Smart Agriculture Insurance Based on Artificial Intelligence - Object Detection

Climate change and extreme weather events, such as hailstorms, increasingly threaten high-value crops like grapes, causing substantial yield losses and economic risks for farmers. Traditional damage assessment methods, typically based on manual field inspections, are time-consuming, subjective, and error-prone, leading to delays in compensation and inefficient risk management. To address these limitations, innovative approaches combining artificial intelligence (AI) and edge computing are required to ensure objective and timely evaluations of vineyard damage. This study developed and tested a low-cost hardware-software system that integrates edge computing and deep learning to automate grape detection and spatial variability mapping. Two state-of-the-art models (RT-DETR and YOLOv8) were trained on 8,931 labeled images compiled from three open-source datasets and a dedicated acquisition campaign. The best-performing model, YOLOv8, was deployed on a Raspberry Pi 5 equipped with dual cameras and GPS to acquire geo-referenced data in a commercial vineyard in central Italy. The resulting grape counts were analyzed using univariate geostatistics and ordinary kriging to generate prediction and standard deviation maps. YOLOv8 achieved the best performance, with precision = 0.89, recall = 0.78, and mAP50 = 0.87. During field testing, 671 data points were collected every four seconds, each containing an image, GPS coordinates, and AI-based grape inferences. The geostatistical analysis confirmed data symmetry (skewness = 0.78) and identified the spherical model as optimal (RMSE = 0.24). The resulting variability maps showed grape counts ranging from 1 to 12 and standard deviation values between 1 and 3. The proposed system demonstrates that low-cost edge devices can effectively support objective, rapid and reproducible vineyard damage assessments, offering valuable tools for climate insurance applications. In addition to damage evaluation, the system can also provide detailed spatial information on total grape distribution, supporting precision viticulture practices such as differential fertilization and optimized harvest planning. Future developments will focus on integrating multispectral and thermal imaging and incorporating distance sensors to estimate grape weight, further enhancing both agronomic and insurance-related decision-making.