How do we train deep learning models in scarce-label environments?
Suggested elective courses: Advanced image analysis
Deep learning algorithms have gained increasing popularity in remote sensing due to their accuracy. Yet, one of the main challenges of these algorithms is their application in scarce-label environments. One of the solutions to this challenge is the implementation of advanced transfer learning methods. Transfer learning is the strategy that starts by training a classifier on a different study area (or the same area but a different time) called source domain and fine-tuning it on a target domain. Thus, the challenge is to build a predictive model from a rich set of labeled samples available in a particular area and, successively, adapt this model to another area, with a limited set of labeled samples.
The main objective of this work is to implement and test advanced transfer learning methods for the identification of target classes from areas where we do not have access to a large number of labeled samples. Examples of transfer learning methods include (but are not restricted to) few-shot learning. These methods can be applied to different environmental or societal problems: agriculture or urban areas mapping and monitoring (just to name a few). Depending on the selected application domain and input satellite images (single date vs time series), various deep learning networks -Convolutional Neural Networks (CNN), Fully Convolutional Neural Network (FCNN), Long Short-term Memory (LSTM) or Transformers- can be used.
Sun, X., Wang, B., Wang, Z., Li, H., Li, H., & Fu, K. (2021). Research Progress on Few-Shot Learning for Remote Sensing Image Interpretation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 2387-2402
Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., & He, Q. (2020). A comprehensive survey on transfer learning. Proceedings of the IEEE, 109, 43-76