Soil moisture detection based on data fusion between near-infrared spectroscopy and machine vision

The aim was to find out a way to accurately, rapidly and stably measure soil moisture and extend the model adaptation. Data fusion technology of machine vision (MV) and near-infrared spectroscopy (NIRS) was introduced to analyze soil moisture. Three kinds of soils (paddy soil, yellow brown soil and tidal soil) were collected from Hubei province to construct soil moisture analysis model based on NIRS; Soil surface image characteristics technique was used to build soil moisture analysis models using those three kinds of soils. NIRS was found to be influenced by sample state, so fusion technology of MV and NIRS was used. The results showed that soil moisture analysis model based on NIRS was quite accurate, but the model error of analysis of loess soil samples which were not included in the modeling sample set, was greater than 4%; Image parameters such as H, S and V were taken as input for the home network optimal prediction model, and decision coefficient R2 was obtained as high as 0.9849, but comparatively large error occurred when the model was applied to water-saturated samples (soil moisture＞20%); However, the problem was successfully solved by implementing BP fusion neural network model with R2=0.9961 and validation analysis error of water samples was less than that produced either by MV or NIRS.