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

Authors: DE OCA, MONTES - University Of California; MAGNEY, TROY - University Of California; VOUGIOUKAS, STAVROS - University Of California; RACANO, DARIO - University Of California; TORRES-OROZCO, ALEJANDRO - University Of California; FENNIMORE, STEVEN - University Of California; Martin, Frank; EARLES, MASON - University Of California

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/5/2025
Publication Date: 6/11/2025
Citation: de Oca, M., Magney, T., Vougioukas, S., Racano, D., Torres-Orozco, A., Fennimore, S., Martin, F.N., Earles, M. 2025. Strawberry fruit yield forecasting using image-based time-series plant phenological stages sequences. Computers and Electronics in Agriculture. 237(B). Article 110516. https://doi.org/10.1016/j.compag.2025.110516.
DOI: https://doi.org/10.1016/j.compag.2025.110516

Interpretive Summary: This manuscript discusses the development of strawberry yield prediction models using high resolution georeferenced images collected from the bed top for estimating yield of strawberry. The objective it to provide the grower with an estimate of the current harvest yield as well as a forward looking projection of future yield over the next 2-3 weeks.

Technical Abstract: Yield forecast is crucial for growers, enabling efficient resource management and informed decision-making that leads to increased productivity and cost savings. Such decisions impact storage, product processing, and logistics operations, that heavily rely on accurate yield forecasts. To address this need, this work presents the development and testing of a reliable method for yield forecasting. The proposed methodology combines high-resolution object detection with a multi-sequence input forecasting model that accurately computes the yield for incoming harvests. The forecasting approach incorporates a physically-constrained forecasting model based on a Long Short-Term Memory (LSTM) network. This model dynamically applies weights to timeseries data that include counts for the various phenological stages: flower, green, small white, large white, and red. These phenological sequences are generated using detections from a YOLOv10s model. 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) sequences with the original counts and b) phenological weighted sequences. The results indicate that physically-constrained input sequences outperform untreated sequences, as shown by the improved R2 and Root Mean Square Error (RSME) metrics.