Unsupervised anomaly localization with high-resolution and time-series satellite imagery: from global disaster dataset to zero-shot model via SAM（ISPRS2026）

AnomalyCD is a two-stage anomaly change detection framework for multi-temporal remote sensing imagery. It first performs bitemporal change detection with SAM (Segment Anything Model), then distinguishes normal changes from anomalous changes using historical normal observations.

This repository provides a self-contained implementation for inference and evaluation. Core dependencies (SAM and GeoTIFF I/O utilities) are bundled locally, so no external project paths need to be configured.

---

Overview

AnomalyCD follows a two-stage pipeline:

Stage	Script	Description
Stage 1	run_stage1.py	Detect pixel-level changes between the latest normal image and the anomaly image
Stage 2	run_stage2.py	Identify anomalous changes using all normal temporal images and Stage 1 outputs
Evaluation	run_eval.py	Per-event recall, precision, and F1 for Stage 1 and Stage 2

Stage 1 — Candidate Change Detection

1. Feed the latest normal image and the anomaly image into SAM.
2. Generate object masks and image-encoder feature maps for both images.
3. Compare features within SAM segments and produce a continuous change map.

Stage 2 — Anomaly Identification

1. Load the Stage 1 change map and all normal temporal images together with the anomaly image.
2. Extract SAM features for each temporal observation over changed regions.
3. Measure temporal feature dissimilarity relative to the anomaly image and produce the final AnomalyCD map.

---

GDL Dataset

Download

The GDL (Global Disaster Land surface) dataset can be requested through the following form:

Download GDL dataset

Submit your name, organization, and email to receive download instructions.

Global Coverage

The dataset covers normal control sites and anomaly events across six major categories worldwide:

ID	Category	Color in map
0	Normal	Black
1	Explosion	Red
2	Collapse	Orange
3	Landslide	Green
4	Fire	Yellow
5	Dam break	Light blue
6	Others	Grey

Event folder names are prefixed with the category ID (e.g., 0_ for normal, 1_ for explosion).

Directory Layout

After downloading and extracting the dataset, each event is stored in its own folder with multi-temporal GeoTIFF files sorted by filename.

Anomaly event example:

data/
└── 1_Equatorial_Guinea_Explosion-20210313_30/
    ├── anomaly_20200913.tif          # anomaly image        (term_names[0])
    ├── anomaly_20200913_label.tif    # pixel-wise annotation (term_names[1])
    ├── normal_20040712.tif           # normal temporal image
    ├── normal_20140913.tif
    └── normal_20170913.tif           # latest normal image  (term_names[-1])

Normal control event example (ID 0):

data/
└── 0_Egypt_30/
    ├── normal_20180622.tif
    ├── normal_20181222.tif
    ├── normal_20190622.tif
    └── normal_20201222.tif           # latest image (term_names[-1])

Normal events contain only normal_* images. Stage 1 compares the last two temporal images; Stage 2 uses the full time series.

File Naming Convention

Prefix	Meaning
anomaly_*	Anomaly-time remote sensing image
*_label*	Pixel-wise annotation raster
normal_*	Historical normal-time images
Event folder prefix 0_	Normal control event (category ID 0, no label file)
Event folder prefix 1–6	Anomaly event category (see table above)

Inference vs. evaluation: Stage 1 and Stage 2 process both anomaly and normal events. Only run_eval.py skips normal control events (0_), because they have no binary anomaly labels for recall / precision / F1.

Annotation Format

The label raster is not a binary map. Each pixel value encodes a semantic category:

• 0: background
• Different positive integer values: different anomalous land-cover / object categories within the scene

Multiple anomaly instances or land-cover types in the same event may therefore carry different label IDs in a single annotation file.

Note on evaluation: For TPR / precision / F1 evaluation, the code converts labels to a binary anomaly mask via binarize_label() (value 1 is treated as background; all other positive values are merged into the anomaly class). This follows the original evaluation protocol and does not change the multi-class nature of the raw annotations.

