CASID: A Large-scale Climate-Aware Satellite Image Dataset for Domain Adaptive Land-Cover Semantic Segmentation

This is the official full release of the CASID dataset.

📄 Associated Paper

Title	A large-scale climate-aware satellite image dataset for domain adaptive land-cover semantic segmentation
Venue	ISPRS Journal of Photogrammetry and Remote Sensing, vol. 205, pp. 98–114, 2023 (top journal in photogrammetry & remote sensing)
Authors	Songlin Liu¹·², Linwei Chen³ (co-first authors), Li Zhang², Jun Hu⁴, Ying Fu³ (corresponding)
Affiliations	¹ Wuhan University · ² State Key Laboratory of Geo-information Engineering, Xi'an · ³ Beijing Institute of Technology · ⁴ Sun Yat-sen University
DOI	10.1016/j.isprsjprs.2023.09.007
arXiv	arXiv:2610.21238
Code	https://github.com/Linwei-Chen/CASID

Please cite the paper if you use this dataset (see Citation below).

🌍 What is CASID?

CASID is the first climate-aware land-cover semantic segmentation dataset with multiple domains. It is designed to study — and to benchmark methods against — the climate domain gap: the severe performance degradation that segmentation models suffer when applied to a different climate zone than the one they were trained on.

• 980 satellite images at 5000×5000 px (1 m/px, from Google Earth; two Singapore images are 5000×4656), collected from 30 regions across Asia (Tokyo, Seoul, Taipei/Yilan, Chongqing, Jinhua, Chiang Mai, Vientiane, Bhopal, Kuala Lumpur, Singapore, Palembang, …), covering more than 24,500 km²;
• 4 sub-datasets (domains) by climate zone — the first dataset to expose this axis:

Climate zone	Regions	Images	Forest share (paper stats → this release)
Temperate monsoon (温带季风)	7	284	45.4% → 45.5%
Subtropical monsoon (亚热带季风)	7	284	51.4% → 55.1%
Tropical monsoon (热带季风)	6	202	46.0% → 53.1%
Tropical rainforest (热带雨林)	10	210	76.2% → 76.3%
Total	30	980

The paper's published per-class statistics were computed with the same exact-color-matching script that caused the tile bug below, so two zones under-counted forest. This release restores those pixels; the difference equals the rescued forest exactly (see Data Verification).

• Pixel-wise annotations, 5 classes — annotated by experts:

id	class	palette color
0	background	black (0,0,0)
1	building	yellow (255,255,0)
2	forest	green (85,255,0)
3	road	red (255,0,0)
4	water	blue (0,112,255)

The original color coding is counter-intuitive (building = yellow, road = red); we keep it so visualizations match the original dataset and paper figures.

📉 Why it matters: the climate domain gap

Training and testing DeepLabV2 within/between climate zones (mIoU):

Train \ Test	TemMs	SubMs	TroMs	TroRf
TemMs	39.9	9.4	10.9	12.2
SubMs	10.0	60.9	12.8	15.3
TroMs	12.2	12.8	56.5	9.1
TroRf	8.5	10.2	12.8	49.0

A model trained in one climate zone collapses in another. The paper analyzes the gap from five angles (t-SNE global features, GLCM texture, category distribution, spectral values, object shapes) and benchmarks 12 segmentation + 7 unsupervised domain adaptation methods (up to DAFormer) as baselines. CASID is both a training resource and a challenging benchmark for generalizable segmentation.

