Peer-Reviewed Articles
1201 Mapping Selective Logging in Mixed
Deciduous Forest: A Comparison of Machine
Learning Algorithms
Christopher D. Lippitt, John Rogan, Zhe Li, J. Ronald Eastman, and Trevor G. Jones
Abstract Download
Full Article
This study assesses the performance of five Machine Learning
Algorithms (MLAs) in a chronically modified mixed deciduous
forest in Massachusetts (USA) in terms of their ability to detect
selective timber logging and to cope with deficient reference
datasets. Multitemporal Landsat Enhanced Thematic Mapperplus
(ETM+) imagery is used to assess the performance of three
Artificial Neural Networks - Multi-Layer Perceptron, ARTMAP,
Self-Organizing Map, and two Classification Tree splitting
algorithms: gini and entropy rules. MLA performance evaluations
are based on susceptibility to reduced training set size,
noise, and variations in the training set, as well as the
operability/transparency of the classification process. Classification
trees produced the most accurate selective logging
maps (gini and entropy rule decision tree mean overall map
accuracy = 94 percent and mean per-class kappa of 0.59 and
0.60, respectively). Classification trees are shown to be more
robust and accurate when faced with deficient training data,
regardless of splitting rule. Of the neural network algorithms,
self-organizing maps were least sensitive to the introduction of
noise and variations in training data. Given their robust
classification capabilities and transparency of the class selection
process, classification trees are preferable algorithms
for mapping selective logging and have potential in other
forest monitoring applications.
1213 Subpixel Urban Land Cover Estimation:
Comparing Cubist, Random Forests,
and Support Vector Regression
Jeffrey T. Walton
Abstract Download Full Article
Three machine learning subpixel estimation methods
(Cubist, Random Forests, and support vector regression)
were applied to estimate urban cover. Urban forest canopy
cover and impervious surface cover were estimated from
Landsat-7 ETM+ imagery using a higher resolution cover
map resampled to 30 m as training and reference data.
Three different band combinations (reflectance, tasseled cap,
and both reflectance and tasseled cap plus thermal) were
compared for their effectiveness with each of the methods.
Thirty different training site number and size combinations
were also tested. Support vector regression on the tasseled
cap bands was found to be the best estimator for urban
forest canopy cover, while Cubist performed best using the
reflectance plus tasseled cap band combination when
predicting impervious surface cover. More training data
partitioned in many small training sites generally produces
better estimation results.
Color Figures (Adobe PDF format):
[figure 1.] [figure 2.] [figure 3.] [figure 6.]
1223 Bayesian Learning with Gaussian Processes
for Supervised Classification of
Hyperspectral Data
Kaiguang Zhao, Sorin Popescu, and Xuesong Zhang
Abstract Download
Full Article
Recent advances in kernel machines promote the novel use of
Gaussian processes (GP) for Bayesian learning. Our purpose is
to introduce GP models into the remote sensing community for
supervised learning as exemplified in this study for classifying
hyperspectral images. We first provided the mathematical
formulation of GP models concerning both regression and
classification; described several GP classifiers (GPCLs) and the
automatic learning of kernel parameters; and then, examined
the effectiveness of GPCLs compared with K-nearest neighbor
(KNN) and Support Vector Machines (SVM). Experiment results
on an Airborne Visible/Infrared Imaging Spectroradiometer
image indicate that the GPCLs outperform KNN and yield
classification accuracies comparable to or even better than
SVMs. This study shows that GP models, though with a larger
computation scaling than SVM, bring a competitive tool for
remote sensing applications related to classification or
possibly regression, particularly with small or moderate sizes
of training datasets.