---

Environment Setup

1. Install Dependencies

Create a Python 3.7+ environment and install the required packages:

cd AnomalyCD
pip install -r requirements.txt

Or activate the pre-configured RSAD conda environment:

conda activate RSAD
pip install -r requirements.txt

Core dependencies:

Package	Version (tested)	Purpose
numpy	1.21.6	Array operations
torch / torchvision	1.13.0 / 0.14.0	SAM inference (CUDA)
opencv-python	4.8.0.76	Image processing
scikit-learn	1.0.2	Evaluation metrics
scikit-image	0.19.2	Morphological post-processing
GDAL	3.0.2	GeoTIFF read/write
tqdm	4.65.2	Progress bars

Note: PyTorch with CUDA and GDAL are recommended to install via conda when pip install fails:

conda install pytorch==1.13.0 torchvision==0.14.0 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install -c conda-forge gdal=3.0.2
pip install -r requirements.txt

2. SAM Checkpoint

Download the SAM ViT-B weights and set the path in config.py:

SAM_CHECKPOINT = "/path/to/sam_vit_b_01ec64.pth"

3. Data and Output Paths

Update the default paths in config.py:

DEFAULT_DATA_ROOT = "/path/to/GDL/data"
DEFAULT_OUTPUT_ROOT = "/path/to/output/AnomalyCD"

---

Repository Structure

AnomalyCD/
├── README.md
├── config.py                   # global configuration
├── img_io.py                   # GeoTIFF read/write (GDAL)
├── segment_anything_origin/    # bundled SAM library
├── utils.py                    # shared utilities
├── stage1_analysis.py          # Stage 1 core algorithm
├── stage2_analysis.py          # Stage 2 core algorithm
├── eval_metrics.py             # evaluation metrics
├── run_stage1.py               # Stage 1 entry point
├── run_stage2.py               # Stage 2 entry point
├── run.py                      # full pipeline entry point
├── run_eval.py                 # evaluation entry point
├── quick_test.py               # single-patch quick test
├── requirements.txt            # Python dependencies
└── docs/
    └── images/
        ├── framework.png
        └── dataset_distribution.png

---

Quick Start

cd AnomalyCD
conda activate RSAD

1. Quick Test (~30–40 seconds)

Verify SAM weights, CUDA, and data paths before full inference:

python quick_test.py

Expected output: QUICK TEST PASSED.

2. Full Pipeline

# Run Stage 1 and Stage 2 sequentially on all events
python run.py --stage all --device cuda:0

# Stage 1 only
python run.py --stage 1 --device cuda:0

# Stage 2 only (requires Stage 1 outputs)
python run.py --stage 2 --device cuda:0

# Single event
python run.py --stage all --device cuda:0 --event "1_Equatorial_Guinea_Explosion-20210313_30"

3. Stage-wise Execution

Stage 1 — Change Detection

python run_stage1.py --device cuda:0
python run_stage1.py --device cuda:0 --no-skip-existing
python run_stage1.py --device cuda:0 --event "1_Equatorial_Guinea_Explosion-20210313_30"

Stage 2 — Anomaly Change Detection

python run_stage2.py --device cuda:0
python run_stage2.py --device cuda:0 --event "1_Equatorial_Guinea_Explosion-20210313_30"

4. Evaluation

After Stage 2 finishes, run the official evaluation:

python run_eval.py

Each event prints Stage 1 and Stage 2 recall, precision, and F1:

1_xxx_event_name
  Stage1  recall=0.8500  precision=0.7200  F1=0.7800
  Stage2  recall=0.9100  precision=0.8000  F1=0.8520

Save per-event records to CSV:

python run_eval.py --output-csv eval_records.csv

---

Command-Line Arguments

run_stage1.py

Argument	Default	Description
--data-root	config.DEFAULT_DATA_ROOT	Dataset root directory
--output-root	config.STAGE1_OUTPUT_DIR	Stage 1 output directory
--device	cuda:0	GPU device
--patch-size	2048	Sliding-window patch size
--event	all events	Process a single event folder
--no-skip-existing	—	Re-run events that already have outputs

run_stage2.py

Argument	Default	Description
--data-root	config.DEFAULT_DATA_ROOT	Dataset root directory
--stage1-root	config.STAGE1_OUTPUT_DIR	Stage 1 output directory
--output-root	config.STAGE2_OUTPUT_DIR	Stage 2 output directory
--device	cuda:0	GPU device
--event	all events	Process a single event folder
--skip-existing	False	Skip events with existing AnomalyCD_map.tif

