Satellite image time series (SITS) data are increasingly being used for vegetation-related remote sensing applications. For example, rich phenological information derived from SITS is often used for crop type classification [1].

Machine Learning (ML) techniques are the key tools for SITS classification. However, the performance of these techniques is highly reliant on the availability of a sufficient amount of labeled training data and the quality of them. This is particularly challenging when we are dealing with cross-regional and large-scale SITS classification problems [2]. The lack of sufficient and well spatially distributed training data calls for domain adaptation methods that adapt a given ML model trained with training data from a region that is geographically different from the study area (region of interest) [3]. 

The spectral and temporal characteristics of the same vegetation (crop) can considerably differ among regions due to changes in local conditions, such as the soil, climate, and human-related factors. Estimation of the temporal shift among different regions is of particular importance for cross-regional crop type mapping [2]. This will help adapting an ML technique trained for a specific region to any unlabeled target region. In this work, we would like to first investigate if there are environmental indicators (i.e., exploratory variables from other data sources) that can be used to address the temporal shift problem. Then, we aim to define an automatic approach that allows the adaptation of an ML model trained in one area to another geographically distant.

We will use an open-access dataset for cross-region adaptation with SITS from four different regions in Europe. This dataset was released in 2021 and can be freely downloaded from Zenodo.

This MSc thesis topic is a collaboration with Dr. Claudia Paris at the Department of Natural Resources.