Multi-Task Spatio-Temporal Modeling: Towards Accurate and Robust Plant Growth Prediction in the Real World

Accurately forecasting plant growth is vital for precision agriculture and sustainable food production. While prior work has explored predictive models for growth traits and yield estimation, direct image-level prediction of plant development provides richer and more intuitive insights. We introduce the Classification-assisted Multi-task Prediction (CAMP) model for photorealistic plant growth forecasting. The proposed methodology integrates spatiotemporal modeling with environmental classification within a unified deep learning framework, thereby enhancing prediction robustness by incorporating physical context, such as irrigation conditions. To evaluate biological plausibility beyond traditional metrics, such as the structural similarity index measure and mean absolute error, we propose a pixel of interest-based bipolar regression method, which provides interpretable, physically grounded feedback on model performance. Experiments on the KOMATSUNA dataset and a customized Arabidopsis dataset demonstrate that CAMP achieves competitive accuracy with state-of-the-art baselines on conventional pixel-level metrics and substantially improves the plant-aware POI-MAE metric over the recurrent PredRNN baseline. These findings establish multi-task learning and physically informed evaluation as effective strategies for real-world plant growth modeling.