Satellite image super-resolution for forest localization

With the ever-increasing amount of satellite missions in orbit, Earth observation and remote sensing have advanced rapidly and are applied in a multitude of fields, such as ecosystem monitoring and natural disaster prevention. At the same time, more and more machine-learning architectures are developed, which attempt to map and detect pixel-scale changes on the Earth’s surface. Such an application is forest monitoring, which is deemed nowadays highly critical, considering the threat of deforestation on a global scale. However, the spatial resolution of available open-source satellite imagery is limited. One way to address this issue is to augment the remote sensing imagery resolution using Image Super Resolution Networks. The goal of this work is to perform the Semantic Segmentation task with Sentine1-2 satellite images in order to detect forest areas. Three Deep Neural Network architectures are trained and tested. Moreover, we use a Super Resolution Deep Network to augment the resolution of the images. We create two datasets, one with original Sentine1-2 images and one with super-resolved images. We train the three architectures on the two datasets independently and we test them on a test set which combines original and super-resolved images. We show that Swin Transformer produces the best results, while training on the super-resolved images enhances the modeling power of all architectures.