The cloud and cloud-shadow filters of Landsat and Sentinel-2 satellite images (Level-2A) are still far from perfect. However, for many applications (agriculture, vegetation growth) a 100% correct cloud/shadow filter procedure is essential. The trick is that a proper cloud/shadow detection method should mask all clouds and shadows, but should not mask abrupt changes in the landscape like, ploughing or grass mowing.
For the Groenmonitor.nl all Sentinel-2 and Landsat-8 images are processed to a NDVI green index. Cloud/shadow masking is an essential part of the processing chain. Currently it is carried out on the basis of a decision tree model. However, for a 100% correct cloud/shadow screening a manual check and adjustment is still necessary. This is problematic in terms of scalability of the procedure worldwide and in terms of labour costs.
The objective of the thesis is to improve currently used cloud/shadow detection tools or develop a new method for this purpose. Subsequently, the improved method is to be applied on Sentinel-2 and Landsat-8 images in the Groenmonitor.nl.
Different strategies to develop or improve the cloud/shadow detection method can be followed. The currently existing atmospheric correction models SEN2COR and ATCOR can be adjusted/improved. The currently available method used in the Groenmonitor.nl can be studied and improved (for example by establishing and using a set of historic cloud-free reference images). Or a new method can be developed, for instance on the use of a random forest decision tree model. Or another type of machine learning technique may be tested.
Develop a cloud/shadow detection algorithm or method, based on (one of) the following principles:
- Improving the currently in the groenmonitor.nl used decision tree model with manual check
- Improve one of the existing atmospheric correction models (Sen2cor, Atcor), which provide also cloud and shadow masks
- Develop a random forest decision tree model
- Develop another method
Apply the improved or developed method on Sentinel-2 and Landsat-8 images in the Groenmonitor.nl and assess the accuracy of the developed method.
- Sen2cor model for atmospheric correction of Sentinel-2 images
- Atcor model for atmospheric correction of Landsat-8 (and other) images
Theme(s): Sensing & measuring, Modelling & visualisation