SISTEMAS INTELIGENTES EMBARCADOS EM DISPOSITIVOS MOVEIS E BASEADOS EM DEEP LEARNING PARA DETECÇÃO E MONITORAMENTO ESPACIAL DE PRAGAS AGRÍCOLAS

Abstract

Accurate and timely detection of insect pests remains one of the major challenges in modern agriculture, especially in large-scale soybean and maize production systems. Inefficient monitoring practices often result in delayed control interventions and significant yield losses. Recent advancements in deep learning and mobile computing have opened new opportunities for in-field pest identification using lightweight computer vision models. In this context, this thesis presents an integrated framework for intelligent pest detection and spatial monitoring based on deep learning, geostatistical analysis, and mobile applications. First, two datasets of insect pests were constructed and evaluated: a comprehensive high-resolution dataset curated through double-expert validation, and a smaller sample designed for comparative analysis. State-of-the-art detection architectures (YOLO and Detectron2) were trained on both datasets and subsequently converted into TensorFlow Lite (TFLite) and ONNX formats to enable deployment on resource-constrained devices. Even under the least favorable conditions using the reduced dataset and the lightest ONNX model the results reached a precision of 87.3% and accuracy 95.0%, demonstrating the robustness of the pipeline. Building upon these results, a mobile system named AgroInsect was developed. The application performs real-time, on-device detection of four key pest species relevant to Brazilian soybean and maize production (Diabrotica speciosa, Dalbulus maidis, Diceraeus spp., and Spodoptera frugiperda), automatically extracts geolocation metadata, validates spatial consistency based on field boundaries, and synchronizes detections with a cloud database. Spatial visualization is generated through heatmaps and Ordinary Kriging (PyKrige), enabling high-resolution incidence maps. Field evaluations confirmed strong model performance, with overall accuracy of 95.1%, F1-scores above 0.94 for all species, and only 1.1% false detections. The kriging model achieved R² > 0.94 under dense sampling, accurately reproducing ecological spatial patterns. Additionally, this thesis introduces AgroLabIA, a digital platform designed for the storage, annotation, and dissemination of agricultural pest datasets. It provides curated, multi-format datasets suitable for training machine learning models and supports the continuous expansion of new insect and weed classes. The integrated environment that encompasses dataset generation, mobile detection, spatial verification, and geostatistical mapping demonstrate a scalable and operationally robust solution for precision pest monitoring. The results position the AgroInsect database as an effective tool for accelerating decision-making in integrated pest management, particularly in regions with limited connectivity, thus contributing to the consolidation of Agriculture 4.0.