Key points are not available for this paper at this time.
• UAV and machine learning model for non-destructive rice GMC quantification. • Feature selection improves GMC model’s domain-specific performance. • Reduces labor/time for GMC, optimizing harvest for profit. • Validated over five crop seasons, proving the model’s robustness in agri-tech. • Low-cost, efficient GMC assessment from 22 % to 38 % spatially. Amidst climate change, determining optimal rice harvest timing is increasingly complex. Traditional and common methods, based on farmers’ experience, impact rice’s purchase price and rely on labor-intensive, destructive moisture meters for Grain Moisture Content (GMC) assessment. Addressing these challenges, this study utilizes an Unmanned Aerial Vehicle (UAV) to collect multispectral imagery for GMC estimation. This research integrated Feature Conversion (FC), Feature Selection (FS), and Machine Learning (ML) to develop a non-destructive GMC quantitative estimation model, tested across five crop seasons in both experimental and practical fields. The ground truth data, ranging from 19.71 % to 43.82 % GMC, underscore its extensive applicability. The optimized FC + FS + ML procedure significantly improves, with the Mean Absolute Error (MAE) reduced to 1.15 %, lower than traditional methods (1.74 %). This reduction highlights the effectiveness of FC and FS in minimizing redundant features. Applied in large-scale farming, the model substantially reduces labor and time by eightfold. The paddy-based GMC mapping in this research evaluates GMC spatial distribution, assisting farmers in optimizing harvest schedules, and thereby enhancing both profit and crop quality. This approach promises stable income for farmers and a secure food supply, mitigating risks associated with extreme weather.
Yang et al. (Sun,) studied this question.