ASHS 2025 Conference

Forecasting yield is a timely opportunity to make anticipated harvesting decisions on the grown crop and understand field variability. Such information is a remarkable contribution to the precision agriculture context. However, developing such an approach is challenging for perennial crops such as pecan. These crops present slight canopy changes, which often do not reflect the upcoming yield. Consequently, waiting for the harvest date is the only approach to obtain yield data. Conversely, the advent of image-based data and artificial intelligence techniques has proven their applicability in addressing this issue. Therefore, our objective was to analyze whether UAV multispectral images and AI-based data analysis are suitable for developing forecasting models for yield in pecan trees. Hence, we began collecting multispectral images approximately five months before the harvesting date. Each data collection date had an interval of fifteen days, totaling ten multispectral image sets. Subsequently, we processed the images to generate ten orthomosaics (one for each date). The orthomosaics were used to calculate numerous vegetation indices, texture data, and the canopy area to be used as inputs for the forecasting models. At the harvest date, we measured the yield of 78 individual plants across two pecan fields. Before developing the forecasting models, we performed a correlation analysis to better understand the relationship between the image data and yield. Afterward, we developed the forecasting models using machine learning algorithms, namely, multiple linear regression, decision tree, support vector machine, and random forest. The dataset was split into 70% (n = 55) for training and 30% (n = 23) for testing. The training dataset was used to train the forecasting models, while the testing dataset was used to assess the models’ effectiveness regarding precision (coefficient of determination, R²) and accuracy (mean absolute error, MAE; and root mean squared error, RMSE). All the models produced interesting results and could be implemented to forecast yield in pecan trees. However, random forest outperformed the others (high precision and accuracy) and, therefore, was the remaining model for this study. Forecasting yield in pecan trees presented increased effectiveness, improving the models’ performance early on and establishing higher accuracies closer to the harvesting date. We also performed a feature importance analysis, where predominantly the texture data contributed better to the models’ performance. Certainly, our findings are timely and support pecan growers and stakeholders in making better decisions for harvesting with anticipated and accurate yield data without waiting for the harvesting date.