Avaliação de fósforo em solos utilizando espectroscópia de reflectância e algoritmos quimiométricos

Abstract

Maintaining adequate phosphorus levels in the soil is a determining factor for plant development, influencing vital metabolic processes. Its determination through laboratory methods based on chemical extraction reagents not only generates waste but also makes these techniques costly and time-consuming, rendering their continuous and regular application unfeasible in many agricultural scenarios. This hinders the precise characterization of agricultural areas. The present study aims to evaluate the predictive potential of reflectance spectroscopy for quantifying available phosphorus content in soils, through the application of different chemometric algorithms and transformations of spectral reflectance data acquired by an optical spectroradiometer within the 350–2500 nm range, along with variable selection. For this research, 75 samples of a Neossolo quartzarenico and 75 samples of an Argissolo acinzentado were collected in the Baixo Acaraú irrigation district, and subsequently treated with 15 different concentrations of the commercial fertilizer MAP in 5 replications. For spectral data acquisition, the FieldSpec Pro FR 3 optical sensor was used. In addition to raw reflectance, spectral normalization, Savitzky-Golay filtering, and continuum removal were applied for qualitative and chemometric analyses. Available phosphorus content in the samples was determined using a UV/VIS spectrophotometer based on the Mehlich 1 extraction solution. For the calibration of the phosphorus prediction models, multiple linear regression (MLR), principal component regression (PCR), and partial least squares regression (PLSR) were employed, considering all wavelengths as predictor variables, along with variable selection via Stepwise-Forward. The results showed that spectral normalization combined with PLSR regression on the full spectrum (450–2450 nm) was the most efficient approach for predicting available phosphorus. In the combined RQO + PAcd dataset, R² = 0.95, RMSE = 12.75 mg kg⁻¹, and RPD = 4.31 were obtained; for RQO alone, R² = 0.95, RMSE = 8.30 mg kg⁻¹, and RPD = 4.66; and for PAcd, R² = 0.92, RMSE = 17.76 mg kg⁻¹, and RPD = 3.30 consistently outperforming the PCR models in all cases. Variable selection through correlation filtering and stepwise–forward reinforced the relevance of bands associated with organic matter (650–850 nm), iron oxides (850–900 nm), and 2:1 clay minerals (1400, 1900, and 2200 nm). After variable selection, model performance was heterogeneous and generally inferior to models calibrated with all wavelengths, with PCR outperforming the other regression methods and only the normalized model of the combined dataset achieving an excellent category with R² = 0.84, RMSE = 23.65 mg kg⁻¹, and RPD = 2.49.