Predictive Performance of Mobile Vis–NIR Spectroscopy for Mapping Key Fertility Attributes in Tropical Soils through Local Models Using PLS and ANN

Abstract

Mapping soil fertility attributes at fine spatial resolution is crucial for site-specific management in precision agriculture. This paper evaluated the performance of mobile measurements using visible and near-infrared spectroscopy (vis–NIR) to predict and map key fertility attributes in tropical soils through local data modeling with partial least squares regression (PLS) and artificial neural network (ANN). Models were calibrated and tested in a calibration area (18-ha) with high spatial variability of soil attributes and then extrapolated in the entire field (138-ha). The models calibrated with ANN obtained superior performance for all attributes. Although ANN models obtained satisfactory predictions in the calibration area (ratio of performance to interquartile range (RPIQ) ≥ 1.7) for clay, organic matter (OM), cation exchange capacity (CEC), base saturation (V), and exchangeable (ex-) Ca, it was not repeated for some of them when extrapolated into the entire field. In conclusion, robust mappings (RPIQ = 2.49) were obtained for clay and OM, indicating that these attributes can be successfully mapped on tropical soils using mobile vis–NIR spectroscopy and local calibrations using ANN. This study highlights the need to implement an independent test to assess reliability and extrapolability of previously calibrated models, even when extrapolating the models to neighboring areas.

Keywords

• proximal soil sensing
• on-the-go data acquisition
• soil fertility diagnostic
• precision agriculture
• digital soil mapping