Estimação e classificação do potencial hídrico de cafeeiros utilizando reflectância espectral e técnicas de inteligência computacional

Abstract

The water potential stands as one of the principal indicators of plant water conditions, and therefore, it is extensively used in agricultural production studies. Its direct measurement involves the use of a device called a Scholander pressure bomb, which necessitates a complex and time-consuming procedure. However, literature showcases several studies that correlate water potential with spectral curves of leaves and plant canopies. This work aims to investigate the spectral curves of coffee plant leaves exhibiting different water potentials by utilizing computational intelligence tools and pattern recognition, leveraging spectral samples from coffee plantations cultivated in regions experiencing climatological water deficits. Two databases were employed: the first comprised frequency spectrum and reflected energy data of the coffee plants, while the second contained the water potentials of each plant. The collected samples correspond to coffee plants from Diamantina, Minas Gerais, Brazil, subjected to pre-processing techniques to structure the data. Four machine learning techniques were developed: Multilayer Perceptron Artificial Neural Network (MLP), Decision Tree, Random Forest, and K-Nearest Neighbor (KNN). Two distinct methods—regression and classification—were implemented for the four techniques. To enhance the performance of the machine learning methods, the SMOTE algorithm was executed, generating synthetic samples. The results indicate that Decision Tree outperformed in the regression method, achieving a Root Mean Squared Error (RMSE) of 0.4342 and a Coefficient of Determination (R2) of 0.6993. For classification, the artificial neural networks attained an overall accuracy of 62,05%. The outcomes derived from these methodologies were positive, given that the techniques employed effectively estimated water potential through spectral curves and spectral values.