Keywords

1 Introduction

Hyperspectral image classification is an important research topic in remote sensing. In the presence of commercial hyperspectral sensors e.g. Airborne Visible/Infrared Imaging Spectrometer (AVIRIS), HSI data is easily available to researchers. AVIRIS which is operated by the NASA Jet Propulsion Laboratory covers 224 continuous spectral bands across the electromagnetic spectrum with a spatial resolution of 3.7 m. The information collected by AVRIS is used to classify the objects on earth surface. Supervised or unsupervised classification algorithms have the ability to quickly obtain categorical information from remote-sensing images and classify the objects present in the image. Consequently, such algorithms play an important role in remote-sensing image applications.

The basic purpose of image classification is to classify the labels for each pixel in HSI image, which is a challenging task. The performance of classification techniques is closely affected by high dimensionality of the data, limited labeled samples and spatial variability of spectral information. To overcome such issues, various techniques, such as independent component analysis (ICA) [1], neighborhood preserving embedding [2], linear discriminant analysis (LDA) [3] and wavelet analysis [4], have been proposed for the classification of hyperspectral images. Investigations show that the afore-mentioned techniques did not bring significant improvement in classification accuracy. However, support vector machine(SVM) based methods and Neural networks(NN) present a more attractive solution to image classification in terms of computational cost and classification accuracy [5]. Due to the high diversity of HSI data, it is difficult to determine which feature is more relevant for the classification task.

Moreover, recently introduced deep learning (DL) models automatically learn high-level features from data in a hierarchical manner. Typical deep learning models includes Deep Belief Networks [6], Deep Boltzmann Machines [7], Stacked Denoising Autoencoders [8] and Convolutional Neural network (CNN) [9]. More specifically Autoencoders (AE) [10] has been an efficiently used for the classification of HSI images, basically the input of Autoencoders (AE) is high dimensional vector i.e. flatten the high dimensional image into a vector then feed it to the model later classify it by using logistic regression classifier. A recent state-of-the-art technique proposed by Lee et al. [11], called a contextual deep CNN, consist of nine layers in total, jointly obtained the spatio-spectral features maps and classified by Softmax activation function.

In a similar fashion inspired by [11], in this paper we try to assess the effectiveness of a DL technique namely, Convolutional Neural network (CNN). The basic motivations for us to consider Convolutional approach have two main reasons: the effectiveness of this approach recently proved in numerous remote sensing applications; main characteristics of this technique, which makes it a potential candidate to classify hyperspectral data. In this context, we proposed a Conventional Multi-Layer Perceptron (MLP) network for the classification of remote sensing hyperspectral data. Our proposed structure basically combines the spectral-spatial attributes in initial stage resulting in a high-level spectral-spatial features construction and then implement MLP classifier for probabilistic multiclass HSI classification.

The rest of the paper is organized as follows: In Sect. 2, we provide details of the proposed network. The description of datasets and performance comparison are given in Sect. 3. Finally, Sect. 4 summarizes the process and some probable future work is pointed out.

2 Proposed Architecture

In this section architecture of the proposed system is briefly described. In the first stage the reduction of dimensionality is presented and then the deep structure of CNN and MLP is described.

2.1 Dimensionality Reduction

Usually, HSI data consist of several band/channels along the spectral dimension. Thus, it always has tens of thousands of dimensions resulting in a large amount of redundant information. In most of the cases, the first few band/channels have significant variance and they contain almost 99.9% of information [12]. So in the first layer of our proposed network we introduced PCA, to reduce the dimension to an acceptable scale while reserving the useful spatial information in the meantime. As our main concern is to incorporate the spatial information, so we use PCA along-with the spectral dimension only and retain first several principal components. During our experimentation process on state-of-the-art hyperspectral datasets, we used only 10 to 30 principal components respectively for each dataset.

