基于特征选择和GA-BP神经网络的多源遥感农田土壤水分反演

土壤水分是影响水文、生态和气候等环境过程的重要参数,而微波遥感是农田地表土壤水分测量的重要手段之一。针对微波遥感反演农田地表土壤水分受植被覆盖影响较大的问题,该研究提出了一种基于特征选择和GA-BP神经网络(Genetic Algorithm-Back Propagation neural network)的多源遥感农田地表土壤水分反演方法。首先对Sentinel-1微波遥感数据和Sentinel-2光学遥感数据进行预处理并提取21个特征参数;然后采用差分进化特征选择(Differential Evolution Feature Selection,DEFS)算法从21个特征中选出包含10个参数的最优特征子集,并利用主成分分析(Principal Component Analysis,PCA)法将特征子集进行降维;之后建立BP神经网络,采用遗传算法(GeneticAlgorithm,GA)对BP网络的节点权值进行优化,使用降维后的特征矩阵和部分实测土壤含水量数据对BP网络进行训练;最后利用训练好的GA-BP网络对研究区土壤水分进行反演,并利用实测数据对反演结果精度进行对比验证。试验结果表明,该研究反演结果的决定系数为0.789 3,均方根误差为0.028 7 cm~3/cm~3,相比单纯使用GA-BP神经网络,加入DEFS和PCA之后决定系数提高了0.215 7,同时均方根误差降低了0.029 5 cm~3/cm~3。该结果展示了DEFS和PCA算法在土壤水分反演最优特征集选择的有效性,为多源遥感农田地表土壤水分反演提供了新思路。

Retrieval for soil moisture in farmland using multi-source remote sensing data and feature selection with GA-BP neural network

Measurements of soil moisture have great significance for agricultural production and environmental protection. A variety of technologies have been introduced into the current monitoring of surface soil moisture, particularly widely-used optical remote sensing and synthetic aperture radar (SAR) microwave remote sensing. In this study, retrieval for surface soil moisture in farmland was proposed using multi-source remote sensing data and feature dimension reduction with GA-BP neural networks. Sentinel-1 microwave and Sentinel-2 optical remote sensing data were used with a high spatial and temporal resolution. Three field surveys and sampling were carried out simultaneously during the transit time of the Sentinel-1 satellite. A total of 20 sampling points were set on the ground in the study area to collect soil moisture in the longitude and latitude coordinates. Firstly, Sentinel-1 and Sentinel-2 data were preprocessed to extract 21 feature parameters, including 9 backscatter coefficient, 5 polarization characteristic parameters, surface roughness, and 6 vegetation indices. Then, the differential evolution feature selection (DEFS) was utilized to obtain an optimal feature subset with 10 parameters, including the incident angle, ?0 VV, ?0 VH, ?0 VH/?0 VV, Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Water Band Index (WBI), and Fusion Vegetation Index (FVI). Principal component analysis (PCA) was used to further reduce the dimension of the optimal feature subset. The feature subset was reduced to eight dimensions, where an eight-dimensional feature matrix was obtained. After that, back propagation (BP) neural network was established to describe the nonlinear relationship between characteristic parameters and surface soil moisture, whereas, genetic algorithm (GA) was used to optimize the node weights and accelerated the learning speed of the BP neural network. The feature matrix after dimension reduction and some measured data of soil moisture were input into the GA-BP network for training, where the distribution map of soil moisture was obtained in the study area. Finally, taking a winter wheat field in Henan Province as the study area, three comparative experimental schemes were set to verify the accuracy of inversion using the measured data. The experimental schemes with DEFS and PCA presented the highest accuracy, where the coefficient of determination was 0.789 3, and the root mean square error was 0.028 7 cm3/cm3, compared with the genetic BP neural network, indicating the coefficient of determination increased by 0.215 7, and the root mean square error was reduced by 0.029 5 cm3/cm3. Meanwhile, the frequency distribution of soil moisture inversion was basically consistent with the measured soil moisture of sampling points. The experimental results demonstrated that the GA-BP network combining with DEFS and PCA can eliminate the redundant characteristic parameters for high inversion accuracy, and a high-resolution distribution map of surface soil moisture with a large area. The finding can offer some advantages and application potentials to the surface soil moisture retrieval in farmland.