What a novel temporal multivariate attention network (TMANet) model can explain about vegetation phenology? A corn crop case

Timely and accurate detection of crop phenology is critical for understanding vegetation dynamics, managing agricultural practices, and addressing climate change impacts. Traditional approaches relying on empirical or curve-based models often fail to capture the complex interactions among biophysical drivers and struggle with generalization across spatial and temporal scales. In this study, we propose a novel deep learning architecture, Temporal Multivariate Attention Network (TMANet), integrating vegetation indices, ground observations, and climatic features to improve the prediction and interpretability of corn phenological stages. TMANet introduces a unique dual-attention mechanism that simultaneously captures temporal dependencies and multivariate feature importance, enabling interpretable, stage-specific predictions of corn growth phases. The model was applied to a 23-year dataset (2000–2022) covering three agricultural districts in Iowa, USA, an intensively cultivated corn-growing region, using MODIS-derived spectral indices (NDVI and LSWI), and heat accumulation indices (GDD and CHU), and validated with USDA ground-truth phenology data. Results demonstrate that TMANet outperforms four benchmark models (LSTM, GRU, Random Forest, and SVR), achieving up to 90 % accuracy in classifying phenological stages and reducing phenological transition prediction errors to an average RMSE < 4 days and MAE < 2 days. The model effectively mapped dominant features across growth stages: NDVI was most predictive during early vegetative phases, while GDD emerged as the strongest driver in later reproductive stages. The integrated attention layers provide transparent insights into stage-wise feature relevance, advancing explainability in phenological modelling. This study not only presents a robust and generalizable model for phenology detection but also highlights the potential of attention-guided architectures in ecological modelling. TMANet contributes to the development of scalable, interpretable, and data-driven phenological monitoring frameworks, with direct applications in climate-resilient precision agriculture and ecosystem management.