Mapping proglacial streams in HMA with Planet imagery

The objective is to classify small proglacial stream in High Mountain Asia (HMA) using high-resolution Planet imagery. Raw images are preprocessed from 16-bit to 8-bit false color channels (nir,r,g,b) and are sliced into 512×512 chips before placing into the image segmentation/classification methods. Postprocessing of classified chips include merging of tiles to return the same geospatial information from the raw Planet imagery.

Prerequisites

Core modules:

• Tensorflow
• OpenCV
• GDAL
• Rasterio

Conda Environment:

Setup the conda environment using the environment.yml file. Note: This environment will include installation of core modules above and other modules might need to be updated.

conda env create -f environment.yml

Implementation

Raw PlanetScope images should be preprocessed using chips.py (/master/utils/).

Use the labeled chips generated from annotation tools such as PixelAnnotationTool. The classification scripts calls the training and validation data following the structure below. Note: The pred folder is only used for full scene prediction tasks (see details in Full Scene Mapping section).

data
└───train
│   └───imgs
│   │   │   chip1.tif
│   │   │   chip2.tif
│   │   │   ....
│   └───masks
│       │   chip1.png
│       │   chip2.png
│       │   ....
└───pred
    └───raw_planet
        │   raw_fullscene_SR1.tif
        │   raw_fullscene_SR2.tif
        │   ....

This repo also includes PlanetAPI image lookup, order, and downloads (under /master/planetAPI/), PlanetScope raw image preprocessing, and water classifications implementations.

Sample Results

Some illustration of mapping results between the classification methods from the PlanetScope scenes in HMA. To implement water classification methods, run:

    a) NDWI Thresholding (Simple and Otsu) - thresh.py
    b) Random forest - rf.py
    c) Computer Vision (U-Net) - cv.py

Full scene multi-tile mapping within a PlanetScope strip in HMA using computer vision.

For full classification of raw PlanetScope imagery, run full_pred.py.

This will preprocess tha raw PlanetScope images inside the ./pred/raw_planet folder and identify water pixels using a pre-trained cv model for HMA (under ./log/cv_mul/cv_multi.hdf5). See the referece below for more info about the cv model.

By default, this will create a ./pred_out folder where it will write the water mask outputs from the cv model.

data
└───pred
    └───raw_planet
        │   raw_fullscene_SR1.tif
        │   raw_fullscene_SR2.tif
        │   ....
pred_out
    │   raw_fullscene_SR1_mask.tif
    │   raw_fullscene_SR2_mask.tif
    │   ....

Reference

To read more about this work or if you use this repository and find it helpful, please read/cite the article:

Flores, J. A., Gleason, C. J., Brinkerhoff, C. B., Harlan, M. E., Lummus, M. M., Stearns, L. A., & Feng, D. (2024). Mapping proglacial headwater streams in High Mountain Asia using PlanetScope imagery. Remote Sensing of Environment, 306, 114124. https://doi.org/10.1016/j.rse.2024.114124

If you have any questions or suggestions about this repo, please contact jflores@umass.edu.