IEEE Geoscience and Remote Sensing Magazine - March 2014 - 24

techniques and making it possible to address new important
and challenging applications. The potentials of the technological development are strengthen from the increased
awareness of the importance of monitoring the Earth surface
at local, regional and global scale. Assessing, monitoring and
predicting the dynamics of land covers and of antrophic processes is at the basis of both the understanding of the problems related to climate changes and the definition of politics
for a sustainable development. Nonetheless, the properties
of the images acquired by the last generation sensors pose
new methodological problems that require the development
of a new generation of methods for the analysis of multitemporal images and temporal series of data.
After a general overview of the problems related to the
analysis of multitemporal images and time series of data,
this talk will focus on the very important problem of automatic change detection between multitemporal images. The
development and the use of effective automatic techniques
for change detection are of major importance in many of the
above-mentioned application scenarios. The increased geometrical resolution of multispectral and SAR sensors, the
increased revisit time of high resolution systems, and the
expected availability of time series of hyperspectral images
in the near future result in many different methodological
problems as well as in very important new possible applications. The talk will address these problems by pointing out
the state of the art and the most promising methodologies
for change detection on images acquired by the last generation of satellite sensors. Examples of the use of change-detection approaches in operative scenarios will be provided.
The presentation can be tuned on request on different
kinds of target audience: 1) students; 2) remote sensing
scientists; 3) scientists expert in data analysis.
AdvAnced Methods for clAssificAtion
of hyperspectrAl dAtA
Melba Crawford
Accurate land cover classification that
ensures robust mapping under diverse
conditions is important in environmental studies where the identification of the land cover changes and its
quantification have critical implications for management practices, ecosystem health, and the impact of climate.
Hyperspectral data provide enhanced capability for more
accurate discrimination of land cover, but significant challenges remain for classification, including highly correlated
spectral bands, high dimensionality, and nonlinear spectral
response in nonstationary environments. Advanced methods in machine learning, including nonlinear manifold
learning, semi-supervised learning, and active learning are
promising for classification of hyperspectral data.
Nonlinear global and local manifold learning methods provide natural capability to both accommodate
nonlinear scattering and practical, robust feature extrac24

tion methods in dynamic environments. Adaptive semisupervised approaches train the classifier with labeled
samples in one location/time and adapt supervised classifiers to samples in spatially disjoint areas or at different
times where samples exhibit significantly different distributions [Kim and Crawford 2010]. Active learning techniques that focus on developing informative training sets
with minimal redundancy have been demonstrated to
promote greater exploitation of the information in both
labeled and unlabeled data, while significantly reducing
the cost of data collection [Di and Crawford 2011]. New
developments for feature extraction via manifold learning, semi-supervised classification, and active learning of
hyperspectral data are outlined and demonstrated using
airborne and space-based hyperspectral data.
AdvAnced neurAl AdAptive processing
in interferoMetric And polAriMetric
rAdAr iMAging
Akira Hirose
This talk presents and discusses
advanced neural networks by focusing on complex-valued neural networks (CVNNs) and their applications in the remote sensing and
imaging fields. CVNNs are suitable
for adaptive processing of complexamplitude information. Since active
remote sensing deals with coherent electromagnetic
wave, we can expect CVNNs to work more effectively than
conventional neural networks or other adaptive methods
in real-number space. Quaternion (or Hypercomplexvalued) neural networks are also discussed in relation to
polarization information processing.
The beginning half of the Talk is devoted to presentation of the basic idea, overall framework, and fundamental treatment in the CVNNs. We discuss the processing
dynamics of Hebbian rule, back-propagation learning, and
self-organizing map in the complex domain. The latter half
shows some examples of CVNN processing in the geoscience and remote sensing society (GRSS) fields. Namely, we
present distortion reduction in phase unwrapping to generate digital elevation model (DEM) from the data obtained
by interferometric synthetic aperture radar (InSAR). In
polarization SAR (PolSAR), we apply quaternion networks
for adaptive classification. Another example is ground penetrating radar (GPR) to visualize underground objects to
distinguish specific targets in high-clutter situation. Finally
we discuss the prospect of the CVNNs in the GRSS fields.
Modeling, siMulAtion, inversion And
chAng'e dAtA vAlidAtion for MicrowAve
observAtion in chinA's lunAr project
Ya-Qiu Jin
In China's first lunar exploration project, Chang-E 1 (CE-1), a
multi-channel microwave radiometer in passive microwave
ieee Geoscience and remote sensing magazine

march 2014

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2014