Hyperspectral analysis for automated quantification of total phosphorus in enriched soil samples

Phosphorus is an essential macronutrient for plant development, and its availability in soil directly influences agricultural productivity. However, traditional laboratory quantification of phosphorus is costly, slow, and destructive. This study introduces a system for automated quantification of total phosphorus (TP) using hyperspectral analysis on soil samples enriched with phosphorus fertilizer (P₂O₅). A previously developed acquisition protocol by the authors was employed, involving the design, development, and construction of a platform equipped with a Bayspec OCI-F camera. The lighting system was designed to ensure adequate spectral response in the visible (VIS) and near-infrared (NIR) regions, covering the range from 420 to 1000 nm. A total of 152 soil samples with varying phosphorus concentrations were prepared. From the hyperspectral images (HSI), the spectral response of each sample was extracted. The data were divided into 80% for training and 20% for validation. Partial least squares regression (PLSR) was used to estimate total phosphorus (TP), and variable importance in projection (VIP) analysis reduced the spectral bands from 145 to 78. Subsequently, a forward propagation artificial neural network (ANN) was trained to predict TP content in new samples. The system achieved a coefficient of determination (R²) of 0.99401, a ratio of performance to deviation (RPD) of 9.1, and a ratio of performance to interquartile range (RPIQ) of 13.9, indicating a good fit. Additionally, it achieved a mean absolute percentage error (MAPE) of 12.1% and a root-mean-square error (RMSE) of 7426 ppm, demonstrating reliable estimation of total phosphorus in soils.