⬇️ Download

pip install -U "huggingface_hub[cli]"

# full dataset (~65 GB)
hf download Linwei-Chen/CASID --repo-type dataset --local-dir CASID-dataset

# or in Python
python -c "
from huggingface_hub import snapshot_download
snapshot_download('Linwei-Chen/CASID', repo_type='dataset', local_dir='CASID-dataset')"

Mainland China mirror: export HF_ENDPOINT=https://hf-mirror.com Baidu Netdisk mirrors: temperate Link · subtropical Link · tropical monsoon Link · rainforest Link (code: casi)

Verify integrity — every file ships with its MD5 in release_file_hashes.csv:

python scripts/verify_download.py --root CASID-dataset

📁 Dataset Structure

CASID/
├── images/{climate_zone}/{region_dir}/{region_dir}_{idx:03d}_img.tif   # 5000×5000 RGB, LZW TIFF
├── labels/{climate_zone}/{region_dir}/{region_dir}_{idx:03d}_gt.png    # palette PNG, values = class ids
├── splits/train.txt  (681)          splits/test.txt  (299)
├── regions.csv            # 30 regions, bilingual (English slug ↔ 中文原名, country, counts, split)
├── file_mapping.csv       # original (Chinese) filenames → released filenames, full provenance
├── colormap.json          # class id ↔ name ↔ RGB
├── release_file_hashes.csv# size + MD5 of every released file
├── verification_report.md # full data audit behind this release
└── scripts/               # casid_split.py (tile cutter) + verify_download.py (integrity check)

• climate_zone ∈ temperate_monsoon | subtropical_monsoon | tropical_monsoon | tropical_rainforest
• region_dir = region{NN}_{english_name}, all-ASCII; original Chinese names preserved in regions.csv / file_mapping.csv
• Train/test split follows the paper: whole regions are either train or test (22 regions → 681 images; 8 test regions → 299 images), so there is no cross-region leakage.

🏷️ Labels

Palette-mode PNGs — one line of code returns the class-id map:

import numpy as np
from PIL import Image

img = Image.open("images/temperate_monsoon/region20_north_of_daqing/region20_north_of_daqing_000_img.tif")
ids = np.array(Image.open(
    "labels/temperate_monsoon/region20_north_of_daqing/region20_north_of_daqing_000_gt.png"))
print(ids.shape, ids.dtype, np.unique(ids))   # (5000, 5000) uint8 [0..4]

✂️ Cropping 1024×1024 tiles (DOTA-style)

scripts/casid_split.py reproduces the paper's training setup and supports the common DOTA conventions:

# paper-faithful default: 1024 tiles, stride 994 (30 px overlap) → 5×5 = 25 tiles per image
python scripts/casid_split.py --root CASID/ --out TILES/

# DOTA paper setting: stride 512, zero-pad edges
python scripts/casid_split.py --root CASID/ --out TILES_DOTA/ --gap 512 --pad

# mmrotate-style: overlap 200
python scripts/casid_split.py --root CASID/ --out TILES_gap200/ --gap 200 --pad

# multi-scale (DOTA convention 0.5/1.0/1.5); images bilinear, labels nearest
python scripts/casid_split.py --root CASID/ --out TILES_MS/ --rates 0.5 1.0 1.5

The default grid was verified pixel-exact against the authors' original cropped tiles.

✅ Data Verification

Before release we audited all 980 image/label pairs (full details in verification_report.md):

• every image fully decodes (978 at 5000×5000, two Singapore images at 5000×4656);
• the original labels are RGB color-coded TIFFs. Every color was mapped to a class id and all pixels in canonical annotation colors were verified pixel-perfectly (0 mismatches) against the original 1024×1024 training tiles used in the paper experiments (the tiles cover each image completely: 5×5 grid, stride 994);
• 77 images (mostly subtropical/tropical-monsoon regions) store forest in a slightly different green (76,230,0) instead of (85,255,0); the original tile-cutter matched colors exactly and silently dropped those pixels to background, so the authors' own training tiles lost forest/road/water in those regions. For this release those pixels were restored to their intended class (nearest canonical color, cross-checked against tiles and satellite imagery); per-zone class shares now match the paper's published statistics to ±0.0005 in the unaffected zones;
• every released file is hashed in release_file_hashes.csv for independent verification.