run_eval.py

Argument	Default	Description
--data-root	config.DEFAULT_DATA_ROOT	Dataset root (for reading labels)
--result-root	config.STAGE2_OUTPUT_DIR	Stage 2 result directory
--max-events	80	Maximum number of events to evaluate
--quantile-threshold	0.945	Quantile threshold for Stage 1 binarization
--stage2-threshold	0.08	Fixed threshold for Stage 2 AnomalyCD binarization
--area-ratio	0.0003	Minimum connected-component area ratio
--background-weight	0.1	Background false-positive weight in precision
--normalize-anomaly-map	False	Apply min-max normalization before thresholding
--output-csv	—	Save per-event evaluation records

---

Outputs

Stage 1

Directory: {STAGE1_OUTPUT_DIR}/{event_name}/

File	Description
change_map_continuous.tif	Continuous change map

Stage 2

Directory: {STAGE2_OUTPUT_DIR}/{event_name}/

File	Description
change_map_filtered.tif	Filtered change map (quantile 0.7)
AnomalyCD_map.tif	Continuous anomaly change detection map
AnomalyCD_map.png	Binarized preview using evaluation settings (fixed threshold 0.08 + morphology)

Evaluation

run_eval.py prints per-event metrics only:

• Stage 1: recall, precision, F1 on change_map_filtered.tif
• Stage 2: recall, precision, F1 on AnomalyCD_map.tif

---

Hyperparameters

Key settings in config.py:

Parameter	Stage	Default	Description
STAGE1_PATCH_SIZE	1	2048	Sliding-window patch size
STAGE1_CHANGE_QUANTILE	1	0.9	Change-map binarization quantile
STAGE1_SAM_PARAMS	1	see config	SAM automatic mask generator settings
STAGE2_CHANGE_MAP_QUANTILE	2	0.7	Change-map filtering quantile
STAGE2_MIN_MASK_PIXELS	2	1500	Minimum mask pixel count
STAGE2_SAM_PARAMS	2	see config	SAM automatic mask generator settings
EVAL_QUANTILE_THRESH	Eval	0.945	Stage 1 evaluation binarization quantile
EVAL_STAGE2_FIXED_THRESH	Eval	0.08	Stage 2 AnomalyCD fixed binarization threshold
EVAL_AREA_RATIO	Eval	0.0003	Small-region filtering area ratio
EVAL_BACKGROUND_WEIGHT	Eval	0.1	Weighted precision background weight

Stage 2 automatically selects the patch size: 1024 when the short side is below 2048, otherwise 2048.

---

Evaluation Protocol

The evaluation follows the protocol in the original Eval/compare_tpr_recall_F1_.py:

1. Change map (Stage 1): binarized by quantile threshold (0.945); no morphological post-processing
2. AnomalyCD map (Stage 2): binarized by fixed threshold (0.08), then morphologically closed and filtered by small connected components
3. Recall (TPR): fraction of anomaly pixels correctly detected
4. Precision: weighted precision with background false positives down-weighted by 0.1
5. F1: harmonic mean of recall and precision
6. Normal control events (0_ prefix) are skipped during evaluation only

---

Notes

1. GPU required for Stage 1 and Stage 2 inference. run_eval.py runs on CPU only.
2. Execution order: Stage 1 → Stage 2 → Evaluation.
3. Stage 1 skips existing outputs by default. Use --no-skip-existing to force re-inference.
4. Runtime: full inference on high-resolution images (e.g., 9840×13184) can take a long time. Use quick_test.py first, then debug with --event.
5. Legacy code remains in the parent repository:
   • Stage 1: SAM_Change_Detection/run.py
   • Stage 2: SAM_Anomaly_Change/run.py
   • Evaluation: Eval/compare_tpr_recall_F1_.py

---

End-to-End Example

# 1. Activate environment
conda activate RSAD
cd AnomalyCD

# 2. Edit data paths and SAM checkpoint in config.py

# 3. Verify setup
python quick_test.py

# 4. Debug on a single event
python run.py --stage all --device cuda:0 --event "1_Equatorial_Guinea_Explosion-20210313_30"

# 5. Full inference
python run.py --stage all --device cuda:0

# 6. Evaluation
python run_eval.py --output-csv eval_records.csv