Expand modelcard and readme

MODELCARD.md

@@ -21,6 +21,8 @@ SatCLIP is a model for contrastive pretraining of satellite image-location pairs
 
 SatCLIP includes an *image* and a *location* encoder. The image encoder processes multi-spectral satellite images of size `[height, width, 13]` into `[d]`-dimensional latent vectors. The location encoder processes location coordinates `[longitude, latitude]` into the same `[d]`-dimensional space. 
 
+SatCLIP is a model trained and tested for use in research projects. It is not intended for use in production environments.
+
 ### Downstream Use 
 
 The SatCLIP location encoder learns location characteristics, as captured by the satellite images, and can be deployed for downstream geospatial prediction tasks. Practically, this involves *querying* the location encoder for the `[d]`-dimensional vector embedding of all downstream locations and then using that embedding as predictor during downstream learning. In our paper, we show the useability of the learned location embeddings for predicting e.g. population density or biomes.
@@ -30,6 +32,7 @@ The SatCLIP location encoder learns location characteristics, as captured by the
 Potential use cases of SatCLIP which we did build the model for and did not test for include:
 * The SatCLIP image encoder can in theory be used for helping with satellite image localization. If this application interests you, we encourage you to check work focusing on this, e.g. [Cepeda et al. (2023)](https://arxiv.org/abs/2309.16020). 
 * Fine-grained geographic problems (i.e. problems constrained to small geographic areas or including many close locations) are out of scope for SatCLIP. SatCLIP location encoders are pretrained for global-scale use.
+* Any use outside of research projects is currently out of scope as we don't evaluate SatCLIP in production environments.
 
 ## Bias, Risks, and Limitations
 

README.md

@@ -1,6 +1,7 @@
 # 🛰️ SatCLIP - A Global, General-Purpose Geographic Location Encoder
 
 ![CLIP](/figures/satclip.png)
+*Overview of the pretraining and deployment pipeline for SatCLIP.*
 
 ## Approach
 
@@ -47,8 +48,15 @@ Now, to train **SatCLIP** models, set the paths correctly, adapt training config
 python clip/main.py
 ```
 
+### Use of the S2-100K dataset
+
+The S2-100K dataset is a dataset of 100,000 multi-spectral satellite images sampled from Sentinel-2 via the [Microsoft Planetary Computer](https://planetarycomputer.microsoft.com/). Copernicus Sentinel data is captured between Jan 1, 2021 and May 17, 2023. The dataset is sampled approximately uniformly over landmass and only includes images without cloud coverage. The dataset is available for research purposes only. If you use the dataset, please cite our paper. More information on the dataset can be found in our [paper](https://arxiv.org/abs/2311.17179).
+
 ## Pretrained Models
 
+![CLIP](/figures/globes.gif)
+*Visualization of embeddings obtained by different location encoders for locations around the globe.*
+
 We provide six pretrained SatCLIP models, trained with different vision encoders and spatial resolution hyperparameters $L$ (these indicate the number of Legendre polynomials used for spherical harmonics location encoding. Please refer to our paper for more details). The pretrained models can be downloaded as follows:
 * SatCLIP-ResNet18-L10: `wget https://satclip.z13.web.core.windows.net/satclip/satclip-resnet18-l10.ckpt` 
 * SatCLIP-ResNet18-L40: `wget https://satclip.z13.web.core.windows.net/satclip/satclip-resnet18-l40.ckpt` 
@@ -78,11 +86,13 @@ Examples of how to use SatCLIP for downstream tasks will be added shortly.
 ## Citation
 
 ```bibtex
-@article{klemmer2023satclip,
-  title={SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery},
-  author={Klemmer, Konstantin and Rolf, Esther and Robinson, Caleb and Mackey, Lester and Russwurm, Marc},
-  journal={TBA},
-  year={2023}
+@misc{klemmer2023satclip,
+      title={SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery}, 
+      author={Konstantin Klemmer and Esther Rolf and Caleb Robinson and Lester Mackey and Marc Rußwurm},
+      year={2023},
+      eprint={2311.17179},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
 }
 ```