Estimation and Classification of Coffee Plant Water Potential Using Spectral Reflectance and Machine Learning Techniques

Abstract

Water potential is an important indicator used to study water relations in plants, as it reflects the level of hydration in their tissues. There are different numerical variables that describe plant properties and can be acquired from leaf reflectance. The objective of this study was to estimate water potential in coffee plants using spectral variables. For this, a range of wavelengths that provided analytical flexibility was used. After this, machine learning techniques were employed to build data-driven models. The dataset used presents spectral characteristics (wavelength) of coffee plants, collected through the CI-710 Mini-Leaf Spectrometer equipment and also the water potential of each coffee plant, measured by the Scholander Chamber equipment. The dataset was divided into two crop management groups: irrigated and rainfed. Four machine learning techniques were implemented: Multi-Layer Perceptron (MLP), Decision Tree, Random Forest and K-Nearest Neighbor (KNN). The implementation of machine learning techniques followed two distinct strategies: regression and classification. The results indicate that the decision tree-based model demonstrated superior performance under irrigated conditions for regression tasks. In contrast, the KNN technique achieved the best performance for classification. Under rainfed conditions, the MLP model outperformed the other techniques for regression, while the Random Forest method exhibited the highest accuracy in classification tasks. While no hardware prototype was developed, the machine learning-based methods presented here suggest a possible pathway toward future intelligent, user-friendly, and accessible sensing technologies for coffee plantations.

Code and data availability

The paper's Data Availability Statement explicitly states that the study's datasets (coffee leaf spectral reflectance and water potential measurements) and the MATLAB analysis codes are publicly available at the authors' UFLA repository, which is an allowed URL. This is a paper-specific, public, actionable asset.