The conservation of tropical forests is a current subject of social and ecological relevance due to their crucial role in the global ecosystem. Unfortunately, millions of hectares are deforested and degraded each year. Therefore, government or private initiatives are needed for monitoring tropical forests. In this sense, this work proposes a novel framework, which uses of distribution estimation algorithm (UMDA) to select spectral bands from Landsat-8 that yield a better representation of deforestation areas to guide a semantic segmentation architecture called DeepLabv3+. In performed experiments, it was possible to find several compositions that reach balanced accuracy superior to 90% in segment classification tasks. Furthermore, the best composition (651) found by UMDA algorithm fed the DeepLabv3+ architecture and surpassed in efficiency and effectiveness all compositions compared in this work.