An efficient video-based rainfall intensity estimation employing different recurrent neural network models
Farshid Rajabi,Neda Faraji,Masoumeh Hashemi
DOI: https://doi.org/10.1007/s12145-024-01290-x
2024-03-26
Earth Science Informatics
Abstract:The evaluation of precipitation using conventional approaches encounters various challenges stemming from limited financial resources, restricted geographic coverage, and reduced accuracy. In contrast, today urban areas are widely equipped with surveillance cameras, which, combined with advancements in deep learning techniques, has allowed for the creation of a highly effective and precise deep learning model capable of estimating rainfall intensity based on captured videos. In this article, a hybrid regression model based on Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) is developed to estimate rainfall intensity from one-minute videos. Three recurrent units named SimpleRNN, GRU and LSTM have been employed in our hybrid CNN+RNN model, which were tested with Adam, RMSProp and SGD optimizers. In particular, the mean absolute percentage error (MAPE) reports the value between 3.55%-6.95%, which are far better results compared to competitive rainfall estimation models. Also, the proposed model shows a very good performance in terms of complexity by recording the time of 0.47 seconds to calculate 100 frames. This prompt and precise measurement of precipitation is of potential to be applied in effectively managing water resources and mitigating the potential risks of flooding. Our regression model provides rainfall intensity estimates as output in a continuous space, in contrast to some methods, which only categorize the rainfall intensities into some discrete levels. Also, it receives one-minute videos as input, which eliminates the need to interpolate labels to obtain rainfall intensities in multi-millisecond scale as required in image-based algorithms.
geosciences, multidisciplinary,computer science, interdisciplinary applications