Temporal deep learning for satellite-based sugarcane monitoring and cycle age estimation

Accurate monitoring of sugarcane growth stages and field age is essential for optimizing yield estimation and harvest logistics within the sugar–energy sector. This study proposes a temporal deep learning approach to infer crop age using multitemporal Sentinel-2 satellite imagery. Object-based masks were generated through NDVI thresholding, and monthly spectral, textural, morphological, and phenological features were extracted for the period from March to October in 2023 and 2024. Five temporal deep learning models—Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Temporal Shift Mixer (TSMixer), and Transformer—were trained and fine-tuned using the Optuna optimization framework. Model performance and spatial generalization were evaluated across multiple regions in São Paulo. The Transformer model achieved the highest cross-region precision (97.7\%), followed by TCN (96.6\%) and TSMixer (96.4\%). These results suggest that attention- and convolution-based architectures effectively capture crop phenological dynamics, enabling the accurate and scalable estimation of sugarcane age from satellite time series for operational agricultural management.