Integrating time-series analysis and deep learning method to reconstruct the RTS dynamics in the Tibetan Plateau over the past nearly four decades

# Remote Sensing Data Analysis and Object Detection Project## Project OverviewThis project is a comprehensive remote sensing data analysis toolkit that includes two core functional modules:1. **Time Series Classification and Segmentation**: Utilizes LSTM models for time series data classification and Random Forest (RF) for time series segmentation, primarily used to analyze vegetation index (such as NDVI) trends.2. **YOLO Object Detection**: Object detection functionality based on the YOLO model, used for object recognition and localization in remote sensing images.## Installation Guide### Requirements- Python 3.7+- Dependencies: numpy, matplotlib, joblib, tensorflow, ultralytics, opencv-python### Installation Steps1. Clone or download this project to your local machine, unzip data.zip2. Install the required dependencies:```bashpip install numpy matplotlib joblib tensorflow ultralytics opencv-python```## Usage Instructions### Time Series Classification and SegmentationUse the `Time_series_classification_segmentation.py` script for time series data classification and segmentation:```bashpython Time_series_classification_segmentation.py```This module will:- Load the LSTM classification model to determine if an event exists- If an event exists, load the Random Forest regressor to predict key time points- Retrieve NDVI Z-scores from Zarr files and generate visualization results### YOLO Object DetectionUse the `Yolo_objection_detection.py` script for object detection in remote sensing images:```bashpython Yolo_objection_detection.py```By default, this script will:- Load the pre-trained YOLO model- Perform object detection on specified test images- Display detection results with object annotations## Project Structure```├── README.txt # Project documentation├── Time_series_classification_segmentation.py # Time series classification and segmentation module├── Yolo_objection_detection.py # YOLO object detection module├── utils.py # Utility functions library├── models/ # Model files directory│ ├── LSTM_time_series_classification.h5 # LSTM classification model│ ├── Yolo_object_detection.pt # YOLO object detection model│ └── time_series_segmentation.joblib # Time series segmentation model├── results/ # Results output directory│ ├── RTS.gpkg # Geographic package file│ ├── RTS_start_year.tif # Start year raster file│ └── TSC_Model_predcition.tif # Model prediction result raster file├── test_images/ # Test images directory│ ├── 25213.jpg│ ├── 4145.jpg│ ├── 4206.jpg│ └── wechat_2025-08-07_165732_336.png└── vegetation_indices.zarr/ # Vegetation indices Zarr data storage```## Data Description- `vegetation_indices.zarr`: Contains time series data for multiple vegetation indices, such as NDVI, ZRNDVI, etc.- `test_images`: Contains sample images for testing YOLO object detection### Results Files DescriptionThe `results/RTS.gpkg` GeoPackage contains three files:1. **RTS**: Retrogressive Thaw Slumps (RTSs) extracted by our algorithm. This dataset is also accompanied by a web-based visualization platform (https://majiapei.github.io/Landsat_VHR_RTSs/), where users can view and compare our results with those derived from very high-resolution imagery2. **RTS_contains_ALDS**: RTSs delineated by manual visual interpretation based on disturbed areas, containing approximately 7,000 RTSs. Note that this dataset may include some deep Active Layer Detachments (ALDs), as deep ALDs are very likely to develop into RTSs. According to Luo et al. (2022), about 98% of RTSs on the Qinghai-Tibet Plateau are formed from ALDs, and some studies suggest that deep ALDs are actually early-stage RTSs (Nesterova et al., 2024).3. **Start_Year_map_from_1986**: A raster map showing the development of RTSs obtained using a time-series segmentation algorithm. The raster value + 1986 indicates the year when the vegetation in that pixel was disturbed (approximately representing the occurrence time of the RTS).## Model Description- `LSTM_time_series_classification.h5`: LSTM model for time series event classification- `time_series_segmentation.joblib`: Random Forest model for time series segmentation- `Yolo_object_detection.pt`: YOLO model for object detection## References- Luo, J., Niu, F., Lin, Z., Liu, M., Yin, G., Gao, Z., 2022. Inventory and frequency of retrogressive thaw slumps in permafrost region of the Qinghai–Tibet Plateau. Geophysical Research Letters, 49, e2022GL099829. https://doi.org/10.1029/2022GL099829- Nesterova, N., Leibman, M., Kizyakov, A., Lantuit, H., Tarasevich, I., Nitze, I., Veremeeva, A., Grosse, G., 2024. Review article: Retrogressive thaw slump characteristics and terminology. The Cryosphere, 18, 4787–4810. https://doi.org/10.5194/tc-18-4787-2024## Contact InformationFor questions or suggestions, please contact the project maintainer.