Fusion of Vis-NIR and XRF spectra for estimation of key soil attributes

Abstract

Precision agriculture (PA) is an integrated solution to optimize the application of farming inputs site-specifically to increase crops efficiency, reduce resources consumption, and protect the environment. Recently, visible-near- infrared (Vis-NIR) and X-ray-fluorescence (XRF) have emerged as efficient and inexpensive proximal soil sensing (PSS) methods for assessment of soil attributes in support of PA applications. The goal of this study was to explore the potential of data fusion in accuracy improvement of predicting key soil attributes using Vis-NIR and XRF spectra. More specifically, local weighted regression (LWR) and partial least squares (PLS) regression methods were examined and compared in prediction of the following soil attributes: pH, organic carbon (OC), phosphorous (P), magnesium (Mg), calcium (Ca), and sodium (Na). Besides the individual-sensor models, we evaluated three data fusion approaches basedon spectra fusion (SF). The first method concatenates the full Vis-NIR and XRF spectra before PLS (SF1-PLS). In the second SF approach (SF2-PLS), the concatenation of the principal components (PCs) of the two spectra types – each resulted from a principal component analysis (PCA) – were used as input to PLS. The last proposed method, SF-CNN, uses a convolutional neural network (CNN) with input in form of a matrix whose columns are the PCs of the Vis-NIR and the PCs of the XRF spectra. To evaluate the proposed methods, a total 267 soil samples of nine arable fields were collected and scanned in laboratory settings using a Vis-NIR and an XRF spectrometer. For each soil attribute, the samples of six fields were used as the calibration set while the samples of the remaining three fields were used for validating the prediction models. The individual-sensor and SF models were evaluated in terms of ratio of performance to inter-quartile (RPIQ), root mean square error (RMSE), and Lin’s concordance correlation coefficient (LCCC). The validation results showed that PLS generally outperforms LWR. Also, the prediction accuracy obtained from Vis-NIR spectra was better than that of XRF for all the attributes, except for Ca. The proposed SF schemes, especially the SF-CNN, outperformed the individual models in the prediction of all attributes, except P and Ca. The best improvement of the prediction results, compared to the individual-sensor models, was obtained for Mg (RPIQ = 1.60, RMSE = 6.88 mg/100 g, LCCC = 0.93), successively followed by pH (RPIQ = 2.52, RMSE = 0.36, LCCC = 0.68), and Na (RPIQ = 1.69, RMSE = 2.55 mg/100 g, LCCC = 0.65). Overall, the results of this study suggest the CNN-based fusion for the prediction of the studied key soil fertility attributes in order to enable making accurate decisions and control loop in PA applications using high-resolution data on spatiotemporal variability of soil.