Project Summary

• Development of Finding DEM0, a zero-shot framework for converting monocular depth maps into metrically accurate Digital Elevation Models (DEMs) without supervised training.
• Integration of a foundation model (DepthAnything V2) with global geodetic datasets (Copernicus DEM and GEDI LiDAR) through a lightweight calibration strategy.
• Demonstration of improved elevation accuracy (up to ~2 m in urban areas and ~1.5 m in vegetated regions) compared to baseline DEMs, enabling scalable and cost-effective terrain monitoring.

Development Tools

• DepthAnything V2: foundation model used for monocular depth estimation from high-resolution RGB imagery.
• Copernicus GLO-30 DEM: global elevation dataset used as a continuous reference for calibration.
• GEDI LiDAR L2A: spaceborne LiDAR data providing sparse but highly accurate elevation measurements.
• Python ecosystem: core implementation developed in Python, including numerical processing (NumPy) and data handling (GDAL, rasterio).
• Machine learning tools: lightweight regression models implemented for depth-to-elevation calibration.
• Geospatial datasets from IGN: high-resolution orthophotos and validation LiDAR data.

Development Outputs

• Generated datasets: high-resolution (0.5 m) calibrated DEMs over ~2,500 tiles across France.
• Experimental results and benchmarks: quantitative evaluation reports comparing baseline and calibrated DEMs.
• Scientific publication: ISPRS abstract and ongoing work toward full paper submission.

Here is a brief overview of the visual outputs:

Project Description

The project focuses on addressing a critical limitation in Earth Observation: the generation of accurate and up-to-date Digital Elevation Models (DEMs) at high spatial resolution. Traditional methods such as airborne LiDAR and photogrammetry, while highly precise, are expensive, time-consuming, and limited in temporal coverage. This creates a gap in applications requiring frequent updates, such as environmental monitoring, urban planning, and disaster management.

To overcome these limitations, this work introduces Finding DEM0, a novel zero-shot calibration framework. The core idea is to leverage recent advances in vision foundation models, specifically monocular depth estimation, and bridge the gap between relative depth predictions and absolute elevation measurements.

The workflow begins with high-resolution RGB orthophoto imagery (0.5 m resolution), from which depth maps are generated using the DepthAnything V2 model. These depth maps capture detailed geometric structures of the scene but lack a consistent metric scale. To address this, the proposed framework introduces a calibration step that aligns the predicted depth values with real-world elevation data.

Calibration is performed using two complementary geodetic data sources: the Copernicus GLO-30 DEM, which provides continuous global elevation coverage, and GEDI LiDAR data, which offers sparse but highly accurate elevation samples. The depth maps are divided into grid cells (30 m × 30 m), and within each cell, depth values are paired with corresponding elevation references. When GEDI data is available, it is used preferentially due to its higher vertical accuracy; otherwise, Copernicus DEM values are used.

These depth–elevation correspondences are used to train a simple linear regression model on a per-image basis. This lightweight approach avoids the need for large training datasets or computationally intensive retraining processes. Once calibrated, the model is applied pixel-wise to the entire depth map, producing a high-resolution DEM with absolute elevation values.

The framework was evaluated on approximately 2,500 image tiles covering diverse landscapes across France, including urban, agricultural, and forested areas. Validation against high- resolution LiDAR ground truth data demonstrated consistent improvements over baseline DEM products. In particular, the method significantly reduces elevation errors in complex environments such as cities and vegetated regions.

An additional advantage of the proposed approach is its computational efficiency. The system runs on standard CPU hardware and achieves near real-time performance, making it accessible to a wide range of users without requiring specialized infrastructure.

Overall, the collaboration with ESA Φ-lab resulted in a scalable, training-free solution for generating accurate DEMs from optical imagery alone. This opens the door to more frequent and cost-effective updates of elevation data, reducing reliance on traditional survey methods and enabling continuous global monitoring.