Key points are not available for this paper at this time.
Soil health is essential to global sustainable food production. Beyond its role in food production, soil also plays a crucial role in maintaining ecosystem health and mitigating climate change. Monitoring and improving the health of agricultural soils requires insight into spatial variation in soil properties and associated ecosystem functions. Measuring this variation via classic sampling and analysis on field, regional or global scale is challenging due to high spatial variability inherent to soils and to the lack of affordable and reliable measurement methods. We present here a novel and worldwide applicable approach combining NIR spectroscopy using proximal sensors, remote sensing data and deep learning models to predict the main soil properties controlling soil health in the field. These include the soil texture (clay, sand, silt), soil pH and buffered cation exchange capacity, the organic and inorganic carbon content and soil nutrient contents for nitrogen, phosphorus (P) and potassium (K). The designed model infrastructure is shown to predict all soil properties (except for P and K) on the LUCAS dataset well (R2>0.8), and that predictive performance of field-state samples can be made comparable to lab-dried performance through transfer learning and sensor fusion with globally available covariates. These findings show that proximal soil sensing has high potential for soil health assessments and tailor-made recommendations regarding crop, soil and fertiliser management measures.
Kok et al. (Wed,) studied this question.