HSKL: Hyperspectral-scikit-learn

Hyperspectral image analysis using scikit-learn

Installation

The package can be installed from pip:

pip install hskl

Usage

Training a pixel-level classifier for segmentation:

import os

from hskl.demo import dl_hyrank, load_hyrank
import hskl.classification as classification
import hskl.utils as utils

# Download, unpack, and load HyRANK dataset from current directory.
path = os.getcwd()
if not os.path.exists("HyRANK_satellite"):
    dl_hyrank(path)    
images, labels, _ = load_hyrank(path)

# Dimensional reduction using PCA, retain 99.9% image variance
pca = utils.pca_fit(images[0])
train, _ = utils.pca_apply(images[0], pca, 0.999)
test, _ = utils.pca_apply(images[1], pca, 0.999)
label = labels[0]
test_mask = labels[1]>0

# Train a classifier and predict test image labels
cl = classification.HyperspectralClassifier(
         method_name="LinearDiscriminantAnalysis")
cl.fit(train, label)
prediction = cl.predict(test)

# Visualization of training data, test prediction, and test ground truth
fig_objs_train = utils.overlay(train,label)
utils.save_overlay(fig_objs_train, "hyrank_train.png")

fig_objs_predict = utils.overlay(test,prediction*test_mask)
utils.save_overlay(fig_objs_predict, "hyrank_predict.png")

fig_objs_test = utils.overlay(test,labels[1])
utils.save_overlay(fig_objs_test, "hyrank_test.png")

Output:

Training image and ground truth labels:

Test image and ground truth labels:

Test image and predicted labels:

Notes:

1. Shape of train and test arrays are (DimX, DimY, SpectralChannels).
2. Shape of label and prediction arrays are (DimX, DimY).
3. Labeling convention for classifiers: (a) Datatype: label.dtype == np.uint8. (b) Labeled classes start from integer 1. Pixels with label == 0 are ignored (masked out).
4. Dimension(s) of train and label must be consistent: train.shape[0] == label.shape[0] and train.shape[1] == label.shape[1].
5. Inputs: train, test, and label can also be lists of np.ndarrays with each element satisfying the preceeding requirements.

Planned Features

In the near-term:

• Test scripts and data
• Grid search cross validation

In the long-term, support for:

• Pipelines
• Patch-based featurizer
• Dask-enabled parallelism
• Deep learning (PyTorch) models

References

Karantzalos, Konstantinos, Karakizi, Christina, Kandylakis, Zacharias, & Antoniou, Georgia. (2018). HyRANK Hyperspectral Satellite Dataset I (Version v001). Zenodo. http://doi.org/10.5281/zenodo.1222202

Some functionalities in this package are provided by Spectral Python (SPy): https://github.com/spectralpython/spectral