Abordagens de aprendizado de máquina em dados espectrais para estimativa de elementos químicos no solo

Abstract

The need for more efficient, sustainable methods to analyze soil chemical components is driving the integration of innovative technologies in agriculture. This study evaluated the potential of different machine learning (ML) approaches for estimating Total Organic Carbon (TOC) and Total Nitrogen (N) in agricultural soils, using near-infrared spectroscopy data and considering transformations to improve the algorithms' predictive performance. A total of 235 soil samples from the UEPG Soil Fertility Laboratory were used, with spectral data in the 1,100–1,648 nm range, comprising 277 predictive attributes. The methodology included data preprocessing with logarithmic transformation, standardization (Z-score), normalization (Min-Max Scaler), and a Savitzky-Golay filter. Seven ML algorithms were used, and their performance was evaluated in two approaches: using all spectral attributes, and after applying the SPA (Successive Projections Algorithm) attribute selection technique. The results indicated that the models Partial Least Squares Regression (PLSR) and Orthogonal Matching Pursuit (OMP), which are robust to multicollinearity, presented superior performance, with R² of up to 0.9124 for N and 0.8156 for TOC, when using the complete database transformed by the successive combination of logarithmic transformation, standardization (Z-score), and Savitzky-Golay filters. The use of feature selection by the SPA method significantly improved the performance of Linear Regression (R² = 0.8975) for N estimation, but did not bring improvements when associated with the PLSR algorithm. The analysis of variable importance in the PLSR model identified that the spectral bands between 1,394 nm and 1,412 nm were the most relevant for N prediction, while for TOC, the most important bands were between 1,396 nm and 1,450 nm. It is concluded that: (i) the combination of NIR spectroscopy with ML, through the joint use of logarithmic transformation methods, standardization (Z-score) and Savitzky-Golay filters, and associated with the PLSR and OMP algorithms, consists of a promising methodology for the analysis of TOC and N content in the soil, (ii) it is possible to generate simple models based on linear regression for estimating N in the soil, when the SAP attribute selection method is previously used, (iii) the spectral range between 1,394 nm and 1,450 nm is of special interest for the prediction of N and TOC