AIoT-Driven Pest Monitoring Approach for Real-Time Tuta absoluta Detection and Population Prediction in a Controlled Greenhouse Environment

South American tomato leaf miner Tuta absoluta (Meyrick) (Lepidoptera: Gelechiidae) is responsible for significant biological invasions of tomato crops, perturbing global food production. The larvae feed within leaves, stems, and fruits, creating characteristic mines that reduce photosynthetic capacity by causing leaf necrosis and rendering fruits unmarketable. Synthetic pesticide applications used to suppress pest populations have detrimental impacts on the environment and human health and incur high production costs to farmers. These prevailing situations highlight the urgency of sustainable pest management practices. AI-driven pest monitoring systems assist farmers by enabling early pest detection and predicting expected pest outbreaks before their devastating attacks on crops. These data-driven insights facilitate farmers in implementing proactive pest control strategies, permitting reduced but impactful pesticide use through tracking pest population levels in the field. To develop smart pest monitoring systems to control T. absoluta outbreaks in tomato crops in greenhouse settings, this study highlights the use of AI-based pest detection and population prediction models to detect pest incidences and predict outbreaks in advance. A real-time image dataset was collected in greenhouse conditions to train various object detection models such as YOLOv10, YOLOv11, and YOLOv26 for the early detection of T. absoluta. The YOLOv26 model outperformed its counterpart approaches in terms of pest detection accuracy and inference speed by delivering 98.4% (mAP50) and 0.053 seconds, respectively. For early prediction of pest outbreaks, abiotic parameter data were collected through IoT sensors alongside pest biological data to generate early forecasts. Feature engineering and feature selection approaches were used for data preparation to model pest population dynamics. Various ML and ensemble learning approaches were implemented to model pest dynamics in relation to different meteorological parameters, where a blended ensemble learning technique resulted in the best prediction performance by delivering 4.99517, 62.415492, and 7.900348 values for MAE, MSE, and RMSE metrics, respectively.