Deep Learning for Semantic Segmentation of Remote Sensing Data

The present project aims to investigate different strategies to handle the insufficient availability of labeled training data, towards fully exploiting the potential of deep architectures for remote sensing image analysis.
Regarding the application domain, the project aims to employ the solutions to be developed in relevant, concrete problems:

(1) Crop mapping in tropical regions from temporal multisensor image sequences
Food security is a matter of global concern, and Brazil is one of the major food producers in the world. In this context, regular update of spatial and temporal information on agriculture activities is of paramount importance for food policy decision making. This application aims at automatically determining crop types and other vegetation classes from temporal series of multispectral optical and synthetic aperture radar (SAR) data.

(2) Land cover mapping in the Brazilian savannas
The Brazilian savannas, also known as “Cerrado”, comprise the second largest biome in the country and one of the world´s biodiversity hotspots. Vegetation mapping in the Cerrado is extremely important for studies related to carbon balance, nutrient cycling and water resources; selection of new areas for conservation; monitoring of protected areas; and understanding of the vegetation dynamics. This application aims at automatically mapping vegetation classes and geomorphological features from multitemporal series of multispectral and hyperspectral images.

(3) Update of topographic geospatial databases
Detailed and current topographic data are at the core of many land development and environmental monitoring issues, they also play a major role in assisted and autonomous driving. Even after years of research and development, reliable automatic update is still a difficult challenge. In this application, we deal with updating the corresponding databases using multispectral and multitemporal aerial and satellite images. The advantage of such an approach is that the database potentially provides a lot of training data, which is why deep learning carries a lot of promises for solving this task. However, some of them are outdated. Thus, image classification needs to be coupled with possibilities to either eliminate the outdated data as blunders or to properly model the transitions have occurred in the landscape.