2.2 Classification Framework

For CNN, Image input data is expressed as a 3-dimensional matrix of width * height * channels (h * w * c). In order to input an HSI image, we have to decompose HSI into patches, each one of which contains spectral and spatial information for a specific pixel. Our proposed network contains 12 convolutional layers. First convolutional layer in network contains 32 features with a filter whose dimension is 3 * 3. The batch size of 30 samples is used and the block size is set to 11. In first convolutional layer, we use a filter of dimension 3 * 3 and get feature maps in subsequent layers as shown in Fig. 1. In a similar manner for further layers filter size remains same but the number of feature maps is increased. For preserving local spatio-spectral correlation we do not increase the filter size. The first convolutional layer is followed by further hidden layers in the network.

Fig. 1.
figure 1

Filter size and feature map representation.

Full size image

During the training, network parameters keep changing repeatedly which cause a change in activations, this refers to as “internal covariate shift”. To resolve this problem we adopt Batch normalization (BN) [13] which allows us to use much higher learning rate.

figure a

The algorithm given above presents Batch normalization (BN) transforms where \( \beta = \left\{ {x_{1} \ldots x_{m} } \right\} \) are the values over mini-batch. Equation (3) implements normalization operation while Eq. (4) implements scaling and shifting learned by γ and β parameters to get the final result \( y_{i} \). The main characteristic of BN is that it is based on simple differentiable operations, which can be inserted anywhere in CNN network to normalize improper network initialization. BN boost up the performance as well.

After convolving the image fed the neurons to max-pooling layer, the purpose is to take the maximum values from the input and shorten the size of selected features. The pool size is 2 * 2. Next, pooling layer is followed by the Flatten layer which converts the 2D matrix to a vector called Flatten. It allows the output to be processed by standard fully connected layers. ReLU (Rectified linear unit) and dropout are also employed here. The threshold value for dropout is 0.3. The purpose of using ReLU is that it is much faster than other nonlinear functions and Dropout is used to prevent overfitting and complex co-adoptions phenomena.

For classification purpose Softmax activation [14] function issued to output probability-like predications according to the number of classes. Softmax is a generalization of logistic function, and its output can be used to represent the categorical distribution, which is basically a gradient-log-normalizer:

$$ p\left( {y = j|z^{(i)} } \right) = \phi_{soft\,\,\hbox{max} } \left( {z^{(i)} } \right) = \frac{{e^{{z^{(i)} }} }}{{\sum\nolimits_{j = 0}^{k} {e^{{z_{k}^{\left( i \right)} }} } }} $$
(5)

where \( z \) is the net input can be defined as

$$ z = w_{0} x_{0} + w_{1} x_{1} + \ldots + w_{m} x_{m} = \sum\limits_{l = 0}^{m} {w_{l} x_{l} } = \varvec{w}^{T} \varvec{x} $$
(6)

where \( w \) is the weight vector, \( w_{0} \) is for bias and \( x \) is the feature vector. \( z^{(i)} \) is basically a classification function of \( j - th \) class which takes “x” as an input and compute probability “y” for each class label. Therefore, Softmax is adopted here because it is a potential candidate for probabilistic multiclass HSI classification problem.

Stochastic gradient descent (SGD) is a classical approach for training deep learning architecture is employed here. SGD algorithm is used to calculate the error and propagate it back to adjust the MLP weights and filters. The architecture of our proposed approach is presented in Fig. 2.

Fig. 2.
figure 2

Graphical representation of proposed network for HIS classification.

Full size image

3 Experimental Results and Comparative Analysis

3.1 Datasets

AVIRIS and ROSIS sensor datasets are the classical datasets [15]. Particularly, in our experiment the Indian Pines, Salinas and Pavia university datasets are used. Indian Pines dataset depicts a test site in North-western Indiana and consists of 145 * 145 pixels with 224 spectral reflectance bands in the wavelength range from 0.4 to 2.5 µm while spatial resolution is 20 m. Basically, it contains 16 classes but we only use 8 classes because they have a large number of samples among others.

The University of Pavia dataset depicts the scenes acquired by the ROSIS sensor during a flight campaign over Pavia, northern Italy whose number of spectral bands are 102 contains 610 * 340 pixels. It contains 9 classes.

