cells. These independent estimates provide a 
source of useful data for landscape level 
monitoring. However, the estimates from the two 
systems are made with different definitions of 
forest, in different years, and can be 
contradictory. Differences in schedules among 
independent monitoring systems can be 
accommodated by annual estimates (Section 7); 
different classification systems can be 
accommodated by calibration. 
Calibration for misclassification error in remote 
sensing requires plots for which reference and 
remotely sensed classifications are known 
(Section 3). Calibration for "misclassification” 
error caused by differences in classification 
systems requires plots which are independently 
classified by two independent monitoring systems. 
If sample units from other monitoring systems are 
accurately registered to the 40 km 2 sample units, 
then multivariate calibration models can estimate 
the quantitative statistical relationship between 
areal estimates from another agency, and areal 
estimates from the landscape monitoring system. 
These estimates that are "calibrated" for 
differences in definitions might be further 
calibrated using the calibration model for remote 
sensing errors. 
This would allow several agencies to share areal 
estimates applicable to aggregations of sample 
units, while maintaining their own classification 
systems. These shared estimates might be made at 
the level of individual 640 km 2 cells using small 
area estimation, which takes the form of a 
composite estimator. Shared statistical 
estimates might improve the efficiency and 
compatibility of participating monitoring 
systems, without major disruptions to any one 
existing system. However, statistical 
calibration is not a panacea. Calibration will 
propagate statistical errors (Section 3.4), but 
these can be are minimized by making the 
independent classification systems as compatible 
as possible. 
8.2 The cell as a stratum with sample size one 
A stratum is a contiguous, homogeneous geographic 
area. However, calibration models superimpose 
additional stratification criteria, such as 
Landsat scene boundaries, or sets of sample units 
photointerpreted by one individual. These 
differences are needed to correct for different 
misclassification probabilities, and these 
criteria can change over time. When ancillary 
data from other monitoring systems are combined, 
a stratum is further subdivided by the geographic 
criteria used by each other system It is likely 
the number of strata will eventually approach the 
number of sample units. 
The estimated status of a 640 km 2 cell might be 
considered a combination of the estimated status 
of the one 40 km 2 sample unit it that cell, and 
ancillary estimates from other agencies, which 
apply to aggregations of cells (Section 8.1). 
Estimation error associated with each 40 km 2 
sample unit includes propagated and correlated 
errors from a regional calibration model (Section 
5), propagated and correlated prediction errors 
from a regional deterministic prediction model 
(Section 7), and sampling error from use of one 
40 km 2 sample unit in the cell. Sampling error 
might be estimated using aggregations of 40 km 2 
plots and assuming independence and homogeneity, 
or geostatistical methods, such as Kriging and 
spatial correlograms. 
9. LANDSCAPE DETECTION AND EVALUATION MONITORING 
One objective might be monitoring "environmental 
health." "Detection" monitoring might use 
quantitative indicators of response and exposure 
to classify each 40 km 2 sample unit as "healthy" 
or "unhealthy". Unhealthy sample units could be 
further subclassified as to probable cause during 
"evaluation monitoring". Sample units classified 
based on their health can be used to make areal 
estimates of environmental health for regional 
assessments. Therefore, there is interest in 
individual sample units that might not be 
necessary if statistical estimates of regional 
status were the sole objective. 
This is analogous to a psychologist’s judgment 
(i.e., detection) whether a patient in a random 
sample (i.e., a sampling unit) is mentally ill 
(i.e., unhealthy) based on blood chemistry and 
psychological profile tests (i.e., response 
indicators), and history of chemical abuse or 
family mental health problems (i.e., exposure 
indicators); diagnosing probable cause(s) for the 
patient’s condition (i.e., evaluation); and 
making an estimate of the suspected prevalence of 
various types of mental illnesses in the 
population (i.e., assessment) using a large 
sample of patients. 
Quantitative indicators are needed to identify 
unhealthy sample units. Causal hypotheses might 
be suggested by exploratory statistical methods, 
such as scatter plots or principal components 
analyses, or geostatistical methods that might 
show similar spatial associations in unhealthy 
sites and indicator values. Hypotheses might be 
more difficult to formulate if landscape 
processes are nonlinear, with time lags and 
feedback mechanisms that obscure direct cause and 
effect relationships. Process oriented 
deterministic models contain a collection of 
individual hypotheses regarding landscape 
structure and function. If exposure indicators 
associated with individual sample units are 
included among driving variables for a landscape 
level model, and the model can predict response 
indicators that are measured on sample units, 
then aggregate hypotheses in the deterministic 
model can be scientifically tested. 
The residual difference between model predictions 
and direct observations represents model 
prediction error, i.e., lack of agreement in 
predicting measurements of landscape structure 
and function. A model and direct measurements 
are imperfect caricatures of a system, and 
prediction errors are expected. However, 
residuals are expected to be random if the model 
and" measurements are reliable. If spatial or 
temporal patterns exist in the residuals, then 
important processes are not included in the 
model, or there are unrecognized problems with 
the measurement process. 
Such an unexpected situation should trigger a 
search for hypotheses that might explain the 
apparent nonrandom patterns. If the prediction 
model, rather than measurements, is judged to be 
the problem, alternative hypotheses might be 
incorporated in the prediction model, and tested 
with independent monitoring data. Therefore, 
analysis of data from a landscape monitoring 
system, and predictions from a landscape model, 
can be a crucial step in the cycle of hypotheses 
development, hypothesis testing, and hypotheses 
refinement to help understand the condition and 
functioning of landscapes. 
377