Combining Machine Learning and Vis-NIR Spectroscopy to Estimate Nutrients in Fruit Tree Leaves

Traditional chemical analysis of plant tissue is time-consuming, costly, and poses risks due to exposure to toxic gases, highlighting the need for faster, low-cost, and safer alternatives. Vis-NIR spectroscopy, combined with machine learning, offers a promising method for estimating leaf nutrient levels without chemical reagents. This study evaluated the potential of Vis-NIR spectroscopy for nutrient estimation in leaf samples of banana (n = 363), mango (n = 239), and grapevine (n = 336) by applying spectral pre-processing techniques—smoothing (SMO) and first derivative Savitzky–Golay (SGD1d) alongside two machine learning methods: Partial Least Squares Regression (PLSR) and Random Forest (RF). Plant tissue samples were analyzed using sulfuric and nitroperchloric wet digestion and hyperspectral sensors. The prediction models were assessed using concordance correlation coefficient (CCC) and mean squared error (MSE). The highest accuracy (CCC > 0.80 and MSE < 2 g kg−1) was achieved for Ca in banana, P in mango, and N and Ca in grapevine across both machine learning methods and pre-processing techniques. The predictive models calibrated for ‘Grapevine’ exhibited the highest accuracy—characterized by higher CCC values and lower MSE values—when compared with the models developed for ‘Mango’ and ‘Banana’. Models using SMO and SGD1d showed better performance than those using raw spectra (RAW). The high amplitudes and variations in nutrient levels, combined with large standard deviations, negatively affected the predictive performance of the models.