Enhanced Generalized Regression Neural Network With Backward Sequential Feature Selection for Machine-Learning-Driven Soil Moisture Estimation: A Case Study Over the Qinghai-Tibet Plateau.

Soil moisture (SM) is affected by many factors, such as soil characteristics, land cover, and meteorological conditions, making accurate remote sensing SM estimation a tough task. To fully explore the complementary information of multisource remote sensing data in SM estimation, it is necessary to explore the multiple feature variable selection method. Traditional filter methods may lead to feature redundancy and low accuracy, and embedding methods usually require complex parameter optimization. To overcome the above issues, we propose an enhanced generalized regression neural network with backward sequential feature selection (EBSFS) method for SM estimation. By using $k$-fold cross-validation to obtain the training set and validation set, and using the Pearson correlation coefficient to design evaluation criteria and an objective function, EBSFS searches for feature variables that minimize the objective function and updates the feature subset during iteration. EBSFS can adaptively obtain the optimal number of feature variables based on the evaluation criteria. Moreover, EBSFS does not require parameter optimization and can be flexibly and conveniently embedded into ensemble learning framework. Experiments conducted over the Qinghai-Tibet Plateau (QTP) from April 2015 to March 2016 demonstrate that EBSFS greatly reduces the feature redundancy, produces a more compact feature subset, and achieves higher estimation accuracy. Precisely, EBSFS presents better performance with R = 0.9544 and RMSE = 0.0310 under 13 input feature variables.