Estimating soil moisture from hyperspectra in saline soil based on EPO-PLS regression

The information of surface soil moisture is of great importance for the irrigation and production of agriculture. Researches have shown that surface reflectance spectra of soils are often jointly affected by soil moisture content and salt content, whichhas not yet been sufficiently addressed. In this study, we investigated the external parameter orthogonalization (EPO) method to eliminate the effect of soil salinity by preprocessing soil spectral reflectance and establishing EPO-PLS (partial least squares regression after EPO pre-processing) model to predict soil moisture content. Soil salt composition and texture were obtained by taking soil samples in Hetao Irrigation District, Inner Mongolia, China in July 2014. The components of soil salt were mixed to artificially create 11 levels (percentage by weight, g/(100 g)) of salt salinity in the soil samples: 0.1 (natural soil salt content), 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 1.0, 2.0 and 5.0%. The moisture contents of total soil samples were designed as relative weight 38%. Filling 11 replicate dishes (12 cm in diameter) with each level of salinity soil, respectively. Each dish was filled with about 374 g wet soil with 3cm depth and a bulk density of 1.3. A controlled laboratory experiment was conducted by a way of continuously monitoring changes of soil moisture and salt content. Soil reflectance spectra were measured for each level of salt salinity samples in a darkroom using Analytical Spectral Device FieldSpec 3 Hi-Res (ASD, USA) spectrometer covering wavelengths from 350 to 2 500 nm at an interval of 1 nm. Reflectance spectra and weight of each soil samples were measured every day until the weights remained unvaried (completely air-dried). Based on laboratory controlled experiments, this paper is mainly focused on the changes of slightly and moderately salt-affected soil reflectance spectra in the process of evaporation. We quantitatively analyzed the changes in soil reflectance of overall bands and the results suggested that a combination of salt and moisture in soil caused confusion of soil reflectance spectra. The reflectance spectra of soils are leveling off and almost converge to 0.2 after moisture content reaches 25% where the effect of soil moisture dominates in the reflectance spectra of soil. However, for salt-affected soil reflectance spectra, a clear increasing pattern is noted along with a decreasing of soil moisture, in which the severer the salinity is, the higher the reflectance value and the faster the rising speed of average reflectance would be. The effect of soil salt on soil reflectance spectra became predominant when more and more salt accumulated on the soil surface in a form of white crusts with high reflectance because of moisture evaporation. EPO is a method to reduce the space dimensionality in regard to external parameters which is referred to soil salt in this paper. PLS and EPO-PLS models were established to predict the moisture of salt-affected soil, respectively. The prediction results of PLS model show a significant deterioration and bias with an increase of soil salt content. It is clear that soil salt has a strong influence on the prediction of soil moisture content. Direct application of PLS models leads to an over-predicted results of moisture content of salt-affected soil, based on the spectra of non-saline soil samples. Through comparing PLS with EPO-PLS model, R2 and RPD rose from 0.722 and 1.976 to 0.898 and 3.145 for validation data, respectively. RMSE was reduced from 5.087 g/(100 g) to 3.237 g/(100 g). Results show the model quality of EPO-PLS for prediction of soil moisture increases significantly. EPO is verified to eliminate the effect of soil salt on spectra successfully. In this way more precise information of soil moisture can be predicted by establishing the partial least squares regression after EPO pre-processing and the approach should realize the soil moisture estimation well in saline area.