Internship — Deep Learning for TomoSAR

Host: EO Analytics, Salzburg University, Austria • Supervisor: Dr. Karima Hadj-Rabah

Internship Report: Forest Canopy and Ground Height Estimation using TomoSAR and Deep Learning Models

Intern: Sahar Mohamed
Period: July 2025 – September 2025

Lab & Context

The work used TomoSAR data (TropiSAR) for forest height estimation in Paracou, French Guiana. The aim was to estimate canopy and ground heights using deep learning, compare methods, and test generalization on unlabeled stacks.

Tasks

1. Training a Tomographic SAR Neural Network (TSNN)
   Trained a TSNN to estimate canopy and ground heights. Dataset split into training/validation/test; models evaluated on unlabeled data and tuned for performance.
2. Literature Review and Performance Comparison
   Benchmarked TSNN against recent state-of-the-art methods.
3. Development of Modified UNet (M-UNET)
   Designed M-UNET to estimate heights from patches (capturing spatial context) and tested generalization without lidar.
4. Comparison with Traditional Methods
   Compared DL models (TSNN, M-UNET) with spectral/tomographic methods (e.g., CAPON, MUSIC).

Learning & Notes

The internship deepened my understanding of TomoSAR data formats (covariance matrices, full-pol SLC) and practical DL model training for geophysical variables. Combining DL with traditional approaches improved estimation robustness.

Download Internship Report (PDF) Open GitHub Repo