Classification of Hyperspectral Satellite Image Using Deep Convolutional Neural Networks. This is re-implementation of the paper
[1] K. Makantasis, K. Karantzalos, A. Doulamis and N. Doulamis, "Deep supervised learning for hyperspectral data classification through convolutional neural networks," 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan, 2015, pp. 4959-4962.
-
For each non-zero labelled pixel, we extract 5 x 5 x c neighbourhood and corresponding label.
-
Dimensionality reduction using PCA is performed. Final dimension is 5 x 5 x cr.
-
Training using CNN is performed with the following architecture: conv1-conv2-conv3-conv4-hidden1-hidden2-16way-softmax
-
Training : testing split ratio is maintained at 0.8 : 0.2
Table 1 : Comparison of accuracy for various classification methods
| Dataset | No. of Components | RBF-SVM | CNN [1] | Our CNN |
|---|---|---|---|---|
| Indian Pines | 30 | 82.79 | 98.88 | 98.94 |
| Pavia University | 10 | 93.94 | 99.62 | 99.66 |
- Data preparation : Matlab (Mat file)
-Download publicly available data mat files from following link and place them in /Matlab-Sat-Data/data/
http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes
-Then run /Matlab-Sat-Data/script_prep_data.m
- CNN classification : Theano + Lasagne+ Nolearn
-Run train.py for training and testing accuracy
For PCA, this matlab file exchange implementation was used: https://jp.mathworks.com/matlabcentral/fileexchange/38300-pca-and-ica-package/content/pca_ica/myPCA.m
MIT License Copyright (c) 2016 Subhajit Chaudhury