IEEE Geoscience and Remote Sensing Magazine - December 2017 - 14

Re Lo
gr gis
es tic
sio
n

g
in
ol

3
C -D
on
v.

g
in
ol

3
C -D
on
v.

Po

Po

Label

Neighborhood
of the Pixel
Hyperspectral
Vector
Image
FIGURE 5. A flowchart of the 3-D CNN architecture proposed in [19] for spectral-spatial hyperspectral image classification. Conv.: convolution.

UNSUPERVISED DEEP LEARNING
To allow less dependence on the existence of large annotated collections of labeled data, unsupervised feature
extraction is of great interest. The authors of [35] propose
an unsupervised convolutional network for learning spectral-spatial features using sparse learning to estimate the
network weights in a greedy layerwise fashion instead of
end-to-end learning. Mou et al. [33], [34] present a network
architecture called a fully r- esidual conv-deconv network for unsupervised spectral-spatial feature learning of hyperspectral images. They report an extensive study of the filters
learned (see Figure 6).

(a)

(b)

FIGURE 6. The object-detection maps using learned filters of

the first residual block in the unsupervised residual conv-deconv
network [33], [34], where some neurons own good description
power for semantic visual patterns at the object level. The feature
maps activated by the convolutional filter numbers 52 and 03, e.g.,
in the first residual block can be used to precisely capture (a) metal
sheets and (b) vegetative covers, respectively.

an end-to-end, band-adaptive spectral-spatial-featurelearning network to address the problems of the curse of
dimensionality. In [30], to allow a CNN to be appropriately
trained using limited labeled data, the authors present a
novel pixel-pair CNN to significantly augment the number
of training samples.
Following recent vision developments in 3-D CNNs [31],
in which the third dimension usually refers to the time axis,
such architecture has also been employed in hyperspectral
classification. In other words, in a 3-D CNN, convolution
operations are performed spatial spectrally, while in 2-D
CNNs, they are done only spatially. The authors in [19] introduce a supervised, , 2-regularized 3-D CNN-based model (see Figure 5), while the authors of [32] follow a similar
idea for spatial-spectral classification.
14

RECURRENT NEURAL NETWORKS FOR
HYPERSPECTRAL IMAGE CLASSIFICATION
In [36], the authors propose an RNN model with a new
activation function and modified gated recurrent unit
for -hyperspectral image classification that can effectively
analyze hyperspectral pixels as sequential data and then
determine information categories via network reasoning
(see Figure 7).
ANOMALY DETECTION
In a hyperspectral image, the pixels whose spectral signatures are significantly different from the global background
pixels are considered anomalies. Because the prior knowledge of the anomalous spectrum is difficult to obtain in
practice, anomaly detection is usually solved by background
modeling or statistical characterization for hyperspectral
data. So far, the only attempt to address this problem via
deep learning can be found in [37], where Li et al. propose an
anomaly detection framework in which a multilayer CNN is
trained using the differences in values between neighboring
pixel pairs in the reference image as input data. Then, in the
test phase, anomalies are detected by evaluating differences
between neighboring pixel pairs using the trained CNN.
In summary, deep learning has been widely applied
to multi/hyperspectral image classification, and some
promising results have been achieved. In contrast, for
other hyperspectral data analysis tasks, such as change
and anomaly detections, deep learning is just beginning
to make its mark [37], [38]. Some potential problems to
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2017
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017
