README.md

# Remote Sensing Dataset: Substation Dataset

## Description

The first dataset is curated by TransitionZero and sourced from publicly available data repositories, including OpenSreetMap (OSMF) and Copernicus Sentinel data. The dataset consists of Sentinel-2 images from 27k+ locations; the task is to segment power-substations, which appear in the majority of locations in the dataset. Most locations have 4-5 images taken at different timepoints (i.e., revisits) and each image is of dimension 228x228 pixels. Each image has 13 spectral bands and each band has been linearly interpolated to a spatial resolution of 10m. Lastly, there is one ground truth mask for each location.

### Key Features

- **Source:**  OpenSreetMap (OSMF) and Copernicus Sentinel data
- **Resolution:** 10m per pixel
- **Bands:** 13 Sentinel-2 Bands
- **Size:** Approximately 70GB

We utilize this dataset in this [project](https://arxiv.org/abs/2409.17363). In this work, we focus on an applied research question of relevance to climate change mitigation -- power substation segmentation -- that is representative of applied uses of pre-trained models more generally. Through extensive tests of different multi-temporal input schemes across diverse model architectures, we find that fusing representations from multiple revisits in the model latent space is superior to other methods of using revisits, including as a form of data augmentation. We also find that a SWIN Transformer-based architecture performs better than U-nets and ViT-based models. 

---
license: apache-2.0
---