Termografia aérea e aprendizado de máquina na avaliação de coberturas de solo e predição da produtividade do cafeeiro

Abstract

Coffee farming is of high socioeconomic importance in Brazil and other developing countries, but faces growing challenges associated with climate change, requiring more efficient management strategies and advanced monitoring tools. This study evaluated the influence of different soil covers (polyethylene film, brachiaria grass, and exposed soil) and two Arabica coffee cultivars (Arara and Mundo Novo) on canopy thermal dynamics, as well as the use of artificial intelligence techniques for the automatic classification of flowering levels and their relationship with productivity. Thermal monitoring was performed using aerial thermography with remotely piloted aircraft (RPAs) in monthly campaigns throughout a complete phenological cycle between June 2024 and June 2025. The thermal images were processed in Pix4D Mapper® and QGIS®, and the data were statistically analyzed in R®, using time series, analysis of variance, and regression models. The brachiaria grass cover promoted greater thermal attenuation of the canopy, with reductions of up to 1.76 °C, in addition to significantly reducing the relationship between soil and canopy temperature, limiting the thermal increase to 0.38 °C for the Mundo Novo cultivar and 0.19 °C for Arara. During the flowering period, the Arara cultivar had canopies that were up to 1.1 °C cooler than Mundo Novo, highlighting physiological differences and differences in response to the microclimate. At the same time, RGB aerial images were used for the automatic classification of flowering levels (low, medium, and high) using Convolutional Neural Networks (CNNs). The dataset consisted of 1,593 segmented and standardized images, and four CNN architectures were evaluated, including a customized network and models based on transfer learning (VGG16, ResNet50, and MobileNetV2). The VGG16 architecture performed best, with an accuracy of 92.02% and a loss of 0.2563 in the test set. The extreme flowering classes were classified with greater accuracy, while the intermediate class showed greater confusion, associated with visual overlap between adjacent phenological stages. Agronomic validation demonstrated a positive correlation between the classified flowering levels and actual productivity (kg plant⁻¹), with r = 0.717 and R² = 0.515 considering the two cultivars together. The Arara cultivar showed a stronger association (r = 0.772; R² = 0.596), while Mundo Novo showed a weaker relationship (r = 0.663; R² = 0.439). The simple linear regression adjustment showed moderate error (RMSE = 1.03 kg plant⁻¹; MAE = 0.87 kg plant⁻¹). It can be concluded that aerial thermography is efficient for thermal monitoring of coffee trees, that brachiaria plays a fundamental role in modulating the microclimate of the canopy, and that CNNs are a robust tool for classifying flowering and supporting early productivity estimates, contributing to the sustainable management of coffee farming. Keywords: Coffee farming; Remote sensing; Aerial thermography; Convolutional neural networks.