zcash/DEM-SuperRes-Model

DEM Super-Resolution

This repository contains a pipeline for generating synthetic high-resolution Digital Elevation Models (DEMs) by super-resolving 30m SRTM data to 10m resolution, fused with Sentinel-2 imagery. The model is trained on high-resolution LiDAR DEM data from McKinley Mine, NM, and applied to generate DEMs for Marrakech, Morocco.

Overview

The implementation uses an adapted DeepDEM model with a U-Net architecture (ResNet34 encoder) that takes 7 input channels:

• SRTM DEM (30m)
• Sentinel-2 RGB bands (10m)
• Sentinel-2 NIR band (10m)
• NDVI
• Nodata mask

The model predicts residual corrections to be added to a smoothed SRTM trend, producing 10m synthetic DEMs.

Requirements

• Python 3.8+
• PyTorch 2.0+
• PyTorch Lightning
• Segmentation Models PyTorch
• Rasterio
• Geopandas
• Albumentations
• Earth Engine API (for data acquisition)
• GDAL
• Boto3 (for LiDAR data download)

Installation

1. Clone the repository:

   git clone https://github.com/nfl0/DEM_SuperRes.git
   cd DEM_SuperRes
   

2. Install dependencies:

   pip install torch torchvision torchaudio pytorch-lightning segmentation-models-pytorch rasterio geopandas albumentations scipy gdown earthengine-api boto3
   

3. Install system dependencies:

   apt-get install libspatialindex-dev libgdal-dev gdal-bin
   pip install gdal
   

Data Acquisition

The notebook handles data acquisition from:

• SRTM 30m DEM: CGIAR/SRTM90_V4 via Google Earth Engine
• Sentinel-2 10m imagery: COPERNICUS/S2_SR_HARMONIZED via Google Earth Engine
• High-resolution LiDAR DEM: OpenTopography (McKinley Mine, NM)

Authenticate with Google Earth Engine and ensure access to required datasets.

Usage

1. Open DEM_SuperRes.ipynb in Google Colab or Jupyter.

2. Run cells sequentially to:

   • Acquire and preprocess training data (McKinley)
   • Train the model
   • Acquire and preprocess inference data (Marrakech)
   • Generate synthetic DEM
   • Run validation checks

3. Key outputs:

   • Trained model: Models/deepdem_model.ckpt
   • Synthetic DEM: synth_dem_marrakech.tif

Model Training

• Architecture: U-Net with ResNet34 encoder, 7 input channels, 1 output channel (residuals)
• Loss: L1 loss
• Optimizer: Adam (lr=1e-4)
• Training: 5 epochs on random crops from McKinley DEM
• Data Augmentation: Random crops, rotations, flips, noise

Validation

The notebook includes checks for:

• Input data statistics and validity
• Training fit (MAE/RMSE on validation crops)
• Output alignment and correlation with SRTM trend