近红外光谱和机器视觉信息融合的土壤含水率检测

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

为了精确、快速和稳定测定土壤含水率以及扩大所建模型的适应性,该文提出了机器视觉与近红外光谱技术融合的土壤含水率分析方法。通过试验建立了湖北地区主要土壤基于近红外光谱的土壤含水率分析模型、基于土壤表层图像特征参数的含水率分析模型和机器视觉与近红外光谱信息融合的土壤含水率分析模型。结果表明,基于近红外光谱含水率分析模型虽然具有较高的精度,但该模型预测非建模样品黄绵土误差均大于4%;以图像特征参数H,S和V所建BP人工神经网络非线性预测模型最优,模型的决定系数R2为0.9849,但当土壤水分饱和(达到20%以上)时存在分析误差;而所建立的土壤的近红外光谱与机器视觉BP神经网络信息融合模型可预测非建模样品黄绵土与水分饱和达20%以上土壤,决定系数R2可达到0.9961,融合模型分析精度均高于单独使用近红外光谱或机器视觉分析模型。

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 Rz 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 R^2=0.9961 and validation analysis error of water samples was less than that produced either by MV or NIRS.