zhangrenchao commited on
Commit
80bd41b
·
verified ·
1 Parent(s): aed6f6f

Update English model card

Browse files
Files changed (1) hide show
  1. README.md +58 -63
README.md CHANGED
@@ -1,19 +1,19 @@
1
  ---
2
- frameworks: PyTorch
3
  language:
4
  - en
5
- license: apache-2.0
6
  tags:
7
  - OneScience
8
  - Earth Science
9
- - Remote Sensing
10
- - Multi-Task Image Understanding
11
- - SatlasPretrain
12
- - arxiv:2211.15660
13
- tasks: []
14
  datasets:
15
- - allenai/satlas-pretrain
 
 
16
  ---
 
17
  <p align="center">
18
  <strong>
19
  <span style="font-size: 30px;">SatlasPretrain</span>
@@ -22,50 +22,48 @@ datasets:
22
 
23
  # Model Introduction
24
 
25
- SatlasPretrain is a large-scale remote sensing pre-training approach for learning from multi-temporal imagery and seven label types. Its unified SatlasNet model supports semantic segmentation, regression, point detection, polygon instances, polylines, object properties, and image classification for Earth observation and downstream remote sensing applications.
26
-
27
- Paper: SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding
28
 
 
29
  https://arxiv.org/abs/2211.15660
30
 
31
  # Model Description
32
 
33
- SatlasPretrain was proposed by the Allen Institute for AI and trained on Sentinel-2 and NAIP imagery with 302 million labels in 137 categories and seven label types.
34
- The model is suitable for multi-temporal remote sensing segmentation, regression, detection, property prediction, and classification tasks.
35
 
36
  # Use Cases
37
 
38
  | Scenario | Description |
39
  | :---: | :--- |
40
- | Remote sensing multi-task training | Train SatlasPretrain with multi-temporal imagery and segmentation, regression, detection, property, and classification labels. |
41
- | Local quick validation | Use synthetic data to validate loading, multi-task training, inference, evaluation, and qualitative visualization. |
42
- | Hugging Face / OneCode execution | Download the standalone model package, install dependencies, and run the scripts directly. |
43
- | Multi-GPU training | Launch distributed multi-process training with `torchrun`. |
 
 
44
 
45
  # Usage Guide
46
 
47
- ## 1. OneCode Usage
48
 
49
  Experience intelligent one-click AI4S programming through the OneCode online environment:
50
 
51
  [Click to Experience Intelligent One-Click AI4S Programming](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home)
52
 
53
- ## 2. Manual Installation and Usage
54
-
55
- **Hardware Requirements**
56
-
57
- - A GPU or DCU is recommended.
58
- - CPU can be used for import and small-scale connectivity verification; full training and inference will be slow.
59
- - DCU users must install DTK in advance. DTK 25.04.2 or above, or the OneScience recommended version matching the cluster, is recommended.
60
-
61
- ### Download the Model Package
62
 
63
  ```bash
64
  hf download OneScience-Group/SatlasPretrain --local-dir ./SatlasPretrain
65
  cd SatlasPretrain
66
  ```
67
 
68
- ### Install the Runtime Environment
 
 
 
 
 
 
69
 
70
  **DCU Environment**
71
 
@@ -89,49 +87,55 @@ pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simp
89
 
90
  ### Training Data Introduction
91
 
92
- The original SatlasPretrain dataset is published by the Allen Institute for AI and organized as Web-Mercator tiles. Sentinel-2, Sentinel-1, and NAIP imagery and bands are stored as PNG files. Point, polygon, polyline, property, and classification labels are stored in `vector.json`; semantic segmentation and regression labels are grayscale PNG files; split and image-time metadata are JSON files.
93
 
94
- Official data resources:
95
 
96
- - https://github.com/allenai/satlas/blob/main/SatlasPretrain.md
97
- - https://huggingface.co/datasets/allenai/satlas-pretrain
98
 
99
- The original dataset is not included in this model repository. The provided scripts use synthetic NPZ data to validate the full training, inference, evaluation, and visualization pipeline. This NPZ protocol is only a compact pipeline check and is not the original SatlasPretrain format.
 
