基于生成对抗网络模型的SMAPL4土壤水分产品降尺度分析

Downscaling analysis of SMAPL4 soil moisture products based on generative adversarial network models

土壤水分是地表和大气水热过程交换的重要纽带,对于农业生产以及优化种植结构具有重要意义,NASA卫星下的SMAPL4是一种以被动微波遥感技术为手段对土壤湿度监测的产品,具有可穿透云层和全天候监测等能力,但其较低空间分辨率很难满足小尺度或小区域范围的实际研究需求。鉴于此,根据云南省姚安县高原灌区特殊的地理位置,引用相关系数推演得出与研究区土壤水分空间分布有关的解释变量,沿用随机森林算法,耦合1 km包含地表温度和归一化植被指数的MODIS地表产品,建立基于RF全局窗口线性回归的1 km级被动微波土壤水分空间降尺度模型;而后堆叠地表温度(LST)、归一化植被指数(NDVI)、降水量(Prec)、地表蒸散量(ET)等4个变量形成条件生成对抗网络框架,并使用均方误差(RMSE)和条件生成对抗性损失函数训练神经网络来建立低分辨率和高分辨率映射关系,随即获得降尺度后土壤水分空间分布结果;最后将实际采样和监测站点提供数据做空间平均聚合后,与SMAPL4原始结果的CGAN、RF降尺度结果进行对比分析。结果表明:LST、NDVI、Prec、ET与土壤水分的相关性均值均大于0.44,具有相关关系,条件生成对抗网络降尺度结果对指标R^(2)和Bias表现效果最好,均值分别为0.7和0.032;RF降尺度结果对RMSE的效果最好,均值为0.006。同比SMAPL4原始数据,RF结果空间分布更为平滑,但极值差异性较大;CGAN结果能有效表征土壤含水空间分布状况,其数据变异性和极值表征能力更为突出。经RMSE与对抗性损失函数训练后,认为0.2~0.28的值域分布为降尺度后的研究区土壤水分数值分布结果。

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 moisture,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 coefficients 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 temperature 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 temperature(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 using 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 R^(2) 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 SMAPL4,and the RF results had the best effect on the RMSE raw data,the RF results had a smoother spatial distribution,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 distribution of soil moisture in the study area after downscaling.