Study on Estimation of Deserts Soil Total Phosphorus Content by Vis-NIR Spectra with Variable Selection

In this paper, 300 samples of desert soil collected in the Ebinur Lake Wetland Nature Reserve of Xinjiang were used as the research subject, and the visible/near-infrared spectra data about the soil obtained with the ASD Field Spec (R) 3 HR spectrometer and the data about total phosphorus in the soil obtained through chemical analysis were used as the data sources; following Savizky-Golay smoothing, standard normal variation transformation and the first-order differential pretreatment, the combination of ant colony optimization interval partial least squares (ACO-iPLS) and genetic algorithm interval partial least squares (GA-iPLS) were employed to extract the characteristic wavelengths of the total phosphorus content in desert soil, before the partial least squares regression model for predicting the total-phosphorus content in soil was constructed; and this model was compared with the full-spectrum partial least squares model, ACO-iPLS and GA-iPLS. According to the results: through filte-ring with ACO-iPLS, the total-phosphorus characteristic wavebands in the desert soil were 500 similar to 700, 1 101 similar to 1 300, 1 501 similar to 1 700, and 1 901 similar to 2 100 nm; through further variable selection with GA-iPLS, 13 effective wavelengths with the minimum co-linearity were selected, which were respectively: 1 621, 546, 1 259, 573, 1 572, 1 527, 564, 1 186, 1 988, 1 541, 2 024, 1 118, and 1 191 nm. According to the comparison of modeling methods, the most accurate model was the one based on the characteristic variables selected with the combination of ACO-iPLS and GA-iPLS, followed by the ones with genetic algorithm, ant colony optimization algorithm and the full spectrum method. For the total phosphorus content in soil model established with the combination of ACO-iPLS and GA-iPLS, the root mean square error of cross validation (RMSECV) and the root mean square error of prediction (RMSEP) were respectively 0. 122 and 0. 108 mg.g(-1), and the related coefficient for cross validation (R-e) and the related coefficient for prediction (R-p) were 0.535 7 and 0.55 59, respectively. Therefore, it can be seen that the model constructed through Savizky-Golay smoothing, standard normal variation transformation and the first-order differential pretreatment and by using the combination of ACO-iPLS and GA-iPLS has simple structure, high prediction accuracy and good robustness, and can be used for estimating the total phosphorus content in desert soil.