 
 
 
 
 
 
 
 
 
 
 
 
100
 
101
- Generate synthetic data:
102
 
103
  ```bash
104
  python scripts/fake_data.py
105
  ```
106
 
107
- The generated `data/fake_satlaspretrain.npz` contains four samples with `images` in `[B, T, C, H, W]` layout and labels for all seven task heads. Verify that `conf/config.yaml` points to this file before training.
108
-
109
  ### Training
110
 
111
- Single GPU or CPU:
112
-
113
  ```bash
114
  python scripts/train.py
115
  ```
116
 
117
- Multi-GPU:
118
 
119
  ```bash
120
- torchrun --nproc_per_node=8 --nnodes=1 --rdzv_id=1000 --rdzv_backend=c10d --max_restarts=0 --master_addr="localhost" --master_port=29500 scripts/train.py
121
  ```
122
 
123
- Training outputs:
124
 
125
  ```text
126
- data/checkpoints/best.pth
127
- data/checkpoints/last.pth
128
  ```
129
 
130
- `best.pth` stores the checkpoint with the lowest joint multi-task validation loss, while `last.pth` stores the final training epoch. Checkpoints generated from synthetic data validate the pipeline only and are not pretrained weights reproducing the paper results.
131
-
132
  ### Training Weights
133
 
134
- This repository does not include synthetic, trained, or official pretrained weights. Running the training script generates `best.pth` and `last.pth` under `data/checkpoints/`. For real downstream use, train with appropriately converted SatlasPretrain data or use official AllenAI foundation model weights under their applicable license.
135
 
136
  ### Inference
137
 
@@ -139,35 +143,25 @@ This repository does not include synthetic, trained, or official pretrained weig
139
  python scripts/inference.py
140
  ```
141
 
142
- Prediction outputs:
143
 
144
  ```text
145
- result/output/segmentation.npy
146
- result/output/regression.npy
147
- result/output/point.npy
148
- result/output/polygon.npy
149
- result/output/polyline.npy
150
- result/output/property.npy
151
- result/output/classification.npy
152
  ```
153
 
154
- These files store class probabilities, dense regression values, confidence maps, and category probabilities for the seven compact task heads.
155
-
156
  ### Evaluation and Visualization
157
 
158
  ```bash
159
  python scripts/result.py
160
  ```
161
 
162
- Evaluation outputs:
163
 
164
  ```text
165
- result/output/satlaspretrain_metrics.json
166
- result/output/qualitative_comparison.ppm
167
  ```
168
 
169
- The metrics follow the task families reported in the paper: segmentation F1, regression MAE, point mAP, polygon mAP, polyline GEO, property accuracy, and classification accuracy. The qualitative comparison displays input imagery and paired ground-truth/prediction panels. Metrics and figures produced from synthetic data are pipeline checks, not paper results.
170
-
171
  # Official OneScience Resources
172
 
173
  | Platform | OneScience Main Repository | Skills Repository |
@@ -177,5 +171,6 @@ The metrics follow the task families reported in the paper: segmentation F1, reg
177
 
178
  # Citation and License
179
 
180
- - This repository is a reproduction of the original SatlasPretrain paper.
181
- - The reproduction code is released under Apache License 2.0. The original SatlasPretrain data and official model weights remain subject to their respective source licenses.
 
 
1
  ---
2
+ license: apache-2.0
3
  language:
4
  - en
 
5
  tags:
6
  - OneScience
7
  - Earth Science
8
+ - Remote Sensing Multi-Task Learning
9
+ - Multi-Temporal Remote Sensing
10
+ frameworks: PyTorch
 
 
11
  datasets:
12
+ - SatlasPretrain
13
+ - NAIP
14
+ - Sentinel-2
15
  ---
16
+
17
  <p align="center">
18
  <strong>
19
  <span style="font-size: 30px;">SatlasPretrain</span>
 
22
 
23
  # Model Introduction
24
 
25
+ SatlasPretrain is a multi-task pretraining model for large-scale remote sensing understanding. It uses a dual-stream hierarchical backbone to process multi-temporal high-resolution RGB and Sentinel-2 multispectral imagery and jointly learns dense prediction and global classification tasks.
 
 
26
 
27
+ Paper: SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding
28
  https://arxiv.org/abs/2211.15660
29
 
30
  # Model Description
31
 
32
+ SatlasPretrain was proposed by a research team from the Allen Institute for AI. The model is trained with NAIP-style high-resolution imagery, Sentinel-2 multispectral imagery, and multiple types of remote sensing labels from SatlasPretrain. It is suitable for remote sensing multi-task learning, including semantic segmentation, regression, point, polyline, and polygon object prediction, property recognition, and scene classification.
 
33
 
34
  # Use Cases
35
 
36
  | Scenario | Description |
37
  | :---: | :--- |
38
+ | Multi-temporal remote sensing fusion | Jointly process 4 time steps of high-resolution RGB and 8 time steps of Sentinel-2 data. |
39
+ | Remote sensing multi-task learning | Simultaneously predict segmentation, regression, point, polygon, polyline, property, and classification tasks. |
