Abstract
Double-cropping grape systems offer enhanced land productivity but face significant challenges from climate variability, particularly rain stress and pest outbreaks during critical phenological stages. Accurate phenological prediction is essential to synchronize management practices with crop development and improve ecological resilience. This study presents a novel deep learning framework that integrates MobileNet with an augmented version of Dream Optimizer [Augmented Dream Optimizer (ADO)] to model grape phenology using satellite-derived time series of Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and rainfall from the CropHarvest dataset. The model transforms temporal data into pseudo-images for efficient spatiotemporal feature extraction, achieving 93% classification accuracy and a 6.1-day mean absolute error in stage prediction. ADO enhances convergence and generalization by optimizing key hyperparameters through a hybrid metaheuristic search. The system further identifies high-risk periods for rain damage and pest infestation, enabling proactive interventions. The model statistically significantly outperforms machine learning and deep learning baselines (p< 0.01) across three different agroecological zones [Mediterranean (California, USA, and southern Europe), subtropical (South Australia), and temperate (central Europe)] through spatially stratified fivefold cross-validation on 3,000 held-out test samples of the CropHarvest dataset. This work demonstrates the potential of optimized lightweight neural networks in sustainable viticulture, providing a scalable tool for precision management in climate-resilient double-cropping systems.