in temperate 
s. In the US, 
y as October, 
ius forming a 
ics, there are 
th one or two 
ically. It has 
nza seasonal 
ve across the 
al migration 
of influenza 
nd migration 
ther study in 
ulation state 
ls southward 
lonso et al., 
among the 
in influenza 
NTS 
iting disease 
cal Rainfall 
TRMM is a 
in built the 
raft in 1999. 
[ microwave 
), Cloud and 
hting Image 
cipitation is 
neasurement 
m resolution 
igher spatial 
arid regions 
| growth is 
with plenty 
oth provided 
'oradiometer 
g the visible 
a and Aqua 
owever, can 
ith red and 
nces in band 
orbits and 
rs must first 
1 addition to 
itellites also 
tification of 
| risks. For 
of satellites 
Nssion and 
anced Land 
nultispectral 
Microwave 
id. Synthetic 
as that are 
geoscience 
and ranging 
' vegetation 
n imagery, 
rldView can 
4. MODELING TECHNIQUES AND EXAMPLES 
Many factors including environmental and meteorological 
conditions affect the transmission of malaria, dengue and 
seasonal influenza. The environmental and meteorological 
conditions can indeed be considered the driving factors for 
these diseases when other factors are stable. This is especially 
evident for vector-borne diseases where the vector propagation 
is directly influenced by the environmental and meteorological 
conditions. This is the essential premise of why remote sensing 
can be used to predict disease risks. In general, statistical and 
biological models can be used to predict disease risks. Both 
types of models accept remotely sensed environmental and 
meteorological parameters as input. In some applications, 
remote sensing not only can be used for predicting risks, but 
also for detecting and reducing risks. In addition, remote 
sensing-based model can be used to project disease risk under 
the impact of global warming. 
In general, statistical and mechanistic, or processed-based 
approaches are used to model disease risks based on satellite 
observations of meteorological and environmental parameters. 
In the statistical approach, epidemiological data are correlated 
with satellite data. The unknown parameters in the models are 
determined using statistical goodness of fit criteria, such as 
mean squared errors or Akaike Information Criteria (AIC). 
Once the model is trained, it can then be applied to other 
situations than those used for deriving the model parameters. 
How the pathogens actually transmit the disease under different 
meteorological and environmental conditions is not explicitly 
modelled in this approach. The common methods in this 
category include regression, time series analysis, and neural 
network. Examples are given in the following for using neural 
network to model malaria cases in Thailand and dengue cases in 
Indonesia; using Autoregressive Integrated Moving Average 
(ARIMA) to model influenza in Hong Kong; and using a 
discrete event simulation model to simulate malaria prevalence 
among 23 households. 
4.1 Malaria in Thailand 
Figure 1 shows an example for using neural network to predict 
malaria cases for the border provinces in Thailand (Kiang et al., 
2006). The objective is to predict malaria cases in the near 
future in order to forewarn public health stakeholders on the 
expected transmission intensity. The main meteorological and 
environmental parameters used in modelling include 
precipitation, NDVI, and surface temperature. Excellent 
agreement between the actual and hindcast case rates is seen. 
  
  
  
  
  
  
Figure 1. Actual (left) and predicted (right) malaria 
case rates in Thailand 
4.2 Dengue in Indonesia 
Figure 2 shows the results for using ARIMA to model the 
dengue cases in Jakarta, Indonesia using TRMM data and dew 
point temperature. All data except the last 12 months were used 
for the training. The last 12 months of data were used for 
deriving prediction accuracy. Close association between the 
actual and prediction distributions can be seen. Inconsistency 
between model output and the data at the final year is 
potentially due to the vector control effort that was 
implemented at the beginning of year 2005 and not accounted 
for in the model. 
  
8000 
» Data -— ARIMA Output 
E 
7000 
6000 
5000 
4000 
3000 
Dengue monthly cases 
2000 4: 
     
1000 | | 4 
  
  
SQ PS am & 
E E 
GA" Qe Wm uem 
of 
Av ur Y N 
SU m uS ow 
a 
Figure 2. Modelled dengue cases in Jakarta, Indonesia 
4.3 Influenza in Hong Kong 
Figure 3 shows the results for using ARIMA and radial basis 
function neural network to model seasonal influenza cases in 
Hong Kong (Soebiyanto et al, 2010). Data from the last 
influenza season was used for testing modelling accuracy. The 
rest of the data were used for training. Input to these models 
include land surface temperature, rainfall, air and dew point 
temperature. 
  
— DATA mo ARIMA FIT 
7 -o-ARIMA FORECAST ~~ = RBFNN FIT 
^ RBFNN FORECAST 
Uv o 
Weekly Positive Influenza 
{log scale) 
E 
  
  
  
  
Figure 3. Modelled influenza cases in Hong Kong 
using ARIMA and radial basis function neural network 
4.4 Prevalence of Malaria in a Cluster of Households 
Figure 4 shows the expected malaria prevalence rate among 23 
households and 92 residents using a discrete event simulation 
model. In this model, detailed interactions among vector life 
cycles, plasmodium sporogonic cycles, and human infection 
cycles were simulated under the influences of intrinsic and 
extrinsic effects. The simulations compare well with the shaded 
field measurements for vivax and falciparum malarias (Zollner, 
unpublished data).