AIDAID202

AID202 · Fall 2026

Course Project

A research-oriented project in pairs, worth 35% of the AID202 course grade.

Fall 2026 project theme: ML for Earth Observation. Projects may involve images, signals, time series, text, or combinations of these modalities in order to extract useful information from Earth observation data.

Overview

An integral part of the course is the class project, which gives students a chance to apply the algorithms discussed in class to a research-oriented problem. You may use a dataset available on the web or collect your own data. If you choose to collect your own data, remember that data collection is often more time-consuming than expected.

The complexity and ambition of the project topic will be taken into consideration during grading. A good project should ask a clear machine learning question, establish sensible baselines, and analyze the results rather than simply train a large model.

2026 theme: ML for Earth Observation

Earth observation (EO) is broader than satellite image classification. Modern observing systems produce optical and radar imagery, hyperspectral measurements, geospatial time series, environmental sensor data, maps, metadata, and sometimes associated textual information such as event reports, captions, or geographic descriptions. This year's theme therefore includes projects related to vision, signal processing, time-series modeling, text processing, and multimodal learning.

Your project does not need to invent a completely new model. A careful empirical study of an interesting EO problem is perfectly suitable, especially when it includes meaningful baselines, ablations, error analysis, or a study of generalization.

Example project directions

Possible directions include, but are not limited to, the following. You are welcome to propose your own variation if it fits the theme and is feasible within one semester.

1) Images and spatial data

  • Land-use / land-cover classification. Predict land cover categories from optical or multispectral satellite imagery.
  • Flood, wildfire, or burned-area mapping. Detect regions affected by natural hazards using pre/post-event imagery or multi-sensor observations.
  • Change detection. Identify meaningful changes between observations acquired at different times, such as urban growth, deforestation, or disaster damage.
  • Building, road, or object detection. Detect or segment objects and infrastructure in aerial or satellite images.

2) Signals and spectral data

  • Hyperspectral pixel or patch classification. Treat each pixel or patch as a high-dimensional spectral signal and compare PCA, feature selection, SVMs, random forests, shallow neural networks, or 1-D CNNs.
  • Spectral anomaly detection. Identify unusual spectral signatures, rare materials, or anomalous surface conditions using one-class or unsupervised methods.
  • SAR / radar-based learning. Study classification, flood detection, land monitoring, or feature extraction from SAR observations, including the role of speckle and preprocessing.
  • Classical ML versus learned features. Compare hand-crafted spectral, texture, or vegetation-index features with learned representations.

3) Time series and forecasting

  • Crop or vegetation analysis from satellite time series. Predict crop type, vegetation condition, or other agricultural variables from sequences of observations.
  • Phenology or seasonal pattern analysis. Use EO time series to model seasonal behavior or distinguish land-cover classes by temporal signatures.
  • Forecasting environmental variables. Predict NDVI, temperature, rainfall, soil moisture, air quality, or another geospatially relevant signal.
  • Temporal change and event detection. Detect abrupt changes or evolving patterns in time-series observations.

4) Text and language

  • Environmental event detection from text. Detect floods, wildfires, earthquakes, droughts, or other events from news, reports, or social media, and optionally connect them to geographic regions.
  • Remote sensing captioning or retrieval. Study image-text matching, caption generation, or text-based retrieval for EO imagery.
  • Text classification for EO applications. Categorize disaster reports, land-use descriptions, or environmental documents using classical NLP or modern text encoders.
  • Weak supervision from text. Explore whether text descriptions, tags, or reports can provide labels or auxiliary supervision for an EO task.

5) Multimodal EO

  • Optical + SAR fusion. Compare or combine optical, multispectral, and SAR observations, for example Sentinel-1 and Sentinel-2.
  • EO + tabular or environmental metadata. Combine imagery with weather, elevation, map information, or other structured features.
  • EO + text. Combine imagery with captions, geographic descriptions, or event reports.
  • Generalization across sensors or regions. Study how models transfer across countries, climates, seasons, sensors, or data sources.

Datasets and data sources

You may use an established benchmark or construct a manageable dataset from openly available sources. The following are only starting points.

