ABSTRACT

Precipitation estimation with high spatial and temporal resolution is very important for monitoring floods and natural disasters. At present, a couple of quantitative precipitation estimation products and research methods can successfully estimate precipitation at one hourly temporal resolution. In this study, a deep learning model based on Convolutional Neural Network (CNN) was proposed to estimate the precipitation intensity based on the hyperspectral satellite FengYun-4/Advanced Geostationary Radiation Imager (FY-4A), and the temporal resolution is reduced to half an hour. Firstly, the importance of different channels and channel differences for precipitation intensity estimation was determined by ablation experiments. Secondly, compared with the existing model Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks–Convolutional Neural Networks (PERSIANN-CNN) and U-Net. The experimental results show that Small Wisely Network (SW-Net) provides more accurate precipitation intensity estimation, compared with PERSIANN-CNN (U-Net) in the same spatial and temporal resolutions. SW-Net outperformed PERSIANN-CNN (U-Net) by 5.9439% (5.6298%) and 6.3600 (5.8400) percentage points in the loss value and Mean Intersection over Union (MIoU), demonstrating the better feature extraction performance of the model. Furthermore, the False Alarm Ratio (FAR) of precipitation estimation with respect to Global Precipitation Measurement (GPM) Integrated Multi-satellite Retrievals for GPM (GPM-IMERG), for SW-Net was lower than that of PERSIANN-CNN (U-Net) by 49.2132% (49.4302%), showing the higher accuracy of proposed model.

摘要

具有高时空分辨率的降水估算对于监控洪水和自然灾害非常重要。目前，有两种定量的降水估算产品和研究方法可以在一小时的时间分辨率下成功估算降水。本研究提出了一种基于卷积神经网络（CNN）的深度学习模型，以基于高光谱卫星风云4号/先进对地静止辐射成像仪（FY-4A）估算降水强度，并将时间分辨率降低到一半一小时。首先，通过消融实验确定了不同通道和通道差异对降水强度估计的重要性。第二，与现有模型相比，使用人工神经网络-卷积神经网络（PERSIANN-CNN）和U-Net从遥感信息中进行降水估计。实验结果表明，与在相同时空分辨率下的PERSIANN-CNN（U-Net）相比，小型Wisely网络（SW-Net）提供了更准确的降水强度估算。SW-Net的损失值和联合平均交集（MIoU）的表现优于PERSIANN-CNN（U-Net）5.9439％（5.6298％）和6.3600（5.8400）个百分点，证明了该模型具有更好的特征提取性能。此外，SW-Net相对于全球降水量测量（GPM）集成多卫星检索（GPM-IMERG）的降水估算的虚警率（FAR）低于PERSIANN-CNN（U-Net） ）的49。