Applying Super-Resolution to Sentinel 2 Imagery

Alexander Vu and Ben Gaskill

Description:

The goal of our project is to research, adapt, and fine-tune super-resolution deep learning frameworks to resample Sentinel-2 imagery from its native resolution of 10 meters up to 1.5 meters per pixel. Our main focus is on the WorldStrat super-resolution model, which we tested using multi-temporal stacks of input imagery for 6 selected sites across Zambia.

Please refer to for the full description of our project.

Visualization of Data and Results:

Please refer to the notebook for dynamic visualizations (as well as our . Below is a static visualization of our inputs and outputs.

---

---

Input 1: Sentinel 2 (10 meter resolution, Site 0)

---

Output 1: Super Resolution (1.5 meter resolution, Site 0, Per-Band Normalization)

A closer look:

Sentinel 2

Super Resolution

Planet

---

Input 2: Sentinel 2 (1.5 meter resolution, Site 1, Cross-Band Normalization)

Output 2: Super Resolution (1.5 meter resolution, Site 1, Cross-Band Normalization)

A closer look:

Sentinel 2

Super Resolution

Planet