利用SAR影像与多光谱数据反演广域土壤湿度

针对基于主动微波遥感途径开展广域土壤湿度反演的过程中,对植被和土壤粗糙度影响难以进行有效估算的问题,该研究联合多极化雷达和原始多光谱数据源,提出一种改进的卷积神经网络(Improved Convolutional Neural Network,ICNN)方法。该方法采用不同尺寸的卷积核对原始变量进行一维卷积运算,自适应提取能反映测区土壤湿度时空差异的高级特征维;同时,去除了传统卷积神经网络结构中的池化层,保证提取的特征信息完整。试验结果表明,在边长超过100 km的四川盆地研究区域内,模型预测值与样本数据相关系数达到0.934,预测值偏差服从均值趋近于0的正态分布,均方根误差为1.45%,误差分布范围小于6.3%,结果具有较高的可靠性。该方法可为精准农业、旱涝灾害等领域的广域监测研究提供一定的支撑。

Wide-area soil moisture retrieval using SAR images and multispectral data

Abstract: As an important physical quantity reflecting the state of the ground, soil moisture not only reflects the drought conditions of the land but also is a determinant of crop water supply. Because the microwave spectrum has high permeability with respect to vegetation and is sensitive to changes in soil moisture, soil moisture inversion models and algorithms based on active microwave remote sensing have developed rapidly in recent years. This study proposed an Improved Convolutional Neural Network (ICNN) method through combined multi-polarized SAR and original multispectral data, which achieved wide-area soil moisture inversion. Firstly, this study statistically analyzed the correlation among soil water content and VV, VH polarized SAR and the original red, near-infrared band data and incident angle. Statistics results confirmed that Soil moisture was affected by multiple variables and that the applied single input variable method could not effectively inverse soil moisture. Therefore, it was necessary to use various original influence factors in the inversion to improve the inversion accuracy and the robustness of the inversion model. Secondly, to better eliminate the influence of vegetation and soil roughness on the retrieval of soil moisture in wide-area, this study used 1×1, 2×1 and 3×1 convolution kernels to perform one-dimensional convolution operations on the original input variables and to adaptively extract the advanced feature dimensions that reflected the spatial and temporal differences in soil moisture. Simultaneously, the pooling layer in the traditional Convolutional Neural Network (CNN) structure was removed to ensure the completeness of the extracted feature information. Finally, to verify the validity of the model, the Sichuan Basin area with complex vegetation cover, many river systems and buildings was selected as the typical experimental area, and the "CLDAS-V2.0" soil moisture product of China meteorological administration was selected as the label for model training and the true value during model validation. The experimental results showed that with the increase of the number of iterations, the residual loss value and the training accuracy converged quickly, after 500 iterations, the convergence speed slowed down, in which the training accuracy fluctuated slightly, and the overall tended to rise slowly. Within the range of the study area which sides were longer than 100 km, the model predicted values had good consistency with the sample data with the correlation coefficient reached 0.934, and the results could be accurately reproduced in arid and humid areas. Analyzed the inversion results showed that the deviation range was mainly -6.3%-5.1%, the RMSE value was 1.45%. Although the average error was not 0, the value was less than the resolution of the inversion result, therefore, it could be considered that the error as a whole was a random deviation. In addition, from the perspective of the spatial distribution of the inversion results, the parts tended to have the largest positive and negative deviations had obvious concentration distributions. Most of the deviations in the areas with smaller errors had positive values, part of the reason for this phenomenon might be the influence of urban clusters and scattered buildings, that was, some high-noise samples still participated in model training. For negative deviation areas, it might be caused by factors such as dense vegetation and residual water noise samples. ICNN had good application potential in wide-area monitoring of precision agriculture, drought and flood disasters, and could provide some support for related research.