| Student: | S. Mohammadi MSc |
|---|---|
| Timeline: | May 2020 - 1 May 2024 |
With the current availability of multiple Earth Observation satellites, the access to multi-temporal remote sensing images at high spatial and spectral resolution has greatly improved. In addition to advances in the technology used for remote sensing data collection, there have been important scientific and methodological developments to convert the vast quantities of multi-temporal remote sensing data into useful information. Most developments have been based on supervised machine learning that rely on training samples. Mainstream supervised machine learning methods, i.e. deep learning, Random Forest, Dynamic Time Warping, work efficiently when applied on a user-defined study area, but they fail when transferred to other geographic areas and/or across different time periods. This happens because (1) the target classes are highly variable and usually embedded in a heterogeneous and complex natural or anthropogenic landscape, (2) insufficient training samples are available to adequately represent this high level of variability.
This research focuses on developing innovative, efficient and transferable methods based on deep learning for crop mapping and monitoring from optical multi-temporal remote sensing data. The deep learning-based methods will be developed in a way that they can get the most out of the data that is at hand (labeled and unlabeled) to learn generalizable features to address the problem of scarcity of labeled training samples in the target study area.
Meet the team