40
+ | Land-cover and crop mapping | Use multi-temporal imagery for semantic segmentation tasks such as land cover and crop type mapping. |
41
+ | Infrastructure feature extraction | Identify point, polyline, and polygon geographic objects such as buildings, roads, railways, airports, and energy facilities. |
42
+ | Local engineering validation | Use a small amount of synthetic data to check the training, inference, and evaluation workflows. |
43
+ | Multi-GPU training | Launch distributed training with `torchrun`. |
44
 
45
  # Usage Guide
46
 
47
+ ## 1. OneCode
48
 
49
  Experience intelligent one-click AI4S programming through the OneCode online environment:
50
 
51
  [Click to Experience Intelligent One-Click AI4S Programming](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home)
52
 
53
+ ## 2. Download and Installation
 
 
 
 
 
 
 
 
54
 
55
  ```bash
56
  hf download OneScience-Group/SatlasPretrain --local-dir ./SatlasPretrain
57
  cd SatlasPretrain
58
  ```
59
 
60
+ ### Environment Dependencies
61
+
62
+ **Hardware Requirements**
63
+
64
+ - A GPU or DCU is recommended.
65
+ - CPU can be used for small-configuration connectivity validation; full training and inference will be slow.
66
+ - DCU users must install DTK in advance. DTK 25.04.2 or above, or the OneScience recommended version matching the current cluster, is recommended.
67
 
68
  **DCU Environment**
69
 
 
87
 
88
  ### Training Data Introduction
89
 
90
+ By default, 1 training and 1 test synthetic sample are used to validate the engineering workflow. They are saved as `data/train.npz` and `data/test.npz`, respectively.
91
 
92
+ The synthetic data preserves the official multi-temporal configuration of 4 NAIP time steps, 8 Sentinel-2 time steps, a `512x512` spatial size, and seven task-label specifications.
93
 
94
+ Real data must be preprocessed and converted to the following NPZ training protocol. This protocol is consistent with the model input specification but is not the download format of the original datasets.
 
95
 
96
+ ```text
97
+ highres_images: float32 [N,4,3,512,512]
98
+ lowres_images: float32 [N,8,9,512,512]
99
+ valid_highres_times: bool [N,4]
100
+ valid_lowres_times: bool [N,8]
101
+ sample_ids: string [N]
102
+ segmentation: int64 [N,512,512]
103
+ regression: float32 [N,1,512,512]
104
+ point: float32 [N,1,512,512]
105
+ polygon: float32 [N,1,512,512]
106
+ polyline: float32 [N,1,512,512]
107
+ property: int64 [N]
108
+ classification: int64 [N]
109
+ ```
110
 
111
+ `fake_data.py` automatically writes the `protocol` and `source` protocol metadata. These fields must be retained when using real data.
112
 
113
  ```bash
114
  python scripts/fake_data.py
115
  ```
116
 
 
 
117
  ### Training
118
 
 
 
119
  ```bash
120
  python scripts/train.py
121
  ```
122
 
123
+ For multi-GPU training, use:
124
 
125
  ```bash
126
+ torchrun --nproc_per_node=8 scripts/train.py
127
  ```
128
 
129
+ Training jointly optimizes seven remote sensing task types and saves a checkpoint and aggregate training metrics. The default configuration is intended for quick workflow validation. Formal experiments should use the multi-temporal data, complete task labels, model configuration, and training duration corresponding to the paper.
130
 
131
  ```text
132
+ result/checkpoints/satlaspretrain.pt
133
+ result/training/metrics.json
134
  ```
135
 
 
 
136
  ### Training Weights
137
 
138
+ This repository will provide SatlasPretrain training weights in the `weight/` folder. The weight files will be uploaded soon and are expected to be available in the near future.
139
 
140
  ### Inference
141
 
 
143
  python scripts/inference.py
144
  ```
145
 
146
+ Inference loads the training checkpoint, generates predictions for seven task types, retains sample identity and protocol metadata, and saves the results to:
147
 
148
  ```text
149
+ result/output/predictions.npz
 
 
 
 
 
 
150
  ```
151
 
 
 
152
  ### Evaluation and Visualization
153
 
154
  ```bash
155
  python scripts/result.py
156
  ```
157
 
158
+ Evaluation covers segmentation, regression, point, polyline, and polygon objects, property, and classification tasks, and generates a multi-task prediction figure. Results on synthetic data are only for engineering workflow validation and do not represent full-paper performance.
159
 
160
  ```text
161
+ result/evaluation/metrics.json
162
+ result/evaluation/multitask_predictions.png
163
  ```
164
 
 
 
165
  # Official OneScience Resources
166
 
167
  | Platform | OneScience Main Repository | Skills Repository |
 
171
 
172
  # Citation and License
173
 
174
+ This repository is a reproduction of the original SatlasPretrain paper.
175
+
176
+ Use of the code and data in this repository remains subject to the licenses and terms of use of their respective projects.