Sentinel-2 Revegetation Classifier for Reclaimed Mines

ISEF Category: Environmental Engineering

Subcategory: Land Reclamation  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A reclaimed mine can look green from a car window and still be unstable on the ground. Satellites can spot that difference before most people can. You can use those images to test whether early plant recovery really matches the land’s recovery. That gives you a real environmental problem, not just a map.

What Is It?

This project uses satellite images to tell whether a reclaimed mine site is recovering well. Sentinel-2 is a European Earth-observing satellite that takes regular pictures of land in several color bands, including bands humans cannot see. A random forest classifier is a machine learning model that combines many decision trees to sort sites into groups, such as healthy revegetation, weak revegetation, or bare ground.

Think of it like judging a lawn from far away. Your eye sees green, but the satellite sees details in plant color, moisture, and bare soil that your eye misses. If a site has strong vegetation cover, stable soil, and a consistent seasonal pattern, the model should label it differently from a site that still looks patchy or stressed.

The challenge is not just making a map. You also need to define what counts as success. That could mean canopy cover, vegetation index values, change over time, or agreement with field photos and site records. The best projects compare the model’s answer with something real on the ground.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear question with real data, and you do not need a wet lab. You can work with public satellite imagery, site maps, and simple field reference data or historical records. The project connects to mining restoration, erosion control, habitat recovery, and land management, so it has real value. You can also learn remote sensing, classification, validation, and spatial analysis, which are useful research skills in environmental science.

Research Questions

• How does the choice of vegetation index affect classification accuracy for early revegetation success?
• What is the effect of training the model on one Appalachian site versus multiple reclaimed sites?
• Does adding seasonal Sentinel-2 images improve the model’s ability to detect weak revegetation?
• To what extent do slope, aspect, and elevation improve prediction of revegetation success?
• Which random forest feature set best separates healthy recovery from partial failure?
• How does model performance change when you test it on a site the model has never seen before?
• What is the effect of using field photo labels versus satellite-based labels on classifier agreement?

Basic Materials

• Laptop with stable internet access.
• Google Earth Engine account.
• Spreadsheet software for organizing site labels.
• GPS-enabled phone for optional site reference photos.
• Free map sources for reclaimed mine boundaries and background layers.
• Basic knowledge of Python or JavaScript for Earth Engine scripts.
• External hard drive or cloud storage for image exports and backups.
• Notebook for tracking site IDs, label rules, and validation notes.

Advanced Materials

• University or school GIS workstations with enough memory for raster analysis.
• Field GPS unit for accurate ground reference points.
• Camera with consistent geotagging for validation photos.
• Survey-grade site data, if available, for vegetation or cover estimates.
• GIS software such as QGIS for map inspection and QA checks.
• Python environment with geospatial libraries for custom evaluation.
• Access to archived mine reclamation records or ecological monitoring data.
• Optional drone imagery for high-resolution validation of sample plots.

Software & Tools

• Google Earth Engine: Pulls Sentinel-2 imagery, builds image composites, and runs the random forest classifier.
• QGIS: Lets you inspect site boundaries, compare layers, and check whether labels match the landscape.
• Python: Helps you clean data, test model performance, and graph feature importance.
• ImageJ: Measures vegetation cover from field photos if you need image-based validation.
• Google Sheets: Tracks site labels, sampling notes, and error counts in one place.

Experiment Steps

1. Define what early-stage revegetation success means for your project and pick measurable labels that match it.
2. Choose a small set of reclaimed sites and build a clean training and test split that keeps nearby areas from leaking into both sets.
3. Select Sentinel-2 features that may separate recovery stages, such as spectral bands, vegetation indices, and topographic context.
4. Design a random forest workflow that lets you compare different feature sets, label rules, and sampling strategies.
5. Plan a validation method that checks model output against field photos, site reports, or independent visual interpretation.
6. Decide how you will report errors, class imbalance, and site-by-site differences so the results stay honest and useful.

Common Pitfalls

• Using one reclaimed site for both training and testing, which makes the model look better than it really is.
• Labeling sites from color alone, which misses thin vegetation, exposed soil, and seasonal change.
• Mixing images from different seasons without tracking date, which adds noise that can hide recovery patterns.
• Ignoring cloud, shadow, or haze contamination in Sentinel-2 scenes, which distorts the spectral signal.
• Skipping an independent validation set, which leaves you unable to tell whether the classifier works on unseen land.

What Makes This Competitive

A strong version of this project does more than make a pretty map. It tests whether the model still works on sites it has never seen, and it compares simple feature sets against smarter ones. You can raise the quality by using careful ground truth rules, class balance checks, and site-level validation instead of random pixel splits. If you also explain where the model fails, you show real scientific judgment.

Project Variations

• Use Landsat instead of Sentinel-2 to compare how image resolution changes revegetation detection.
• Focus on one mine watershed and test whether slope and drainage patterns predict recovery speed.
• Compare random forest with another classifier, such as support vector machines, to see which one handles sparse vegetation better.

Learn More

• Google Earth Engine Documentation: Search the official documentation for Sentinel-2 tutorials, supervised classification examples, and accuracy assessment guides.
• NASA Earthdata: Find background material on remote sensing, vegetation indices, and satellite data use for land cover monitoring.
• USGS Landsat Collection Resources: Search for free articles and tutorials on land change, reclamation monitoring, and image interpretation.
• NOAA Climate Data and Tools: Use this for climate context, precipitation patterns, and environmental background that may affect revegetation.
• QGIS Training Manual: Read the free manual for basic GIS layer handling, map inspection, and spatial analysis workflows.
• PubMed: Search for review articles on mine reclamation, revegetation monitoring, and remote sensing validation methods.