🔎 Known Quirks (kept for provenance, documented in regions.csv)

• Region ids run 02–25 plus six later rainforest regions without original ids (numbered 26–31 here); id 01 was never used.
• Original filenames mixed full-width and ASCII dashes with Chinese names; all names are ASCII now, with the complete original mapping in file_mapping.csv.
• Region 12 was named "马来西亚-斯里巴加湾南部" (Malaysia – south of Bandar Seri Begawan); BSB is the capital of Brunei, so the country is recorded as Brunei.
• Region 22 was named "朝鲜-首尔东南" (DPRK – southeast of Seoul); Seoul is in South Korea, recorded as South Korea.
• Two images in the Singapore region (region26_lim_chu_kang_036/_040) are 5000×4656; their tiles still form a complete 5×5 grid with the last row clipped to 680 px.
• The source label/ folders contain 35 stray palette PNGs (Taiwan region) with a BGR-ordered palette and unclear id semantics; they are inconsistent with the label TIFFs and are not part of this release.
• The source data also contains the authors' rendered 1024×1024 tiles and GID-style previews; these intermediates are not re-released — regenerate tiles with scripts/casid_split.py.

📄 License

The dataset is released under CC BY 4.0. The code/tools at https://github.com/Linwei-Chen/CASID are MIT.

Citation

@article{LIU202398,
    title = {A large-scale climate-aware satellite image dataset for domain adaptive land-cover semantic segmentation},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {205},
    pages = {98-114},
    year = {2023},
    issn = {0924-2716},
    doi = {10.1016/j.isprsjprs.2023.09.007},
    url = {https://www.sciencedirect.com/science/article/pii/S0924271623002484},
    author = {Songlin Liu and Linwei Chen and Li Zhang and Jun Hu and Ying Fu},
}

---

中文简介

CASID 是首个按气候带划分多域的地表覆盖语义分割数据集，官方完整发布版。对应论文发表于遥感顶级期刊 ISPRS Journal of Photogrammetry and Remote Sensing（205 卷，98–114 页，2023）：A large-scale climate-aware satellite image dataset for domain adaptive land-cover semantic segmentation（刘松林、陈林蔚为共同第一作者，付莹教授为通讯作者，DOI）。

• 980 张 5000×5000 卫星影像（1 米分辨率，Google Earth）+ 逐像素标注，来自亚洲 30 个区域（东京、首尔、台湾宜兰、重庆、金华、清迈、万象、博帕尔、吉隆坡、新加坡、巨港……），覆盖超 24,500 平方公里；
• 按气候带分四个域：温带季风（7 区 284 张）、亚热带季风（7 区 284 张）、热带季风（6 区 202 张）、热带雨林（10 区 210 张）——跨气候带训练测试，DeepLabV2 的 mIoU 会从 50–60 崩到 10 左右（温带→雨林仅 12.2）；
• 标注 5 类：0=背景、1=建筑、2=森林、3=道路、4=水体（调色板 PNG，np.array 一行读出类别 id）；
• 划分沿用论文实验：按区域整体分 train（681）/ test（299），同区域的图不会一边进训练一边进测试；
• 全 ASCII 文件名，regions.csv 保留中文原名对照，file_mapping.csv 记录原始文件名到发布文件名的完整映射；
• scripts/casid_split.py DOTA 式大图滑窗裁切：默认 1024、步距 994（重叠 30），每张图恰好 25 块、与论文实验逐像素一致；也支持 --gap 512 --pad（DOTA 论文设置）、--gap 200（mmrotate 常用）与 --rates 0.5 1.0 1.5 多尺度；
• scripts/verify_download.py 一条命令校验下载完整性（release_file_hashes.csv 含每个文件的 MD5）；
• 发布前已对全部 980 对数据做完整性与逐像素一致性校验（详见 verification_report.md），并修复了原始数据中 77 张图因调色板变体导致的图块森林丢失问题。

使用数据集请引用论文（BibTeX 见上）。GitHub 使用指南：https://github.com/Linwei-Chen/CASID