Spaces:

---

ibm-esa-geospatial

/

challenge

like 36

Running

TerraEye – flexible land mapping

Link to current demo:
https://huggingface.co/spaces/PolslLandMappingDevs/TerraEye

TerraEye is an open-source project. It provides tools for mapping, spatial analysis,
and visualization of geospatial and satellite data. The system combines a classical
web application architecture with the TerraMind foundation model as an analytical layer,
enabling satellite image processing, thematic map generation, and spatial analysis
for environmental, planning, and research applications.

Figure 1: Menu and an interactive globe view used for global navigation and selection of the area of interest.

Figure 2: Zoomed menu and an interactive globe view used for global navigation and selection of the area of interest.

Why it’s interesting?

Geospatial mapping and analysis are essential for modern information systems,
including spatial planning, environmental monitoring, infrastructure change detection,
and land-use assessment. With increasing availability of satellite data, there is growing
demand for tools that not only visualize geospatial information but also automatically
interpret it. TerraEye offers a flexible, modular, open-source platform that leverages
foundation models to support advanced geospatial analysis, suitable for education,
research, and potential industrial use.

How It Works

The system is modular: the backend handles data processing and TerraMind integration,
while the frontend provides an interactive interface for maps and analysis results.
TerraMind performs multimodal satellite image analysis, generates land-use,
land-cover maps and supports spatial metric calculations.

Advanced use of TerraMind

TerraMind is the core analytical engine of the project. The model is trained for Earth
observation tasks, allowing it to process satellite data alongside traditional spectral
indices.

Core TerraMind functions

1. Multimodal satellite analysis
   The model processes satellite images and classifies pixels according to land
   cover and land use, such as forests, water, urban areas, and croplands.

2. Semantic map generation
   Model outputs are used to create LULC maps that support both visual
   interpretation and further spatial calculations.

3. Context-aware analysis
   By providing semantic labels for pixels, TerraMind enhances spectral analyses
   e.g., NDVI (vegetation), NDWI (water), NDBI (built-up areas) enabling
   interpretation of indices in the context of land type.

Spectral indices

Spectral indices are mathematical combinations of satellite spectral bands designed
to highlight specific landscape properties, such as vegetation health, soil brightness,
or water presence. In TerraEye, model outputs can be combined with spectral index
calculations to produce richer analytical layers.

Figure 3: Comparison of two TerraMind variants (e.g., Large and Tiny), illustrating the progression from RGB satellite imagery to TerraMind outputs enhanced with spectral indices.

Evaluation and Model Comparison

To assess different TerraMind variants and spectral calculations, the system uses
standard pixel-based metrics:

• Pixel Accuracy – proportion of „correctly” classified pixels
• Precision / Recall – quality of predictions for individual classes
• Dice Score – balanced measure for small or sparse classes
• Intersection over Union (IoU, mIoU, fwIoU) – overlap quality, including class
  weighting

Figure 4: evaluation metrics calculated between two sets of segmentation maps produced by different TerraMind variants (e.g., Large and Tiny), where one variant serves as the reference (“ground truth”) delineation.

These metrics allow comparison of model variants in terms of accuracy
and computational cost, critical for real-world scenarios and potential on-board
satellite processing.

Demo Features

The current demo offers:

• Interactive visualization of geospatial data
• Presentation of TerraMind-generated analysis results
• Basic spatial metric calculations
• Real-time communication between frontend and backend

Lessons Learned

The project provided practical experience in integrating AI models with GIS systems,
handling satellite data, and evaluating model outputs. Despite no prior experience in AI
or satellite analysis, the team quickly adapted, highlighting the value of modular design,
practical problem-solving, and interdisciplinary teamwork.

Team

▪ Authors: Jakub Bodzioch, Łukasz Wolf, Kajetan Hołdan, Jan Piechota,
Michał Figołuszka, Karol Bieżuński

▪ Affiliation: Silesian University of Technology, Poland

▪ Supervisor: Jakub Nalepa, PhD, DSc

▪ Contact: jb305900@student.polsl.pl