Retrieval of soil moisture at returning green stage of winter wheat using MODIS drought index and RBFNN

Abstract: Accurate and effective drought monitoring using remote sensing technology is essential for regional and global drought warning and forecasting. Especially, soil moisture (SM) is one of the key potential factors affecting agricultural drought. At present, most studies ignore the fact that soil moisture is a complex nonlinear coupling system，the research only using visible light, near infrared, short-wave infrared, thermal infrared and other remote sensing drought index has certain limitations in SM inversion. A new method of retrieving farmland SM of winter wheat at returning green stage based on MODIS and radial basis function neural network (RBFNN) is presented in this paper. Firstly, the adaptability of various MODIS-derived drought indices is analyzed in the period of seedling establishment of winter wheat, including soil water content, crop morphological change in water requirement, crop canopy water content and temperature and other parameters. Secondly, the correlation between the original remote sensing drought indices and the soil relative humidity 10 cm soil layer was analyzed. A comprehensive evaluation set of apparent thermal inertia (ATI), vegetation supply water index (VSWI), enhanced vegetation index (EVI), normalized difference infrared index band7 (NDIIB7), normalized multi-band drought index (NMDI) and temperature condition index (TCI) are selected to invert soil moisture of farmland. Finally, combining MODIS remote sensing drought index with RBFNN, the SM of farmland was retrieved synergistically, and the retrieved results were compared with those of BP-NN and linear regression (LR) models. The experimental study is conducted in Henan province of China, MODIS reflectance products (MOD09A1) and temperature products (MOD11A2) at returning green stage of winter wheat from 2001 to 2012 were used to extract remote sensing drought indices, and the global land cover product of MODIS (MCD12Q1) is selected to obtain cropland distributions. Soil moisture data is derived from the "China crop growth and development and farmland soil moisture ten-day dataset", which contains the soil relative humidity of 10 and 20 cm soil layer observed every 10 days by 17 soil moisture stations. The results show that the average accuracy of SM inversion model using RBFNN and multiple drought indices is 93.27%, which is increased by 2.92 and 9.97 percentage points compared with BP-NN and LR model, respectively. Most of the data points of the three models are concentrated around the 1:1 line, which indicate that there is a good correlation between the predicted value and the measured value. Compared with BP-NN model and LR model, the deviation between predicted value of RBFNN model and the 1:1 line is the smallest, the higher the regression correlation coefficient is, the higher the determination coefficient is. SM is significantly correlated with ATI, NDIIB7 and VSWI in the early growth stage of winter wheat. Multi-band remote sensing drought monitoring indices can comprehensively reflect the changes of crop physiology and morphology under soil water stress, and also determine the retrieval accuracy of the humidity model simultaneously. In this paper, only winter wheat in returning green stage is selected as the research object, and the comprehensive evaluation index set selected is not suitable for SM inversion of other growth stages of winter wheat, the adaptability of different remote sensing drought indices in different growth stages needs further experimental study. The study provides a new case for regional SM inversion from remote sensing-based drought monitoring indices and neural network.