Redes neurais convolucionais e previsão climática no monitoramento e manejo da cercosporiose em culuturas de pimenta

Abstract

Cercospora leaf spot, caused by the fungus Cercospora capsici, represents a major phytosanitary challenge for the cultivation of malagueta pepper (Capsicum frutescens), leading to yield losses and increased use of chemical pesticides. This thesis aimed to promote the use of convolutional neural networks (CNNs) integrated with climate forecast data for the automatic detection, severity quantification, and support of sustainable management of Cercospora leaf spot. The research was conducted in the municipality of Lagarto, state of Sergipe, Brazil (coordinates: 10.91°S, 37.66°W), and was structured into four complementary studies encompassing technical, technological, and social dimensions. The first study consisted of a systematic literature review, analyzing 64 studies addressing the use of CNNs for automatic assessment of plant disease severity. The second study focused on the development, comparison, and validation of eight customized models based on the Mask R-CNN and YOLOv8 architectures, using an original dataset composed of 1,645 leaf images containing 6,282 annotated lesions collected under real field conditions. This dataset was made publicly available to the scientific community. Mask R-CNN-based models achieved superior performance in pixel-level metrics, with the mask_rcnn_R101_FPN_3x model reaching an MIoU of 86.0% and an F1-score of 92.4%. In contrast, YOLOv8 models outperformed in disease severity level classification, with the YOLOv8n-Seg model achieving a precision of 95.2% and an F1-score of 95.1%. The third study integrated the segmentation models with climate data obtained via API into a mobile decision support system named FitoView, implemented as a Progressive Web Application (PWA) on the Google Cloud Platform. Validation of the OpenMeteo API revealed high correlations with INMET data for temperature (Pearson correlation coefficient r between 0.90 and 0.97), relative humidity (r between 0.81 and 0.95), and solar radiation (r > 0.90) across four municipalities in the state of Sergipe. The system combines CNN-based image analysis with 48-hour weather forecasts, generating contextualized recommendations through a decision matrix that considers detected disease severity, climatic risk, and suitable spraying conditions. The fourth study investigated challenges and opportunities for implementing Agriculture 4.0 technologies among family farmers in the municipality of Lagarto, using participatory approaches that included the presentation of the FitoView system, interviews on local radio, and the development of educational materials. The adopted methodology was based on dialogical processes of engagement and dissemination, enabling the identification of structural limitations—such as connectivity constraints and the need for digital skills training—as well as significant opportunities, particularly the widespread adoption of smartphones. In summary, the results of the four studies demonstrate that the integration of artificial intelligence and climate data can effectively support plant disease control and promote more sustainable management practices, with the potential to reduce pesticide use, especially when developed in alignment with the realities, needs, and capacities of local family farming systems.