The number of spectral bands and spatial resolutions are 103 and 1.3 m respectively. While the spectral reflectance range from 0.4 to 0.8 µm.

Third dataset “Salinas” is also acquired by AVIRIS sensor over Salina Valley, California. It consists 224-bands with 512 * 217pixels with high spatial resolution 3.7 m. Number of classes of this data set are 16. For both datasets (University of Pavia, Salinas) we use all the classes for training and testing because they have a relatively large number of samples. For all datasets, selected classes and samples are listed in Tables 1, 2 and 3.

Table 1. Number of training and testing samples along with selected classes used from the Indian Pines DataSet.
Full size table
Table 2. Number of training and testing samples along with selected classes used from the University of Pavia DataSet.
Full size table
Table 3. Number of training and testing samples along with selected classes used from salinas dataset.
Full size table

3.2 Comparative Analysis

For comparison, we randomly select 200 samples per class for training and all remaining samples for testing. The basic purpose of selecting 200 samples per class is to evaluate our proposed method with the state of the art approaches reported in [11]. To successfully accomplish all the experiments the CNN Tensor flow framework [16] is used on GPU GTX1060.

Table 4 provides a comparative analysis of classification among the proposed method and the one reported in [11]. The contextual deep CNN used in [11] has 9 convolutional layers while our proposed network has twelve layers, we can say that our network is much deeper than contextual deep CNN [11]. It is obvious that our network has much better performance as compare to contextual deep CNN on all datasets. To further evaluate our network we compare our performance with state-of-the-art RBF kernel-based SVM method [17], which consist two convolutional and two fully connected layer much shallower than our technique. In recent research [18], for a diversified Deep Belief Networks(D-DBN) has much better performance as compared to [17], we also use (D-DBN) as a baseline to in our comparative analysis. For all the datasets, we also use other types of methods which are evaluated in [11]: two-layer NN, three-layer NN, shallower CNN and LeNet-5.

Table 4. Classification accuracy comparison among proposed networks and the base lines on three datasets(%). The best performances among all methods are indicated in bold
Full size table

Our proposed network out-performs the baseline approaches on all the datasets. More specifically as compared to [11] for Indian Pines dataset the proposed network gained more than 2% accuracy while in the cases of University of Pavia and Salinas datasets, it gained 1.3% and 2.04% classification accuracy respectively. The significant performance of proposed architecture is just because of its deeper nature which proves, that digging more in the convolutional network leads to high classification accuracy. Figure 3 shows the classification maps of each data set corresponding to their ground truth images.

Fig. 3.
figure 3

RGB compositions of resulted classification map by proposed network along with ground truth are shown for University of Pavia, Salinas and Indian Pines datasets.

Full size image

3.3 Impact of Epochs

During network training weights are updated due to back propagation phenomena, One round of updating the network or the entire training dataset is called an epoch [19]. Figure 4 shows validation loss and classification accuracy on the bases of epoch size. From validation loss plotted in Fig. 4a we observe the performance of the proposed network i.e. the number of lost samples decreased when the number of epochs increased meanwhile the classification accuracy is improved significantly as can be seen in Fig. 4b.

Fig. 4.
figure 4

Classification performance under different set of parameters for all experimental data sets. (a)Validation loss vs. Number of epochs, (b) Classification accuracy over the course of epochs.

Full size image

For all the data sets these observations proved that deepness of our network greatly improves overall accuracy meanwhile preserving lower validation loss.

4 Conclusion

In this letter, we propose a CNN-based classification method for remote sensing data. The proposed method is much deeper, faster and utilizes more spatio-spectral features for the classification of hyperspectral images. The proposed method and existing state-of-art techniques are compared using three data sets. It is shown that our method achieves better classification accuracy. Simulation results demonstrate the superiority of the proposed method. The future research prospects include to combine the proposed network with a shallower convolutional based network for more enhanced classification performance.