DailyWildfireCAN commited on
Commit
499bdac
·
verified ·
1 Parent(s): 27835ad

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +134 -13
README.md CHANGED
@@ -1,22 +1,143 @@
1
  ---
2
  license: cc-by-nc-4.0
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
4
 
5
- Checkpoints of models in DailyWildfireCAN benchmark.
6
 
7
- ![image](https://cdn-uploads.huggingface.co/production/uploads/69f8c2efff7c500b6101eddd/kZlcdvFOCwqimxEmgMZFW.png)
8
 
9
- Each model is evaluated across three independent runs, with the random seed derived from the run index. We provide the optimal weights as best_checkpoint.pt, organized using the directory structure modelName/checkpoints_modelName_runN/
10
- where N denotes the run index. For example the UNet checkpoints are provided under unet/checkpoints_unet_run1, unet/checkpoints_unet_run2, and unet/checkpoints_unet_run3.
11
- This strcuture applies to all evaluated models: unet_age, unet_attention, unet_convlstm, unet_segformer, UTAE
12
 
 
13
 
14
- ## Mono-temporal models:
15
 
16
- * Standard UNet(architecture: 'unet'): The baseline spatial U-Net model.
17
- * Age-Encoding UNet(architecture: 'unet_age'): A U-Net that explicitly encodes the satellite age (the time gap in days between the fire event and the satellite acquisition).
18
- * Attention UNet(architecture: 'unet_attention'): A U-Net utilizing attention gates in the skip connections to help the model focus on the most critical spatial features and suppress irrelevant background noise.
19
- * UNet-SegFormer(architecture: 'unet_segformer'): A hybrid vision-transformer architecture that replaces the standard CNN encoder with SegFormer's Mix Vision Transformer (MiT), paired with a standard U-Net decoder for heavy pixel-level accuracy.
20
- ## Multi-temporal models:
21
- * Spatiotemporal UNet(architecture: 'unet_convlstm'): A U-Net featuring aConvLSTMbottleneck for recurrent time-series processing (e.g., 3-day sliding window forecasting).
22
- * UT-AE(architecture: 'utae'): A temporal attention encoder-decoder baseline adapted from the ICCV 2021 U-TAE model for satellite image time series. This baseline uses the time-series offline samples fromTimeseries_Samples/, and the generator now stores sequence positions for the temporal attention encoder when you regenerate those samples.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: cc-by-nc-4.0
3
+ library_name: pytorch
4
+ pipeline_tag: image-segmentation
5
+ datasets:
6
+ - DailyWildfireCAN/DailyWildfireCAN
7
+ tags:
8
+ - wildfire
9
+ - wildfire-spread-prediction
10
+ - remote-sensing
11
+ - spatio-temporal
12
+ - benchmark
13
+ - canada
14
+ arxiv: 2611.14207
15
  ---
16
 
17
+ # 🍁 DailyWildfireCAN Benchmark: Model Checkpoints
18
 
19
+ [📄 Paper](https://arxiv.org/abs/2611.14207) · [🌐 Project page](https://dailywildfirecan.github.io) · [💻 Code](https://github.com/hagerradi/DailyWildFireCAN) · [🗂️ Dataset](https://huggingface.co/datasets/DailyWildfireCAN/DailyWildfireCAN)
20
 
21
+ This repository contains the trained checkpoints for the models in the **DailyWildfireCAN** benchmark for next-day wildfire spread prediction across Canada. Each model was trained three times with different random seeds, and all three runs are released.
 
 
22
 
23
+ ![DailyWildfireCAN benchmark](https://cdn-uploads.huggingface.co/production/uploads/69f8c2efff7c500b6101eddd/kZlcdvFOCwqimxEmgMZFW.png)
24
 