1235 Neuro-fuzzy Based Analysis of Hyperspectral
Imagery
Fang Qiu
Abstract Download
Full Article
A neuro-fuzzy system, namely Gaussian Fuzzy Learning
Vector Quantization (GFLVQ), was developed based on the
synergy of a neural network and a fuzzy system. GFLVQ is
both a fuzzy neural network and a neural fuzzy system
with supervised learning and unsupervised self-organizing
capabilities. In this paper, GFLVQ was further improved to
efficiently and effectively process hyperspectral data through
training data informed initialization and a simplified fuzzy
learning algorithm. A geovisualization tool was developed
to facilitate knowledge discovery and understanding of the
hyperspectral image. A case study was conducted using
a Hyperion image. The results obtained from the improved
neuro-fuzzy system were found to be significantly
better than those from conventional statistics-based and
endmember-based classifiers. The fuzzy spectral profiles
produced from the geovisualization tool provided an extra
insight into the neuro-fuzzy learning process, further
opening up the black box of the neural network.
Color Figures (Adobe PDF format):
1249 Hybrid Segmentation – Artificial Neural
Network Classification of High Resolution
Hyperspectral Imagery for Site-Specific
Herbicide Management in Agriculture
P.R. Eddy, A.M. Smith, B.D. Hill, D.R. Peddle, C.A. Coburn, and R.E. Blackshaw
Abstract Download
Full Article
Site-Specific Herbicide Management (SSHM) in Precision
Agriculture (PA) requires weed detection in crop fields for
directed herbicide application instead of spraying entire fields.
This has significant economic and environmental advantages
for improved agricultural practices that are essential given
forecast increases in global population and food production
needs. In this study, a new hybrid segmentation - Artificial
Neural Network (HS-ANN) method was compared to standard
Maximum Likelihood Classification (MLC) for improving
crop/weed species discrimination in SSHM/PA. Very high
spatial resolution (1.25 mm) ground-based hyperspectral
image data were acquired over field plots of canola, pea, and
wheat crops seeded with two weed species (redroot pigweed,
wild oat) in southern Alberta, Canada. The very high spatial
and spectral resolution image data required development of
a simple yet efficient vegetation index (Modified Chlorophyll
Absorption in Reflectance Index (MCARI)) threshold segmentation
to separate vegetation from soil for classification. The HSANN
consistently outperformed MLC in both single date and
multi-temporal classifications. Higher class accuracies were
obtained with multi-temporally trained ANNs (84 to 92 percent
overall), with improvements up to 31 percent over MLC. With
advancements in imaging technology and computer processing
speed, this HS-ANN method may constitute a viable farm
option for real-time detection and mapping of weed species
for SSHM in agriculture.
1259 Global Optimization versus Deterministic
Pruning for the Classification of
Remotely Sensed Imagery
D. Stathakis and I. Kanellopoulos
Abstract Download
Full Article
The effect of pruning neural network structures in remote
sensing is investigated. Standard pruning methods, i.e.,
Optimal Brain Damage and Optimal Brain Surgeon, are
compared with pruning based on a genetic algorithm. Direct
coding is used to represent the links of the network for
optimization with a canonical genetic algorithm using binary
representation. The results show that the genetic algorithm
is the only method able to discover a significantly better
neural network structure. The main drawback of the genetic
approach is the extensive training time required.
1267 Genetic Algorithms for the Calibration of
Cellular Automata Urban Growth Modeling
Jie Shan, Sharaf Alkheder, Jun Wang
Abstract Download
Full Article
This paper discusses the use of genetic algorithms to enhance
the efficiency of transition rule calibration in cellular
automata urban growth modeling. The cellular automata
model is designed as a function of multitemporal satellite
imagery and population density. Transition rules in the model
identify the required neighborhood urbanization level for a test
pixel to develop to urban. Calibration of the model is initially
performed by exhaustive search, where the entire solution
space is examined to find the best set of rule values. This
method is computationally extensive and needs to consider all
possible combinations for the transition rules. The rise in the
number of variables will exponentially increase the time
required for running and calibrating the model. This study
introduces genetic algorithms as an effective solution to the
calibration problem. It is shown that the genetic algorithms
are able to produce modeling results close to the ones
obtained from the exhaustive search in a time effective
manner. Optimal rule values can be reached within the early
generations of genetic algorithms. It is expected that genetic
algorithms will significantly benefit urban modeling problems
with larger set of input data and bigger solution spaces.