Towards a more sustainable global soil health monitoring through soil spectroscopy

Healthy soil can be defined as a dynamic ecosystem that performs a variety of essential functions such as controlling plant disease, nutrient cycling, improving soil function with positive effects for filtering and storing water, and nutrient capacity, and contributing to improving crop production. Healthy soil also contributes to mitigating climate change by maintaining or increasing its carbon content. Therefore, information on spatial variability of key soil properties is essential for prioritizing and tracking land management interventions, from the small-scale farm level to the global landscape level. In addition, healthy soil is critical for achieving several SDGs such as #2, 3, 6, 13, 15 and 17. However, the determination of soil properties through wet chemistry measurements is often expensive and time-consuming process, and consequently, soil analyses are restricted to a limited number of soil samples. A method that predicts the soil properties fast, inexpensively, and accurately is soil spectroscopy, which can provide immense opportunities for monitoring important soil health indexes. In this study, the evaluation of three algorithms for predicting three key soil properties, soil organic carbon (SOC), pH and Magnesium (Mg) using mid-infrared spectral data were studied using a dataset of more than 3400 samples. The soil samples were collected across Sub-Saharan Africa (SSA) region using the well-established Land Degradation Surveillance Framework (LDSF) method developed from World Agroforestry (ICRAF). The developed trained calibration models were based on the widely used in the soil spectroscopy research Partial Least Squares (PLS) and Random Forest (RF), and the not applied so far Bayesian Regularization for Feed-Forward Neural Networks (BRNN). The dataset was split into calibration (70%) and validation (30%) sets. Furthermore, the threshold of 5% was applied and thus, only the data with value that lie between 5% and 95% of each soil property were included. In this way, the extreme values that will bias the model and the predictions were excluded. Results has shown that the calibration model developed based on the BRNN algorithm yielded the more robust predictions among the three studied soil properties (R² of val 0.90, 0.92, 0.87 for pH, Mg, SOC, respectively). The predictions utilizing soil spectroscopy for determining soil properties in this study are showing its extremely potential to be beneficial in the support of soil health.