multi-view evaluation on semantic segmentation supported by deep neural networks for the Cerrado biome

Abstract

While the semantic segmentation task has long been studied by the remote sensing (RS) community, it is a fact that deep neural networks (DNNs) have drawn attention due to the great interest and success of deep learning in several application domains. Even if there are so many studies and experiments using DNNs for RS semantic segmentation, an in-depth multi-view evaluation considering not only different DNNs but also distinct types of images (optical, multispectral) and satellite sensors with diverse spatial resolutions is still missing. In this article, we present one of such an experimentation where we considered images of three different satellites, i.e. Landsat-8 (30 m of spatial resolution), Sentinel-2 (10 m of spatial resolution), China-Brazil Earth Resources-4A (CBERS-4A; 8 m of spatial resolution), three classical DNNs, i.e. U-Net, DeepLabV3+, PSPNet, and two types of images (optical (RGB) and multispectral). Our study area is the Brazilian Cerrado biome and the choices of our evaluation focused more on the state-of-the-practice. We performed a thorough investigation and results show that DNNs and spatial resolution of satellite sensors are more relevant than the types of images. This conclusion is interesting because, eventually, researchers may rely on images with less number of channels (optical), decreasing the computational effort during training the DNNs.