- GPKG: a folder storing the lake_poygons_test.gpkg file, containing polygon outlines of predicted lakes.
- model_ckp: a folder storing the TensorFlow checkpoints of the model (must be put in the same directory of the trian.py file).
- Predict.ipynb: a Python Jupyter Notebook file predicting the lake outlines. Only need to modify the third code section.
- Preprocess.ipynb: a Python Jupyter Notebook to crop the large satellite image into six regions for training and testing. Recommend crop images when the prediction is slow (See Note 1 for details).
- numpy version = 1.23.3
- pandas version = 1.4.4
- rasterio version = 1.3.8
- cv2 version = 4.6.0
- geopandas version = 0.14.0
- shapely version = 2.0.1
- tensorflow version = 2.10.0
- Download model_ckp and Predict.ipynb files to local.
- Check if the libraries are compatible.
- Change the path in the third code section to load evaluation images.
- Run the following codes to get the results.
- The estimated prediction time for all test regions is about 1 hour using an NVIDIA RTX A4500. Since the raw satellite is very large, it is recommended to split the raw image into 6 regions. We have done this process using Preprocess.ipynb and the cropped images are stored in a Google Drive for your convenience.
- One may need to change the naming function in the second from the last code section, since we use it to fill the Image and Region_Num columns in the geopackage file.
- Other preprocessing and training codes are available upon request.
- Zhihao Wang, zhwang1@umd.edu, University of Maryland
- Yiqun Xie, xie@umd.edu, University of Maryland
- Xiaowei Jia, xiaowei@pitt.edu, University of Pittsburgh