Prediction of disease severity in coffee plants using digital images and multivariate adaptive regression spline models

Assessing disease severity in coffee plants is an essential step for efficient phytosanitary management, especially in production systems aimed at sustainability and high productivity. Traditionally, this assessment is performed visually and subjectively, which can lead to inconsistencies and hinder decision-making. With the advancement of digital technologies and computational intelligence, it becomes possible to automate this process through image analysis, offering greater precision, reproducibility, and speed. Thus, the objective of this work was to present an approach based on Multivariate Adaptive Regression Splines (MARS) for predicting disease severity in coffee plants from digital images. Leaves were collected from the middle third of the plant. Images were captured by RGB cameras, 50 cm from the ground, under ambient temperature and light conditions. The background used was a blue ethyl vinyl acetate (EVA) sheet, measuring 600 x 400 mm. The image resolution was 1280 × 1024 pixels² and 96 dpi. A total of 2,181 images were collected showing different severities of rust (Hemileia vastatrix) and cercospora leaf spot (Cercospora coffeicola). Morphometric variables of the affected leaves, such as area, perimeter, diameter, and average radius, were obtained and used as predictors in the statistical models. MARS models were constructed with different degrees of interaction (product), ranging from 1 to 3, and with different numbers of terms, in order to evaluate the impact of model complexity on predictive performance. The models were evaluated using k-fold cross-validation, using the Root Mean Squared Error (RMSE) as a performance metric. In addition, an analysis of the importance of the variables was performed using different statistical criteria, such as the number of subsets, GCV, and RSS. The results indicate that models with interaction degrees of 1 and 2 exhibit stable and relatively low RMSE, even with an increasing number of terms, suggesting good generalization capacity and robustness. On the other hand, models with degree 3 showed high and constant RMSE, indicating overfitting and low predictive efficiency. This analysis allowed the identification of relevant attributes for predicting severity, highlighting variables such as average radius and area as the most influential. It is concluded that the proposed approach, combining digital images and MARS models, is effective for estimating disease severity in coffee plants and can be integrated into automated field monitoring systems.