Classificação de cana-de-açúcar em imagens multitemporais do sentinel-2 utilizando modelos de aprendizado de máquina e aprendizado profundo

Abstract

This study investigates the classification of sugarcane crops from multitemporal images acquired by the Sentinel-2 sensor, with a focus on the analysis of temporal domain shift between different agricultural years. The methodology is based on an object-oriented approach, with automatic segmentation using NDVI masks and the extraction of multitemporal multispectral and phenological time series for the characterization of agricultural targets. Tabular models, temporal architectures (LSTM, GRU, TCN, and Transformer), convolutional models, and a hybrid model that integrates spatial and temporal information were evaluated. The evaluation was conducted under strict temporal validation, with training in 2021–2022, validation in 2023, and testing in 2024. The results indicate that tabular models are more sensitive to temporal domain shift, whereas temporal models present greater stability and generalization capacity, with emphasis on the Transformer. Convolutional models achieved competitive performance, but with higher computational cost and lower stability. The hybrid model presented the best overall performance, combining high accuracy and temporal robustness. The findings reinforce the importance of explicit temporal modeling for agricultural monitoring based on remote sensing.