S2 Deep Resolve Using NAIP Imagery

This repository contains scripts for super-resolution of Sentinel-2 imagery using high-resolution NAIP (National Agriculture Imagery Program) aerial imagery. The pipeline leverages NAIP's sub-meter resolution to enhance Sentinel-2 data, with cloud masking, patch generation, and data augmentation for deep learning. Tested on Ubuntu with dependencies like geotile and GDAL.

Overview

• Data Preprocessing: Intersect Sentinel-2 and NAIP imagery in Google Earth Engine (GEE) with a 1-day temporal delta. Apply Sentinel-2 cloud masks to NAIP and export to a Google Cloud Storage bucket.

• Patch Generation: Create image patches using the geotile package.

• Data Augmentation: Augment patches for deep learning using PIL.

• Deep Learning: Train a super-resolution model (requires significant GPU resources).

• 

• 

• 

Prerequisites

System Requirements

• Operating System: Ubuntu (tested on 20.04)
• Hardware: GPU with substantial VRAM (e.g., NVIDIA GPUs with CUDA support)
• Storage: Sufficient space for NAIP and Sentinel-2 imagery (several GBs)
• Google Cloud: Access to a Google Cloud Storage bucket
• Google Earth Engine: Account and authentication for GEE

Software Dependencies

• Python: 3.8 or higher
• GDAL: Required for geotile
• Python Packages:
  • geotile: For patch generation
  • PIL (Pillow): For image augmentation
  • Others (e.g., numpy, tensorflow/pytorch for deep learning)

Installation

Step 1: Install GDAL

GDAL is required for geotile. Install on Ubuntu:

sudo add-apt-repository ppa:ubuntugis/ppa
sudo apt update
sudo apt install gdal-bin

Verify:

gdalinfo --version

More info: https://launchpad.net/~ubuntugis/+archive/ubuntu/ppa

Step 2: Install geotile

pip install geotile

More info: https://geotile.readthedocs.io/en/latest/pages/install.html

Step 3: Install Other Packages

pip install Pillow numpy

For deep learning:

pip install tensorflow
# OR
pip install torch torchvision

Step 4: Set Up Google Earth Engine

Authenticate GEE:

earthengine authenticate

Ensure access to a Google Cloud Storage bucket.

Usage

Script 1: Data Preprocessing

Intersects Sentinel-2 and NAIP imagery in GEE, applies cloud mask, and exports to bucket.

1. Configure GEE credentials and bucket details.
2. Run:

python preprocess.py

Input:

• NAIP imagery collection in GEE
• Sentinel-2 imagery collection in GEE
• Temporal delta: 1 day

Output:

• Cloud-masked NAIP imagery in bucket

Script 2: Patch Generation

Generates patches using geotile.

1. Ensure GDAL and geotile are installed.
2. Specify input imagery and output directory.
3. Run:

python generate_patches.py

Input:

• Preprocessed NAIP and Sentinel-2 imagery

Output:

• Patches in specified directory

Script 3: Data Augmentation

Augments patches using PIL.

1. Install Pillow.
2. Specify input patch directory and output directory.
3. Run:

python augment.py

Input:

• Patches from previous step

Output:

• Augmented patches in specified directory

Training

Configure model architecture and parameters in training script. Example:

see notebook

Ensure GPU drivers and CUDA are configured.

Notes

• NAIP Resolution: Sub-meter (<1m) resolution ideal for super-resolution.
• Cloud Masking: Sentinel-2 cloud mask ensures clean NAIP imagery.
• GPU Requirements: Training requires significant GPU power.