Strawberry fruit yield forecasting using image-based time-series plant phenological stages sequences

Abstract

Yield forecasting is crucial for growers, enabling efficient resource management and informed decision-making. Such decisions impact storage, product processing, and logistics, leading to increased productivity and cost savings. However, this heavily relies on accurate yield forecasts. This work addresses such a need by presenting the development and testing of a reliable method for yield forecasting. The proposed methodology combines high-resolution object detection with a multi-variate input forecasting model that accurately computes the yield for incoming harvests. The forecasting approach incorporates a physically-constrained model based on a Long Short-Term Memory (LSTM) network. This model dynamically applies weights to the time-series data composed of counts for the phenological stages: flower, green, small white, large white, pink, and red (ripe fruit). These counts are obtained from detections made by a YOLOv10s, achieving an mAP@50 of 0.74 for all classes. As a result, the forecasting model’s capacity to interpret input data is enhanced, translating it into a valid ripe count forecast. To validate the proposed approach, the forecasting model was trained and evaluated using (a) untreated count sequences and (b) weighted count sequences. The results indicate that phenologically-weighted input sequences outperform untreated sequences, with the following evaluation metrics: R${}^2$ = 0.74, Root Mean Square Error (RMSE) = 12.67, Mean Absolute Error (MAE) = 10.95, and Mean Absolute Percentage Error (MAPE) = 39.4, improving 15%, 19.26%, 17.13%, and 11.3%, respectively.