Portable NIR spectroscopy for soil property prediction and fertility assessment in coffee and cocoa systems

Abstract

The study evaluated the potential of portable Near-Infrared (NIR) spectroscopy to predict soil physicochemical properties in coffee and cocoa production systems in Amazonas, Peru. A total of 126 topsoil samples were initially collected using a stratified sampling design. A total of 125 surface soil samples were analyzed using a stratified sampling design. Soil properties were determined using standard laboratory reference methods, and spectral data were acquired with a NeoSpectra portable spectrometer operating in the 1350–2550 nm (FT-NIR) range. Predictive models were developed using Random Forest (RF), Support Vector Machine (SVM), Neural Networks (NN), and Partial Least Squares Regression (PLSR), and their performance was evaluated using nested five-fold cross-validation (CV). The predictive performance of portable NIR spectroscopy varied according to the soil property, production system, preprocessing technique, and machine learning algorithm. Coffee soils showed higher predictive performance than cocoa soils. The best results were obtained for clay (R²CV = 0.775; RPDCV = 2.276) using RF with first-derivative preprocessing, followed by phosphorus (R²CV = 0.772; RPDCV = 2.037) using PLSR with band-depth preprocessing. Soil pH and silt also showed satisfactory predictive performance in coffee soils. In cocoa soils, potassium exhibited the highest predictive capability (R²CV = 0.454; RPDCV = 1.378) using SVM with second-derivative preprocessing, although the overall model performance was lower than that observed for coffee soils. Portable NIR spectroscopy provided potentially useful estimates for selected soil properties, particularly clay, phosphorus, pH, and silt in coffee soils. However, its predictive performance was strongly dependent on the soil property, preprocessing strategy, machine learning algorithm, and production system.