Classificação de Monoculturas de Café via Redes Neurais com Redução de Dados utilizando Fusão Espectral de Imagens Sentinel

This work proposes an approach for fusing multispectral Sentinel-2 images and radar data (synthetic aperture radar) from Sentinel-1 to classify coffee-growing areas in Machado, Brazil, using universal network (U-Net) and long short-term memory (LSTM) neural networks. The image bands were combined using techniques such as averaging, principal component analysis (PCA), discrete wavelet transform (DWT), and spectral pyramid, aiming to reduce dimensionality while preserving relevant information. U-Net achieved the best results, with 92.3% accuracy and a Kappa index of 0.78 using PCA, highlighting the effectiveness of spectral fusion in agricultural mapping.