Potential Project Ideas

These are some suggestions for potential topics to help you select your end of year assignments.

1. Use IRIS for surface classification

   • Project Overview: This project focuses on the best use of IRIS to generate training data for a wide range of applications.

   • Data Acquisition and Preparation: Students will leverage the skills they developed in weeks 1 and 2 to create ingest satellite imagery within the IRIS framework to help them create classified imagery in a semi-supervised manner.

     • Methodology:

       • For a supervised approach, students are encouraged to label their own dataset using tools like IRIS.

       • The goal is to demonstrate the usefulness of IRIS for surface classification on a new type of application and/or satellite imagery (urban setting with Sentinel1 or Sentinel2, cloud detection using Sentinel3, lower latitude examples…).

       • In addition to using IRIS the student would need to develop a notebook describing their masked images in the context of the input features and compare with other unsupervised or supervised ML techniques.

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • References: IRIS Github Repo

2. Classification of Land Cover and Land Use

   • Project Overview: Students will engage in a classification task to differentiate various land covers (e.g., water, forest, urban areas) using satellite imagery.

   • Data Source: The project will primarily utilise datasets from Sentinel-2 or Sentinel-3 OLCI but you could try other satellite imagery (i.e. microwave from Sentinel-1 -> see harder project ideas).

     • Methodology:

       • Students have the option to manually label images using IRIS for a supervised learning or explore unsupervised classification methods as an alternative approach.

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • References: Land Cover and Land Use Classification

3. Classification of Cloud coverage

   • Project Overview: Students will engage in a classification task to differentiate various cloud detection using satellite imagery.

   • Data Source: The project will primarily utilise datasets from Sentinel-2 or Sentinel-3 OLCI but could expand to using another dataset (i.e. MISR -> see harder project ideas).

     • Methodology:

       • Students have the option to manually label images using IRIS for a supervised learning or explore unsupervised classification methods as an alternative approach.

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • References: Cloud Cover Detection Challenge

4. Regression analysis with satellite synergy

   • Project Overview: This project focuses on applying regression techniques to variables such as lead fraction or melt ponds.

   • Data Acquisition and Preparation: Students will leverage the skills they developed in weeks 5 and 6 to create collocated datasets from Sentinel-2 and Sentinel-3 OLCI images to build a regression model of some key sea ice surface characteristics and try and test their model in a different region and time (i.e. summer).

     • Methodology:

       • For a supervised approach, students are encouraged to label their own dataset using tools like IRIS, or they may opt for unsupervised techniques for initial data classification for Sentinel-2. And the classification results can be turned into target quantities to train the model that uses Sentinel-3 OLCI as the input features.

       • The goal is to apply regression methods on the S3 OLCI data with the quantities from Sentinel-2.

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • Reference: See week 5&6 for reference

5. Interpolation: Along-Track Interpolation

   • Project Overview: This project involves performing along-track interpolation on a variety of environmental variables such as SLA (that we looked at in class) or elevation over sea ice floes or a combination of sea level over floes and over leads to derive the radar freeboard.

   • Data Source: GPOD ready for use or Sentinel-3 official products.

     • Methodology:

       • The project involves direct acquisition or derivation of desired quantities from available GPOD or official ESA product variables, aiming to apply interpolation techniques on that.

       • Possibly with different ML methods and comparisons of performance.

     • Data access: GPOD or Copernicus Dataspace

     • References: GPSat interpolation

6. Interpolation: Gridded Interpolation

   • Project Overview: This project involves performing along-track interpolation on a variety of environmental variables (SLA or freeboard).

   • Data Source: GPOD or ESA official product.

     • Methodology:

       • The project involves direct acquisition or derivation of desired quantities from available variables, aiming to apply interpolation techniques on that and assess the potential of the method to detect previously unresolved features in space (i.e. Eddies) or in time (i.e. freeboard synoptic scale variability).

       • Possibly with different ML methods and comparisons of performance.

     • Data access: GPOD or Copernicus Dataspace

     • References: GPSat interpolation

7. Explainable AI

   • Project Overview: This projects investigates the use of XAI (Explainable AI) for model improvement by selecting key features that allow for a better classification or regression.

   • Data Source: Satellite multi-frequency optical imagery.

     • Methodology:

       • You will leverage what you learned in week 8 and attempt to use the selected features to re-train the classification or regression algorithms you saw in class for better performance.

       • Quantitatively assess your new model vs your old model against validation data.

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • References: See week 9 for example

And if you feel adventurous, here are some more ambitious projects:

8. IRIS for citizen science type classification

   • Project Overview: We have seen in class that IRIS outperforms even the best AI model (Weibin’s ViT). If thousands of students were to classify thousands of images using IRIS then the combined training dataset could outperform any one automatic AI method.

   • Data Source: The project will primarily utilise datasets from Sentinel-2 or Sentinel-3 OLCI but could expand to using another dataset.

     • Methodology:

       • In addition to using IRIS, the student would need to develop a notebook describing their masked images in the context of the input features and compare with other unsupervised or supervised ML techniques.

       • The student would need to investigate possible ways to turn IRIS into a citizen science method open it to several users.

       • Also, the students would need to show how several training datasets can be merged from different single users to contribute to a master training dataset (i.e., covering many different images in space and time).

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • References: IRIS Github Repo

9. Image alignment and/or sea ice motion

   • Project Overview: This project focuses on applying the image correlation code on pairs of images to align them and deduce the offset between them. Expand the work done in class with pairs separated by several hours and modify the algorithm to make it more robust.

   • Data Acquisition and Preparation: Students will leverage the skills they developed in weeks 5 and 6 to create aligned and collocated datasets from Sentinel-2 and Sentinel-3 OLCI images.

     • Methodology:

       • Adapt the code shared and test its performance.

       • Compare with other approaches (pattern matching, feature tracking).

       • Test code with pair of images at different times to infer the sea ice motion (Sentinel-1 imagery could be useful).

     • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

     • Reference: Nansen Centre Sea ice drift

10. Track and image alignment

    • Project Overview: This project attempts to address a recurring issue when comparing satellite altimetry tracks with imagery in that they are never exactly collocated. The goal here is to compare features on the tracks and images and attempt a collocation.

    • Data Acquisition and Preparation: Students will leverage the skills they learned in week 5 with the alignment code and try and find ways to adapt it to 1D vs 2D data.

      • Methodology:

        • Perform altimetry track and satellite imagery alignment.

        • Use examples combining altimetry (i.e., ICESat-2 or CryoSat-2) and imagery (i.e., Sentinel-2).

      • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

      • Reference: Collocation Tutorial

11. Explainable AI (spatial features)

    • Project Overview: This projects investigates the use of XAI for model improvement by selecting key features (in space) that allow for a better classification or regression.

    • Data Source: Satellite multi-frequency optical imagery.

      • Methodology:

        • You will leverage what you learned in week 8 and attempt to use the selected features in space to re-train the classification or regression algorithms you saw in class for better performance.

        • Specifically, you will test XAI for NN or ViT models for the 3x3 or expand to a larger input dimension (11x11 or 33x33) and see what input features contribute the most.

        • Quantitatively assess your new model vs your old model against validation data.

      • Data access: See week 3 for fetching satellite imagery from Google EE or Copernicus Dataspace

      • References: See week 2 for example