Downscaling analysis of SMAPL4 soil moisture products based on generative adversarial network models
Soil moisture is an important link in the exchange of water and heat processes between the surface and the atmosphere,which is of great significance for agricultural production and optimization of planting structure.SMAPL4 under NASA satellite is a passive microwave remote sensing technology as a means of monitoring soil mois-ture,with the ability to penetrate the clouds,all-weather monitoring.However,due to its low spatial resolution,it is difficult to meet the practical research needs at small scale or small area scale.In view of this,according to the special geographic location of the plateau irrigation area in Yaoan County,Yunnan Province,the correlation coeffi-cients were quoted to derive the explanatory variables related to the spatial distribution of soil moisture in the study area,and along the random forest algorithm,the coupled 1 km MODIS surface products containing surface tempera-ture and normalized vegetation index were used to establish a spatially descaled model of 1 km passive microwave soil moisture based on the linear regression of the RF global window,and then the four variables of surface tempera-ture(LST),normalized vegetation index(NDVI),precipitation(Prec),and surface evapotranspiration(ET)were stacked to form a conditional generative adversarial network framework,and the neural network was trained u-sing the mean squared error(RMSE)and the conditional generative adversarial loss function to establish the low-resolution and high-resolution mapping relationship,and then the results of the spatial distribution of soil moisture were obtained after the downscaling.In addition,the spatially averaged aggregated data from actual sampling and monitoring stations were compared and analyzed with the downscaled CGAN and RF results of the original SMAPL4 results.The results showed that the mean value of the correlation between LST,NDVI,Prec,ET and soil moisture was more than 0.44,which was correlated.The downscaling results of the conditional generation adversarial network(CGAN)had the best effect on the indicators R2 and Bias,with the mean values of 0.7 and 0.032,respectively.The downscaling results of RF had the best effect on the RMSE,with the mean value of 0.006.Compared with the original data of SMAPL4,the spatial distribution of the RF results was better than that of the original data of SMA-PL4,and the RF results had the best effect on the RMSE raw data,the RF results had a smoother spatial distribu-tion,but the polar value variability was larger.The CGAN results effectively characterized the spatial distribution of soil water content,and its data variability and polar value characterization ability was more prominent.After the training of RMSE and adversarial loss function,the value range of 0.2～0.28 was considered as the numerical distri-bution of soil moisture in the study area after downscaling.
万方数据
维普
