The Index Optimization Method in Neural Network for Soil Moisture Inversion

Abstract

Soil moisture, an important evaluation index in the field of environmental studies, plays a vital role in the exchange process between global surface and atmosphere. Although its content just takes a small percentage of freshwater resources, it is involved in the surface evapotranspiration process, moisture exchange process and many other cyclic processes. Temperature vegetation dryness index (TVDI) is a major mean that is based on optical and thermal infrared remote sensing to inverse soil moisture. However, its inversion accuracy is affected by soil background and the resolution of the thermal infrared data. Aiming at solving the problem of limited conditions of the data and complicated mathematical relations in modeling, ASTER NIR/TIR data is used in this study, and normalized differential vegetation index (NDVI) is replaced by the modified soil-adjusted vegetation index (MSAVI). Then, piecewise linear model is used to downscale the resolution of land surface temperature (LST), and modified temperature vegetation dryness index (MTVDI) is gained. Finally, back propagation (BP) neural network is established to calculate soil moisture. The result shows that the precision root mean square error (RMSE) of the two-parameter optimization model is higher than that of the none optimization model and single parameter optimization model. Improving the precision of soil moisture inversion by optimizing the input parameters of the neural network is feasible.