VisionNDVI

A script-driven repository for image/time-series processing and model training/inference (organized by scripts in the repository root directory: tif2npy.py, inference.py, various train_OF_*.py files, etc.).

Repository: @KeranLi/VisonNDVI

Table of Contents

• VisionNDVI
  • Table of Contents
  • Introduction
  • Quick Start
  • Main Scripts & Usage Examples
  • Dependencies
  • Data Format & Suggested Workflow
  • Development & Training
  • Contributing
  • License

Introduction

VisionNDVI is a codebase focused on image/time-series data processing and deep learning model training/inference (script-driven repository). The repository includes:

• tif2npy.py: Preprocessing script for converting GeoTIFF or raster images to NumPy format (for training/inference pipelines).
• check_bad_npy.py: Script to check/repair corrupted or anomalous .npy files (e.g., missing values, inconsistent shapes).
• train_OF_*.py: A set of scripts to train different model variants (e.g., CNN, UNet, 3D-CNN+MLP).
• inference.py: Script for running inference/prediction using a trained model.
• make_SOV.py: Computes/generates SOV or evaluation metrics (depending on implementation).
• models/, loss/, utils/: Directories for model definitions, loss functions, and utility functions.

Note: This repository is currently not packaged as a command-line tool (console_scripts); scripts are run via python <script>.py.

Quick Start

1. Clone the repository:

   git clone https://github.com/KeranLi/VisionNDVI.git
   cd VisionNDVI

2. Create/install dependencies and activate the virtual environment:

   conda env create -f environment.yml -n VisionNDVI
   conda activate VisionNDVI

3. Run examples:

   • Convert GeoTIFF to .npy:

     python tif2npy.py --input-dir data/tifs --output-dir data/npy --pattern "*.tif"

   • Check .npy files:

     python check_bad_npy.py --dir data/npy

   • Train a model (see script comments for exact parameters):

     python train_OF_UnetCNN.py --config configs/train_unet.yaml

   • Run inference / generate results:

     python inference.py --model checkpoints/latest.pth --input data/npy --output results/predictions

Main Scripts & Usage Examples

Here are common scripts and recommended usage (example parameters are placeholders, please replace them with actual values based on script implementation):

• tif2npy.py

  • Purpose: Converts raw GeoTIFF/raster data into .npy/array format for training/inference.
  • Example:

    python tif2npy.py --input-dir data/tifs --output-dir data/npy --band-mapping red:3,nir:4 --tile-size 256

• check_bad_npy.py

  • Purpose: Scans and reports (or fixes) corrupted/size anomalies/NaN values in .npy files.
  • Example:

    python check_bad_npy.py --dir data/npy --fix

• train_OF_*.py

  • Purpose: Trains different model architectures (e.g., train_OF_UnetCNN.py, train_OF_CNNMLP.py, etc.).
  • Recommended: Use YAML/JSON config files to manage training parameters (learning rate, batch size, data paths, logging/checkpoint directories, etc.).
  • Example:

    python train_OF_UnetCNN.py --config configs/train_unet.yaml --device cuda:0

• inference.py

  • Purpose: Loads a trained model and performs inference on new data, outputting predictions (e.g., .npy, .png, .geotiff).
  • Example:

    python inference.py --model checkpoints/best.pth --input data/npy --output results --batch-size 8

• make_SOV.py

  • Purpose: Computes SOV or other evaluation metrics (parameters depend on implementation).
  • Example:

    python make_SOV.py --pred-dir results --gt-dir ground_truth --out metrics/sov_report.csv

If needed, I can replace these examples with exact command-line options based on the actual parameters in the script header. Please paste the first few hundred lines of the script or allow me to access the file content.

Dependencies

• Python 3.8+
• numpy
• torch / torchvision
• opencv-python (cv2)
• rasterio
• tqdm
• matplotlib / pillow
• scikit-image

Data Format & Suggested Workflow

• Raw images should ideally be in GeoTIFF format with geospatial information (if GeoTIFF output is needed later).
• Use tif2npy.py to convert raw images into uniform NumPy arrays for training/inference.
• For large-scale or super-large images, it's recommended to use tile-based/window processing to manage memory usage.
• Keep the training/validation/test data paths consistent and clearly define nodata values and masking strategies in the configuration.

Development & Training

• Recommended tools:
  • Code style: black, flake8
  • Testing: pytest (if the repository includes tests)
• Training debugging:
  • Start with small-scale datasets to verify loss reduction and output validity before scaling up to full training.
  • Use checkpointing and periodic validation set evaluation.

Contributing

We welcome bug reports, feature suggestions, and patches via issues or PRs. Suggested contribution process:

1. Fork the repository and create a feature branch.
2. Add/modify code and include necessary documentation and tests.
3. Submit a PR with a description of the changes and their impact.

If you'd like me to add standard contributing guidelines (CONTRIBUTING.md) or replace the README script usage with precise parameters, I can complete the task and submit a PR once file content access is granted.

License

Please refer to the LICENSE file in the repository for licensing information (currently recommended under the MIT License).