25
+ ## Task
26
 
27
+ Given the inputs available up to day *T*, predict which pixels will newly burn on day *T+1*.
28
+
29
+ - **Input:** a 256 × 256 tile at 90 m resolution with 50 channels (see the [dataset card](https://huggingface.co/datasets/DailyWildfireCAN/DailyWildfireCAN) for the channel order). Multi-temporal models take 3 consecutive days: (3, 50, 256, 256).
30
+ - **Output:** one logit per pixel, (1, 256, 256). Apply a sigmoid to get the probability of new burn.
31
+ - **Metrics:** AUC-PR, precision and recall, computed on the Test-Easy (unseen year), Test-Medium (unseen regions) and Test-Hard (unseen regions and year) splits.
32
+
33
+ ## Models and results
34
+
35
+ AUC-PR is the mean over three seeds. Full results with precision, recall and standard deviations are in Tables 5 and 6 of the [paper](https://arxiv.org/abs/2611.14207).
36
+
37
+ ### Mono-temporal models (single day of input)
38
+
39
+ | Model | Folder | `architecture` in code | Size / run | Test-Easy | Test-Medium | Test-Hard |
40
+ |---|---|---|---|---|---|---|
41
+ | U-Net | `unet` | `unet` | 373 MB | 0.285 | 0.218 | 0.175 |
42
+ | U-Net + Time-Gap Encoding | `unet_with_time_gap` | `unet_time_gap` | 374 MB | **0.287** | **0.233** | 0.192 |
43
+ | Attention U-Net | `attention_unet` | `unet_attention` | 377 MB | 0.280 | 0.199 | 0.155 |
44
+ | SegFormer U-Net | `segformer_unet` | `unet_segformer` | 571 MB | 0.205 | 0.181 | 0.162 |
45
+ | Swin-U-Net | `swin_unet` | `asufm` | 413 MB | 0.185 | 0.148 | 0.102 |
46
+ | OlmoEarth | `olmoearth` | `unet_olmo_offline` | 458 MB | 0.277 | 0.221 | 0.183 |
47
+ | DINOv3 | `dinov3` | `dual_dinov3` | 1.7 GB | 0.238 | 0.194 | **0.189** |
48
+
49
+ ### Multi-temporal models (3 consecutive days of input)
50
+
51
+ | Model | Folder | `architecture` in code | Size / run | Test-Easy | Test-Medium | Test-Hard |
52
+ |---|---|---|---|---|---|---|
53
+ | U-TAE | `utae` | `utae` | 139 MB | **0.352** | **0.306** | **0.259** |
54
+ | U-Net ConvLSTM | `unet_with_convlstm` | `unet_convlstm` | 348 MB | 0.340 | 0.278 | 0.222 |
55
+ | SimVPv2 | `simvpv2` | `simvpv2_spatiotemporal` | 66 MB | 0.284 | 0.250 | 0.203 |
56
+ | Video Swin-U-Net | `video_swin_unet` | `video_swin_unet` | 539 MB | 0.272 | 0.235 | 0.178 |
57
+ | Multi-Modal | `multi_mode_model` | `frp` | 4 MB | 0.182 | 0.167 | 0.110 |
58
+
59
+ Model descriptions and references are in the [code repository](https://github.com/hagerradi/DailyWildFireCAN#-part-4-modeling).
60
+
61
+ ## Repository structure
62
+
63
+ ```
64
+ benchmark_models_checkpoints/
65
+ ├── unet/
66
+ │ ├── checkpoints_unet_run_1/best_checkpoint.pt
67
+ │ ├── checkpoints_unet_run_2/best_checkpoint.pt
68
+ │ └── checkpoints_unet_run_3/best_checkpoint.pt
69
+ ├── unet_with_time_gap/
70
+ │ └── ...
71
+ └── <model>/checkpoints_<model>_run_<N>/best_checkpoint.pt
72
+ ```
73
+
74
+ The full repository is about 16 GB. `best_checkpoint.pt` is the checkpoint with the best validation score. It is a dictionary with the keys `model_state_dict`, `optimizer_state_dict`, `scheduler_state_dict`, `epoch` and `best_val_loss`.
75
+
76
+ ## Quick start
77
+
78
+ Clone the code repository, which contains the model definitions:
79
+
80
+ ```bash
81
+ git clone https://github.com/hagerradi/DailyWildFireCAN
82
+ cd DailyWildFireCAN
83
+ pip install -r requirements.txt
84
+ ```
85
+
86
+ Download a checkpoint and run inference:
87
+
88
+ ```python
89
+ import torch
90
+ from huggingface_hub import hf_hub_download
91
+ from src.models.unet import UNet
92
+
93
+ path = hf_hub_download(
94
+ repo_id="DailyWildfireCAN/DailyWildfireCAN-Benchmark",
95
+ filename="benchmark_models_checkpoints/unet/checkpoints_unet_run_1/best_checkpoint.pt",
96
+ )
97
+
98
+ model = UNet(input_channels=50, num_classes=1, hidden_features=[64, 128, 256, 512])
99
+ ckpt = torch.load(path, map_location="cpu")
100
+ model.load_state_dict(ckpt["model_state_dict"])
101
+ model.eval()
102
+
103
+ x = torch.randn(1, 50, 256, 256) # replace with a normalized sample's "x"
104
+ with torch.no_grad():
105
+ prob = torch.sigmoid(model(x)) # (1, 1, 256, 256) probability of new burn
106
+ ```
107
+
108
+ For the other models, build the model the same way as in `src/evaluate.py` with the settings in `configs/default.yaml`, then load `model_state_dict`.
109
+
110
+ Download all runs of one model:
111
+
112
+ ```python
113
+ from huggingface_hub import snapshot_download
114
+
115
+ snapshot_download(
116
+ repo_id="DailyWildfireCAN/DailyWildfireCAN-Benchmark",
117
+ allow_patterns=["benchmark_models_checkpoints/utae/*"],
118
+ local_dir="checkpoints",
119
+ )
120
+ ```
121
+
122
+ ## Training details
123
+
124
+ All models were trained with the same setup, using the full set of input features and the training-set normalization statistics released with the dataset. See Appendix B of the [paper](https://arxiv.org/abs/2611.14207) for hyperparameters and hardware.
125
+
126
+ ## License
127
+
128
+ The checkpoints are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The models were trained on [DailyWildfireCAN](https://huggingface.co/datasets/DailyWildfireCAN/DailyWildfireCAN), which is released under CC BY-NC-SA 4.0 and includes third-party data subject to their own licenses.
129
+
130
+ ## Citation
131
+
132
+ ```bibtex
133
+ @article{dailywildfirecan,
134
+ title={DailyWildfireCAN: A Dataset and Benchmark for Next-Day Wildfire Spread Prediction in Canada},
135
+ author={Ettahri, Mahacine and Abdelwahed, Hager Radi and Olutayo, Toluwaleke and Aloise, Daniel and Caron, Ga{\'e}tan Marceau},
136
+ journal={arXiv preprint arXiv:2611.14207},
137
+ year={2026}
138
+ }
139
+ ```
140
+
141
+ ## Contact
142
+
143
+ For inquiries, contact Mahacine Ettahri (mahacine.ettahri@etud.polymtl.ca) or Hager Radi (hager.radi@mila.quebec).