Improvement of a Near-Real-Time Precipitation Estimation Algorithm Using Deep Learning

Near-real-time estimation of precipitation from geostationary satellites plays a vital role in natural disaster mitigation due to timely monitoring, high spatial–temporal resolution, and large coverage, yet this letter remains a large challenge. In this letter, a novel deep learning-based algorithm entitled precipitation estimation using a multiscale network (DLPE-MS) is proposed to estimate precipitation during summer over the eastern Continental United States (CONUS) of America. When inputting bispectral satellite information (10.3 and 6.2
$\mu \text{m}$
), this algorithm provides near-real-time rainfall rates at hourly and
$0.04^{\circ}\times 0.04^{\circ}$
resolution. In order to emphasize the information of a precipitation region at different scales using satellites’ data, we design a multiscale framework based on convolutional neural networks (CNNs). In addition, a loss function named balanced weight mean square error (BWMSE) is proposed to settle the problem of underestimation caused by a shortage in heavy rainy samples. Compared with the mean square error (MSE), the BWMSE has more balance parameters for different objects when training, which is able to mitigate the deviation between the prediction and ground truth in tailed categories (heavy rainy). Results show that this algorithm achieves the highest probability of detection (POD) and correlation coefficient (CC) with the value of 95.5% and 0.5. The statistical results of the precipitation cases also show that the DLPE-MS can significantly improve the estimated values in tailed categories than other products and methods. After testing, this algorithm is able to estimate the precipitation for the study area within 0.19 s.