A fast and efficient approach to estimate wild blueberry yield using machine learning with drone photography: Flight altitude, sampling method and model effects

Abstract

Precise yield estimation is important for phenotyping and crop management but is challenging for large-scale and high-density crops, such as wild blueberries. Machine learning (ML) technology and unmanned aerial vehicle (UAV) platforms created opportunities for more accurate and efficient yield estimation. RGB cameras are widely used in image processing due to their low cost, easy operation, and data analysis advantages. Here, we aimed to identify an effective, accurate, low-cost, and practical yield prediction method that can be applied to wild blueberry fields and potentially other high-density crops. Deep learning (DL) was overlooked due to its data-intensive training requirements, and the acquisition of image and yield data for wild blueberries is time-consuming and difficult. We evaluated the feasibility of using RF (Random Forest) and XGBoost (Extreme Gradient Boosting) models based on color and texture feature data obtained from UAV RGB images to predict wild blueberry yield, and tested the effect of flight altitude, sampling method, and ML model. We found that the XGBoost model trained on a systematically sampled dataset built from high-altitude color and texture features achieved the best prediction performance (R2 = 0.89, RMSE = 542.01, and MAE = 379.94). Interestingly, images captured at a higher altitude (30 m) performed better compared to that at lower altitudes (5 and 15 m), which provides clear evidence supporting higher and more efficient UAV cruise operations. Our study provides a non-destructive, efficient, fast, and easy-to-use yield prediction method for large-scale and high-density crops, confirming its feasibility with wild blueberries.