Multi‐sensor fusion for the determination of several soil properties in the Yangtze River Delta, China

Abstract

Summary Soil organic matter (SOM), total nitrogen (TN), available nitrogen (AN), available phosphorus (AP), available potassium (AK) and pH are key chemical properties for evaluating soil fertility and quality. This study involved the integration of four soil sensors, visible near‐infrared (vis–NIR) spectrometer, mid‐infrared (mid‐IR) spectrometer, portable X‐ray fluorescence (PXRF) analyser and laser‐induced breakdown spectroscopy (LIBS), to achieve rapid measurement of these soil properties. A genetic algorithm and partial least‐squares regression (GA–PLSR) were used to select characteristic bands to reduce data redundancy. We then calibrated models from three aspects: models using partial least‐squares regression (PLSR) based on single sensor data, models using PLSR based on fused sensor data, involving data combined from the four sensors into a new dataset to create a data fusion (DF) model, and models with Bayesian model averaging (BMA) based on prediction results of fused sensor data, involving prediction results combined from the four sensors into a new dataset to form the BMA model. The results showed the following. (i) For the single sensor, the predictive performance decreased as follows: mid‐IR > vis–NIR > LIBS > PXRF. (ii) Compared with the single sensor approach, the DF approach slightly improved or even reduced prediction accuracy and caused a large amount of redundancy. We suggest that this approach is not able to improve predictive ability. (iii) The BMA approach achieved the best prediction for the six soil properties. Our findings suggest that model averaging of vis–NIR, mid‐IR and LIBS could be a reliable and stable approach for the fast measurement of soil properties. Highlights We used four proximal soil sensors to evaluate six key properties for evaluating soil fertility and quality. GA–PLSR was used to select characteristic bands. We compared predictions of six soil properties from single sensor, DF and BMA approaches. BMA predictions were more accurate than predictions from single and fused sensor data.

Effect of X-Ray Tube Configuration on Measurement of Key Soil Fertility Attributes with XRF

“…Sensors compatible with PSS approaches can be applied in laboratory and field conditions, and some of them can be used for on-line measurements, assembled on mobile platforms [1]. In order to create practical alternatives for soil analysis, efforts have been made worldwide to adapt and re-engineer sensor systems developed in other areas into soil sensing [6,7], as well as assessing different combinations of sensors and data fusion methods [8][9][10]. However, there is still no consensus on an effective technique that enables the development of a generic and robust methodology for soil fertility analysis via proximal sensors.…”

“…X-ray fluorescence (XRF) spectrometry is one of the promising PSS techniques, as it allows analyses with satisfactory performance using minimum sample preparation [11]. Its recent popularization has made it attractive for joint implementation with other popular techniques, such as visible and near infrared spectroscopy (visNIRS) [8,9]. XRF sensors enable measuring the total content of specific elements in the soil (e.g., Fe, Al, Si, Ca, and K) and can be used as a proxy for indirect predictions of other soil attributes (e.g., pH, cation exchange capacity (CEC) and texture) [12].…”

Data fusion of vis–NIR and PXRF spectra to predict soil physical and chemical properties

“…It can improve the prediction because it makes use of complementary information from different sensors (Mouazen, Alhwaimel, Kuang, & Waine, ). Different methods have been used, including (a) directly concatenating various sensor data (Viscarra Rossel, Walvoort, McBratney, Janik, & Skjemstad, ), (b) converting sensor data using principal component analysis (PCA) and then concatenating principal components (PCs) (Ji et al, ), (c) using outer product analysis (OPA) to fuse different spectra (Terra, Viscarra Rossel, & Demattê, ; Xu, Chen, et al, ), (d) different sensors used as fixed effects and random effects respectively (Cardelli et al, ; Wang et al, ) and (e) model averaging to combine model results developed from individual sensors (O'Rourke, Minasny, Holden, & McBratney, ; O'Rourke, Stockmann, et al, ; Terra et al, ; Xu, Chen, et al, ; Xu, Zhao, et al, ).…”

“…In the USA, the combination of PXRF (elemental data) and vis–NIR spectra for soil chemical characterization has been patented (Weindorf & Chakraborty, ). Recently, a few studies have combined PXRF spectra with other sensor data, for example, vis–NIR, mid‐infrared (MIR) and laser‐induced breakdown spectroscopy (LIBS), using model averaging methods to predict soil physical and chemical properties (O'Rourke, Minasny, et al, ; O'Rourke, Stockmann, et al, ; Xu, Zhao, et al, ). However, these model averaging methods were established based on a back‐end approach, whereby different sensors were used independently to predict soil properties, and different weights were assigned to the predicted soil properties from model results to obtain the final prediction.…”

“…O'Rourke, Stockmann, et al () tested several preprocessing techniques and found that a Savitzky–Golay smoothing filter followed by either a standard normal variate (SNV) baseline correction or multiplicative scatter correction (MSC) produced better results. Xu, Zhao, et al () used a Savitzky–Golay algorithm with a window size of 13 and polynomial order 2 on PXRF spectra, which was then down‐sampled to a resolution of 0.15 keV. However, little research has been carried out to systematically compare the performance of different preprocessing algorithms for PXRF spectra.…”