ResGhostU-Net: U-Net compacta para segmentação de eucalipto em imagens multiespectrais da Sentinel-2

Abstract

The mapping of eucalyptus using remote sensing images can be an inaccurate and laborious process, especially when considering the large-scale multitemporal analysis of images. To try to solve this problem, new machine learning approaches have been proposed. In this work, we propose a compact modified U-Net (ResGhostUNet) for the task of semantic segmentation of eucalyptus using Sentinel-2 satellite images. In addition to the simplified architecture that has a reduced number of filters and depth and downsampling convolutions, we introduce the Ghost Residual Block, which allows reducing the computational cost and increasing the training efficiency. This study uses a new dataset that contains images of eucalyptus plantations in different cities in the Brazilian Cerrado biome. The quantitative and qualitative results demonstrate that the proposed method is highly competitive with respect to popular semantic segmentation methods. The ablation study highlights the effectiveness of the proposed component of the method. Furthermore, it demonstrates that using at least four selected bands yields slightly better results compared to utilizing all 13 bands. The proposed method consistently outperforms popular semantic segmentation methods, being simpler in terms of design, lightweight in terms of parameters, and fast in terms of processing. Due to these characteristics, ResGhostU-Net is potentially applicable for large-scale eucalyptus mapping using open-access satellite imagery.