Robust Training of Deep Neural Networks with Weakly Labelled Data
Gianmarco Perantoni, Lorenzo Bruzzone · 2024
The field of remote sensing has recently experienced a remarkable transformation, driven by the integration of deep learning (DL) techniques in many data analysis tasks. Deep neural networks (DNNs) have demonstrated effective in extracting com- plex patterns and valuable insights from a wide range of remote sensing data, in- cluding multi-spectral (MS), hyper-spectral (HS), and radar satellite imagery [1, 2]. The fusion of DL and remote sensing has resulted in new possibilities for applica- tions such as land-cover classification, object detection, change detection, and en- vironmental monitoring. However, beneath these remarkable achievements lies a pervasive challenge: the scarcity of labelled data for supervised learning tasks.