General EO imagery

  • EuroSAT — Sentinel-2 land-use and land-cover classification.
  • BigEarthNet — a large-scale multi-label Sentinel-1 / Sentinel-2 benchmark.
  • SpaceNet — high-resolution imagery and labels for buildings, roads, floods, and multi-temporal urban analysis.
  • Copernicus Data Space Ecosystem — open access to Sentinel imagery and tools for exploring and downloading EO data.

Spectral and signal-oriented datasets

  • Indian Pines, Salinas, and Pavia University — classic hyperspectral benchmarks for pixel and patch classification.
  • TorchGeo benchmark datasets — a convenient catalog of EO datasets covering classification, regression, segmentation, change detection, and related tasks.

Time-series and environmental data

  • Sentinel or Landsat time series assembled for a chosen region and task.
  • Environmental sensor or weather data such as rainfall, temperature, soil moisture, or air-quality time series, optionally combined with EO data.

Text-oriented data

  • Disaster or environmental reports collected from public sources such as official bulletins, open news archives, or humanitarian data portals.
  • Remote sensing caption or description datasets if you would like to study image-text matching, retrieval, or captioning.

Do not choose a dataset only because it is large. Make sure its size, labels, task definition, and compute requirements are realistic for a one-semester project. Subsampling a large benchmark is acceptable if the experimental protocol remains meaningful.

Important experimental considerations

  • Avoid leakage. For EO imagery, random splits of nearby image patches can put nearly identical locations in training and test sets. For time series, random shuffling across time can also create leakage. Prefer official splits when available, or design splits by geographic region and/or time.
  • Respect the data modality. Clearly state which sensors, bands, temporal sampling, text sources, or auxiliary variables are used. Explain preprocessing, normalization, resampling, filtering, feature extraction, and any missing-data handling.
  • Use appropriate metrics. Accuracy alone may be misleading. Depending on the task, consider macro-F1, per-class precision/recall, IoU, Dice, AUROC, MAE/RMSE, ranking measures, or forecasting errors.
  • Include meaningful baselines. Whenever possible, compare against at least one simple method. Logistic regression, SVM, random forest, k-NN, linear regression, ridge regression, isolation forest, HMM-style baselines, or small neural networks can reveal whether a more complicated model is actually necessary.
  • Analyze errors and generalization. Go beyond a single number. Investigate which classes, regions, seasons, events, or language categories are difficult and why.
  • For multimodal projects, show what each modality adds. If you combine imagery with text, signals, or metadata, compare uni-modal and multi-modal versions whenever feasible.

Software and libraries

You are encouraged to use appropriate machine learning, deep learning, and geospatial tools. Useful starting points include:

Suggested reading and resources

Deliverables

  • Project proposal · 23 October
  • Project progress report · 8 December
  • Final project presentation · 22 December
  • Final report and code · 31 December

Progress and final reports should use the provided LaTeX template and be submitted electronically in PDF format. Late submissions will be penalized according to the policy announced during the semester.

Collaboration policy

Each project should normally be completed in pairs. Exceptions may be made depending on enrollment. Students without a team may be assigned to a project group.

Grading

  • Proposal: 3%
  • Progress report: 10%
  • Presentation: 10%
  • Final report and code: 12%

Project proposal

Each project group should submit a proposal that clearly identifies:

  • The research topic to be investigated.
  • The main challenges that will be faced.
  • The data that will be used, including the modality, source, labels, and proposed train/validation/test split.
  • The evaluation metric(s) and at least one intended baseline.
  • Any important decisions about spatial, temporal, or multimodal experimental design.
  • A short, contextualized list of related papers.

Progress report

The progress report should be 5 pages + references and describe the problem, related work, the methodology being used, the experimental setup, and any preliminary results. It should make clear how the project will be evaluated and which datasets and metrics will be used.

Project presentations

Each group will give a short in-class presentation, including questions and answers. A strong presentation should cover the problem and motivation, the key technical ideas, the experimental setup, the main results, and a candid discussion of strengths and weaknesses.

Final report

The final report should be 9 pages + references and structured as a research paper. A typical organization is:

  • Title and authors
  • Abstract
  • Introduction and motivation
  • Related work
  • Approach
  • Experimental setup and results
  • Conclusions and possible future work
  • References

For EO projects, the experimental section should make the data split reproducible and explicitly state the modalities, sensors, time ranges, preprocessing steps, and evaluation metrics used.