venkatesh-thiru/s2l8h-UNet-5depth-upsample

---

license: mit tags: - agriculture - remote sensing - earth observation - landsat - sentinel-2

Model Card for UNet-5depth-Up+Conv: venkatesh-thiru/s2l8h-UNet-5depth-upsample

Model Description

The UNet-6depth-upsample model is designed to harmonize Landsat-8 and Sentinel-2 satellite imagery by enhancing the spatial resolution of Landsat-8 images. This model takes in Landsat-8 multispectral images (Bottom of the Atmosphere (L2) Reflectances) and pan-chromatic images (Top of the Atmosphere (L1) Reflectances) and outputs images that match the spectral and spatial qualities of Sentinel-2 data.

Model Architecture

This model is a UNet architecture with 5 depth levels and utilizes upsampling combined with convolutional layers to achieve high-fidelity image enhancement. The depth and convolutional layers are fine-tuned to provide a robust transformation that ensures improved spatial resolution and spectral consistency with Sentinel-2 images.

Usage

from transformers import AutoModel
# Load the UNet-5depth-Up+Conv model
model = AutoModel.from_pretrained("venkatesh-thiru/s2l8h-UNet-5depth-upsample", trust_remote_code=True)
# Harmonize Landsat-8 images
l8up = model(l8MS, l8pan)

Where:

l8MS - Landsat Multispectral images (L2 Reflectances)

l8pan - Landsat Pan-Chromatic images (L1 Reflectances)

Applications

Water quality assessment

Urban planning

Climate monitoring

Disaster response

Infrastructure oversight

Agricultural surveillance

Limitations

While the model generalizes well to most regions of the world, minor limitations may occur in areas with significantly different spectral characteristics or extreme environmental conditions.

Reference

For more details, refer to the publication: 10.1016/j.isprsjprs.2024.04.026