Soil organic matter determination based on artificial olfactory system and PLSR-BPNN

Abstract

Abstract Soil organic matter (SOM) is a key indicator of soil fertility. For accurate measurement of SOM, a novel method based on an artificial olfactory system (AOS) was proposed. The response curves of soil volatile organic compounds (VOCs) were measured using a metal-oxide semiconductor sensor array, and four features (including maximum value, mean differential coefficient, response area, and the transient value at the 20th second) were obtained from the curves and used to build olfactory feature space. Then, prediction models were established using the pattern recognition algorithm. To further enhance the accuracy of AOS measurement, we used Monte Carlo cross-validation (MCCV) to identify and eliminate the abnormal samples of the soil olfactory feature space. Then, the dimension reduction method of the genetic algorithm (GA)back-propagation (BP) was used to find the appropriate feature vectors, and two types of hybrid models were presented. One was the support vector machine (SVM) and group method of data handling (GMDH) combined model—SVM-GMDH. The other was a combination of partial least squares regression (PLSR) and back-propagation neural network (BPNN)—PLSR-BPNN. The forecasting performances of three single models (BPNN, PLSR, support vector regression: SVR) and two combined models (PLSR-BPNN, SVM-GMDH) were comparatively evaluated. The evaluation indices included coefficient of determination (R 2), root mean square error (RMSE), ratio of performance to deviation and relative prediction error (RPE). It was found that the predictive capabilities of all five tested models were improved after elimination of abnormal samples and feature reduction. Moreover, PLSR-BPNN performed the best in predicting SOM concentrations, with R 2 = 0.952, RMSE = 1.771, PRD = 4.291, and slight variation of RPE within 0–0.185, and thus can offer a reference for predicting SOM via AOS.