Research On Rainfall Intensity Recognition Technology Based On Deep Learning

Due to global climate change and rapid urbanization,urban flooding has become a frequent hazard in China,resulting in significant economic losses and casualties.Extreme rainfall is a direct and critical factor leading to urban flooding,and therefore rainfall data is of great importance for flood modelling and prevention.However,urban rainfall exhibits significant spatiotemporal heterogeneity,with rainfall intensity varying by more than 30% over spatial distances of three to five kilometers.Obtaining high spatiotemporal resolution rainfall data in urban areas at low cost remains a challenging and cutting-edge problem in the field of urban hydrology and flood research.Traditional rainfall measurement methods,such as rain gauges,ground-based radars,and meteorological satellites,cannot provide high spatiotemporal resolution rainfall data within urban areas.To address the above challenges,this thesis adopts an interdisciplinary approach to investigate the technology of rainfall intensity recognition based on deep learning.Specifically,this thesis proposes a deep learning-based rainfall measurement model that utilizes rainfall videos or images,which enables real-time recognition of rainfall intensity in urban rainfall events.This provides a new approach and path for obtaining low-cost,high spatiotemporal resolution urban rainfall data,thus providing an important data foundation and technical support for urban hydrological simulation,urban flooding prediction,and flood prevention and control.The specific research contributions and innovations of this dissertation are outlined below.This dissertation proposes a deep learning-based rainfall intensity recognition theory and framework based on rainfall images to obtain urban rainfall data.Specifically,we first establish a theoretical formula for deep learning-based rainfall measurement,which maps the rainfall image information Z(d,s)(d is the raindrop density and s is the raindrop size)to the rainfall intensity I.Then,we collect rainfall images using existing sensors(such as smartphones or traffic cameras)to train a convolutional neural network(CNN)rainfall measurement model(ir CNN model),and finally,use the trained ir CNN model to efficiently simulate rainfall intensity based on the collected rainfall images.Synthetic rainfall images are used for theoretical validation,while real rainfall images are used for model accuracy validation.Results show that the ir CNN model can accurately estimate rainfall intensity for both synthetic and real rainfall images,with an average relative error of approximately 16.5% for real rainfall image recognition.Moreover,the main advantage of the proposed ir CNN model is that it has the potential to obtain high spatiotemporal resolution urban rainfall data at a low cost.To address the bottleneck of the ir CNN model’s poor performance in nighttime and other challenging scenarios,this dissertation proposes an innovative raindrop extraction-ir CNN twostage rainfall intensity recognition technology to improve the generalization ability of the ir CNN model.A novel raindrop extraction algorithm is proposed in this study,which leverages the information of the background,noise,and raindrop layers to establish a theoretical formula and optimization algorithm for extracting raindrops from the original rainfall image.These raindrop images are then fed into the ir CNN model to simulate rainfall intensity.The proposed two-stage algorithm is validated on real rainfall events captured under various scenarios.The results show that compared to the ir CNN model using the original rainfall image as input,the proposed twostage algorithm significantly improves the generalization ability of the ir CNN model,with the nighttime recognition error reduced from 28.4% to 19.7%.To overcome the difficulties of large recognition errors in ir CNN models when there are rapid changes in background lighting,as well as the issue of label mismatch(i.e.,the mismatch between the instantaneous rainfall intensity captured by the rainfall image and the cumulative rainfall recorded by the rain gauge),this dissertation proposes a theoretical model and method for deep learning-based rainfall intensity calculation using rainfall videos.This study first decomposes rainfall videos into a sequence of images,and then uses CNN to extract rainfall information from these images.Subsequently,a recurrent neural network is employed to extract the temporal features of the video,thus enabling rainfall intensity recognition based on video information.The main feature of this method is that it uses rainfall videos(i.e.,a time-series of rainfall images)as input information to the model,rather than a single rainfall image,which allows it to directly matched with the cumulative rainfall at the same time and solve the problem of label mismatch.Additionally,this approach avoids accidental errors and further improves the accuracy of the model.The proposed method is tested on rainfall videos captured under different lighting scenarios,and the results show that the model can accurately calculate rainfall intensity under all scenarios,with average relative error less than 17%.The deep learning-based rainfall intensity recognition technique proposed in this thesis is applied to an open-source data,and the results showed a high accuracy in simulating rainfall intensity for both rainfall images and videos in the open-source dataset.The average relative error calculated using rainfall images is within 19%,and that using rainfall videos is within 15%,further demonstrating the broad applicability of this method to data from different sources.