Proximal sensor data fusion for Brazilian soil properties prediction: exchangeable/available macronutrients, aluminum, and potential acidity

Abstract

Proximal sensing has achieved widespread popularity recently in soil science and the combination of different sensors and data processing methods is vast. Yet, confusion exists about which sensor (or the combination of sensors) is worthwhile considering the budget, scope, and the goals of the project. Hence, this work aims to test many modeling combinations using pXRF, Vis-NIR, and NixPro™ data and several preprocessing methods to offer a general guideline for exchangeable/available macronutrient (Ca2+, Mg2+, K+, P-rem), exchangeable Al3+, Al3+ saturation and soil potential acidity predictions (H++Al3+). A total of 604 samples were collected across four Brazilian states. Five types of spectra preprocessing, two sample moisture conditions for color, and the addition of extra explanatory variables were tested. The manifold combinations of these factors were modeled as continuous and categorical variables using the random forest algorithm and yielded 9310 models, from which prediction results were validated. The best results were achieved by fusing all sensors, proving the complementary nature of sensor data. However, pXRF data were key to significantly improving the predictions. Exchangeable Ca2+, Mg2+, Al3+, and Al saturation presented the best prediction results (R2 > 0.75), while available K+ and H++Al3+ had poor predictions (R2 < 0.5). Separating models by soil order improved predictions for Ultisols. Binning was the spectra preprocessing method that appeared most frequently in the best-performing models. The dry and moist color showed little effect in predictions. Categorical validation improved the usability of poorer models and maintained the good performance of the best models. Data fusion provided optimal results combining the three sensors, but pXRF provided key data for the good performance of combined sensor datasets.