Update README.md

README.md

@@ -2039,7 +2039,7 @@ size_categories:
 
 In this benchmark, we investigate the potential of deep learning with multiple modalities for high-resolution wildfire forecasting. Leveraging different data settings across two types of model architectures: CNN-based and ViT-based.
 
-- 📝 Paper on [ArXiv](TBC) <br>
+- 📝 Paper on [ArXiv](https://arxiv.org/abs/2506.08690) <br>
 - 💿 Dataset repository on [GitHub](https://github.com/eceo-epfl/CanadaFireSat-Data) <br>
 - 🤖 Model repository on [GitHub](https://github.com/eceo-epfl/CanadaFireSat-Model) & Weights on [Hugging Face](TBC)
 
@@ -2080,7 +2080,7 @@ We describe below the different sources necessary to build the CanadaFireSat ben
 
 ## 📷 Outputs
 
-### 📊 CanadaFireSat Dataset Statistics:
+### 📊 CanadaFireSat Dataset Statistics (Without Test Hard):
 
 | **Statistic**                          | **Value**                 |
 |----------------------------------------|---------------------------|
@@ -2156,7 +2156,12 @@ We describe below the different sources necessary to build the CanadaFireSat ben
 The paper is currently under review with a preprint available on ArXiv.
 
 ```
-TBC
+@article{porta2025canadafiresat,
+  title={CanadaFireSat: Toward high-resolution wildfire forecasting with multiple modalities},
+  author={Porta, Hugo and Dalsasso, Emanuele and McCarty, Jessica L and Tuia, Devis},
+  journal={arXiv preprint arXiv:2506.08690},
+  year={2025}
+}
 ```
 
 ## Contacts & Information