Remote Sensing and Geospatial Analysis for Landscape Pattern Characterization - Landscape Ecology for Sustainable Environment and Culture - page 208

explanatory power tends to be higher for a model that includes more independent
variables. This suggests that any meaningful comparison of model performance
should be based on the identical number of independent variables.
Finally, multiple regression analysis can make use of all or a subset of the
sample in parameter estimation when building a statistical model. Although most
of existing studies have built upon the use of all sample data in parameter esti-
mations, recent development in spatial statistics suggests using a subset of the
sample data can help reveal the variation of the cause-effect relationship across
space (Fotheringham et al. 2002 ). This localized regression technique called
geographically weighted regression has been included in some leading GIS soft-
ware packages, such as ArcGIS 10. Although the software tool is readily available,
it can be very much data demanding, particularly for some environmental data that
can be only acquired through in situ measurements.
11.3.2 Modeling and Predicting Landscape Dynamics
Models developed to simulate and predict landscape dynamics as a physical
process have become quite popular in recent years. This is necessary to understand
the dynamics of complex ecosystems and to evaluate the consequences of land-
scape change on the environment (Yang and Lo 2003 ; Sutherland 2006 ). Ecolo-
gists were among the earliest groups who have demonstrated a strong interest in
developing spatially explicit models to predict the impacts of different landscape
configurations on plant and animal populations (Kareiva and Wennergren ( 1995 ).
While many methods have been developed to predict the impacts at the species or
population level (c.f. Sutherland 2006 ), here we direct our attention on the spa-
tially explicit, dynamic models that are designed to work at the community or
ecosystem level.
Over the past several decades, a variety of spatially explicit models have been
developed by different communities, which can be either stochastic, such as the
logit (e.g. Hu and Lo 2007 ), Markov (e.g. Myint and Wang 2006 ), cellular auto-
mata (e.g. Hagen-Zanker and Lajoie 2008 ), and agent-based models (e.g. Robinson
et al. 2012 ), or processes based, such as dynamic ecosystem models (e.g. Eu-
skirchen et al. 2006 ). Although these models are different by their underlying
mechanism, they share many commonalities. The common approaches are the use
of transition probabilities in a class transition matrix, multinominal logit methods,
cellular or agent-based modeling, and GIS weighted overlay approach. These
models consider different constraints by various biophysical, economical, and
social parameters. Some of these parameters include land transition probabilities,
topography, environmental protection, forest properties, transportation, popula-
tion, economic indicators, human behavior, and policy. The role of remote sensing
and GIS is indispensible in the entire model development process from model
conceptualization to implementation that includes input data preparation, model
calibration, and model validation. Comprehensive reviews on various spatial