good accuracy performance.