Deep Neural Networks for Remote Sensing Image Classification

Nowadays, Artificial Neural Networks are used along with remote sensing technology in order to perform land cover classification tasks which are very useful for change detection and monitoring studies of hydro-geomorphological high risk areas on the Earth surface. Following this framework, several Convolutional Neural Networks (CNNs) have been trained to test an original dataset of images acquired by UAV missions along the Basento River (in Basilicata region, located in the southern Italy). The dataset is made of more than 3000 aerial images which have been divided in classes and downgraded firstly to 80 cm/pixel and then to 20 cm/pixel in order to be comparable with the spatial resolution of the images that are supposed to be acquired by a Very Low Earth Orbit satellite that has been designed in the context of CLOSE - Close to the Earth project. The data used have been gathered in the context of the RPASInAir project which aims to enable innovative services with the purpose of land monitoring through the employment of data collected by Remotely Piloted Aircraft Systems (RPAS). A comparison of the performance of different CNNs will be shown and results will be given in terms of model accuracy and loss. All the results have been derived exploiting the resources belonging to the ReCaS-Bari data center HPC cluster.