全球36 km格网土壤水分逐日估算

土壤水分是陆面生态系统和能量循环的核心变量之一,利用微波遥感技术获得的土壤水分产品的时间分辨率一般是2-3 d,因此精确地获得具有较高时间分辨率的土壤水分成了人们关注的焦点。本文尝试将SMAP (the Soil Moisture Passive and Active）土壤水分和MODIS光学数据相结合,利用广义回归神经网络进行全球36 km土壤水分的估算,提升SMAP土壤水分的时间分辨率。结果显示,广义回归神经网络估算土壤水分与SMAP保持了高相关性(r = 0.7528),但其却保留了较高的误差 (rmse = 0.0914 m³/m³)。尽管如此,估算的土壤水分能够很好地保持SMAP土壤水分的整体空间变化,并且提升了土壤水分的时间分辨率（1 d）。此处,本文研究了SMAP土壤水分与MODIS光学数据之间的关系,这对今后利用机器学习进行SMAP土壤水分降尺度研究提供了重要的参考价值。

Daily Estimate of Global 36 km Grid Soil Moisture

Soil moisture is one of the core variables in land surface ecosystem and energy cycle. For the strong penetration of the cloud, rain and atmosphere, microwave remote sensing has advantages of the high precision in soil moisture retrieval. Currently, there are many passive microwave sensors or satellites used for surface soil (<5 cm) moisture observations, such as NASA's AMSR-E (The Advanced Microwave Scanning Radiometer-Earth Observing System) and SMAP (the Soil Moisture Passive and Active) and the European Space Agency SMOS (The Soil Moisture and Ocean Salinity). Although the use of microwave sensor can get higher precision of soil moisture products. The errors of SMAP 36 km soil moisture products can be less than 0.04 m³/m³. The 2~3 days revisited time restricts the applications that need the soil moisture products with higher temporal resolution (1 days). Therefore, it has been drawn more and more attention to get the accurate soil moisture with higher temporal resolution for the global weather prediction. Although the SM retrieval from MODIS data has higher error than retrieval from passive microwave data, the temporal resolution of MODIS data (1 day) is higher than the passive microwave data. For the different advantages of MODIS and passive microwave data, the combination of the two data for soil moisture retrieval may get the SM products with the MODIS temporal resolution and the similar accuracy or similar spatial variation of passive microwave data. In this study, we attempt to combine SMAP 36 km soil moisture product and MODIS optical/thermal infrared data to estimate the global 36 km soil moisture. This improve the temporal resolution of SMAP soil moisture from the 2~3 days to 1 day. By using the generalized regression neural network (GRNN) method, we simulated the relationship of SMAP soil moisture with MODIS global surface temperature and the surface reflectance products. Then we estimated the global 36 km soil moisture using the GRNN simulated relationship. In order to prevent overfitting of GRNN, all sample data according to the ratio of 0.8:0.2 is divided into training dataset and validation dataset. With the increase of the spread factor, the performance of GRNN prediction of the validation dataset shows a decreasing trend after the first increase, and GRNN obtained the maximal correlation coefficient (r) and root mean square error (rmse) with 0.02 of the diffusion factor. Finally, the well trained GRNN is used to estimate the global 36 km soil moisture. The results show that the accuracy of the GRNN for soil moisture estimate has a high correlation with SMAP (r=0.7528), but it retains a high error (RMSE=0.0914 m³/m³). For the cloud contamination of MODIS data, there has a part of loss of GRNN 36 km soil moisture estimate. Nevertheless, the GRNN estimated soil moisture can be very good to maintain the overall spatial variation of SMAP soil moisture, and enhance the temporary resolution of soil moisture from 2~3 days to one day. Besides, the relationship between SMAP and MODIS data is also studied in this paper, which can provide a significant reference for SMAP 36 km soil moisture downscaling by the machine learning.