STACKED AUTOENCODERS FOR MULTICLASS CHANGE DETECTION IN HYPERSPECTRAL IMAGES

Change detection (CD) in multitemporal datasets is a key task in remote sensing. In this paper, a scheme to perform multiclass CD for remote sensing hyperspectral datasets extracting features by means of Stacked Autoencoders (SAEs) is introduced. The scheme combines multiclass and binary CD to obtain an accurate multiclass change map. The multiclass CD begins with the fusion of the multitemporal data followed by feature extraction by SAE. The binary CD is based on the spectral information by calculating pixel-wise distances and thresholding, and it also incorporates spatial information through watershed segmentation. The data coming from the multiclass CD is filtered by using the binary CD map and later classified by a Support Vector Machine or an Extreme Learning Machine algorithm. The scheme was evaluated over a multitemporal hyperspectral dataset obtained from the Hyperion sensor. Experimental results show the effectiveness of the proposed scheme using SAE for extracting the relevant features of the fused information when compared to other pub- lished feature extraction methods

All the images are avaiable in Matlab (.mat) format, among others. For further information see the readme in the files.

* Hermiston

* Codes were run in Ubuntu 14.04. * Caffe framework 1.0.0-rc3 to perform the feature extraction by means of SAE. We have configured the SAE to obtain 12 features. This is done by two consecutive layers that reduce the dimensionality of the data from 242 to 100 and from 100 to 12 features,respectively. 20% of the pixels were randomly chosen and used for training the SAE with a batch of 64 pixels per iteration and a limit of 300000 iterations. The back-propagation process uses a Stochastic Gradient Descent (SGD) and the ’inv’ learning rate policy corresponding to inv = base lr ∗ (1 + γ ∗ i) (−power) being i the iteration number and with a base learning rate (base lr) of 0.01, and values for the parameters γ and power of 0.0001 and 0.75 respectively.

* NWFE and PCA retaing 12 features.

* ELM and SVM trained with 5% of the reference data available for each class. Training samples randomly chosen in each run SVM classification carried out using the LIB-SVM library and the Gaussian radial basis function (RBF) ELM configured with a sigmoidal activation function.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.