Quantitative Inversion Of Precipitation Based On Satellite Observations And Deep Learnin

Accurate monitoring of precipitation plays an important role in weather analysis and forecasting,and meteorological disaster prevention.However,conventional ground-based precipitation observation methods(rain gauges and radar)have problems with incomplete coverage and temporal and spatial discontinuities.Precipitation retrieval product using satellite remote sensing data can supplement the conventional ground-based precipitation observation to make the precipitation data more complete.Traditional retrieval methods and machine learning methods have certain limitations.Therefore,this paper investigates quantitative precipitation estimation research using geostationary satellite infrared bright temperature data and GPM multi-satellite joint precipitation products based on deep learning methods.The main conclusions are summarized as follows:(1)A PRSOT deep learning model based on attention mechanism is constructed for the precipitation quantitative retrieval task.The comparative experiment of multi-head attention and single-head attention was carried out.It was found that multi-head attention could fully mine the data information and obtain more accurate characteristics related to precipitation,so that the model could make a more accurate quantitative estimate of precipitation events on the basis of accurately identifying the precipitation falling area.(2)Considering the problem of category imbalance in the original data,the original proportional Scenario1 dataset and the equal-scale Scenario2 dataset optimized by sampling technology were constructed for comparative experiments.The PRSOT model trained with the original proportional Scenario1 dataset,can identify the spatial location and morphology of large-scale precipitation,but there is an obvious underestimation of precipitation events above 5mm/h.The PRSOT model trained with the equal-scale Scenario2 dataset,has better capture and estimation ability for heavy precipitation events in the center of the precipitation falling area,and has a higher detection rate and smaller estimation error for precipitation events above 7mm/h.However,the increase in rain samples cause the PRSOT model to overestimate a large number of no rain samples as light rain.Therefore,there is a significant difference between the model estimation and GPM precipitation product in the judgment of the overall precipitation falling area.To sum up,there is scope for continued optimization of the model’s retrieval performance.(3)To address the problem in(2),two optimization strategies are proposed.Firstly,the precipitation falling area recognition subtask is added,and the U-Net convolutional neural network with strong spatial feature extraction ability is introduced to complete the task.Secondly,the bright temperature difference characteristic with more physical characteristic meaning is introduced.These two strategies enable the PRSOT-Unet joint model to enhance the ability to estimate heavy precipitation events based on accurate identification of precipitation falling areas.(4)The generalization of the PRSOT-Unet joint model was evaluated by using independent time period and spatial region data,and it was found that the model could still retrieval the spatial distribution of precipitation on independent data.