In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
SURFACE TEMPERATURE ESTIMATION USING ARTIFICIAL NEURAL NETWORK 
M. R. Veronez 3 G. Wittmann a , A. O. Reinhardt b , R. M. Da Silva 3 
a Graduate Program in Geology, Remote Sensing and Digital Cartography Laboratory, Vale do Rio dos Sinos 
University (UNISINOS), Av. Unisinos, 950, CEP 93022-000, Säo Leopoldo, RS, Brazil veronez@unisinos.br 
KEY WORDS: Artificial Neural Networks; image from NOAA satellite; surface temperature, split window 
ABSTRACT: 
This research presents an alternative method to extrapolation land Surface Temperature (ST) through artificial neural network, using 
positional variables (UTM coordinates and altitude), temperature and air relative humidity. The study region was the Rio dos Sinos 
Hydrographic Basin (BHRS), in Rio Grande do Sul state, Brazil. For training the neural network was used a thermal image from 
NOAA satellite, with pixel size of 1X1 km, with known ST information referring to 12/06/2003. After training many network sets 
were done and one of them with the best performance and composed by a single intermediate layer (with 4 neurons and logistic 
sigmoid activation function) was selected. The training network was tested inside the BHRS where were collected 60 points of ST 
values supported by a portable laser sensor on date 3/18/2008. The average error provided by this model for ST measurement was 
2.2°C and through executed statistical tests was possible to verify that not exist variation between average ST values accepted as true 
and the values provided by the neural model with a significance level of 5%. 
1. INTRODUCTION 
Artificial Neural Networks (ANN) are organized and 
interconnected collections of processing units (neurons or 
nodes), whose operation is analogue to a neural structure of 
intelligence organisms (Müller and Fill, 2003). ANN extract its 
computational power from its solid parallel distribution 
structure and ability to leam/generalize, allowing the resolution 
of complex propositions in many knowledge areas (Haykin, 
2001). 
ANN execution is inspired on human brain (Haykin, 2001) and 
has been used in many applications with success. In agreement 
with Galväo et al. (1999), by the reason of its nonlinear 
structure the ANN can acquire more complex data 
characteristics, which are not always possible using traditional 
statistical techniques. 
According to Müller and Fill (2003), conventional methods 
don’t have the necessity to know the problem intrinsic theory 
and also don’t need to analyze relations that aren’t totally 
known between involved variables in modeling, so ANNs have 
a greater advantage over them. 
Surface Temperature (ST) is established by a fenologic 
parameter that is significantly influenced by climate variations 
and is an indicator of plants hydrous state. Therefore its 
estimation has a large utility to surveys that need to assure the 
observation of hydrous culture demand, contributing in a 
significative way the irrigation programs (Silva, 2007). 
Currently a mostly used method to ST estimation is through the 
use of thermal satellite images. Rivas (2004) recommends the 
use of NOAAAVHRR (National Oceanic Atmospheric 
Administration/Advanced Very High Resolution Radiometer) 
images adapted to split windows equation to estimate ST 
values. Such modeling connects emissivity variables and 
atmospherical data. Is a complex method because beyond it 
processes not simple statistical models, it has the necessity to 
work with digital images in the emissivity determination 
process. 
So is much important to have methods of ST estimation in a 
more convenient way as, for example, the combination of 
temperature, air relative humidity and geographical data 
position (Veronez et al., 2006). 
Some authors realized researches using ANNs with the finality 
of ST estimation based on climate elements (Yang et al., 1997; 
Atluri et al., 1999; George, 2001; Veronez et al., 2006; Mao & 
Shi, 2008), where all of them have found satisfactory results. 
Yang (1997) describe the importance to develope a model 
capable to assist agricultural processes, once that ST estimation 
in distinct depths is complex due its large number of involved 
variables. The author have used as ANN input the following 
variables: daily precipitation, evapotranspiration, air maximun 
and minimun temperature and days of year, been all these 
information easy to obtain on meteorological station. 
Atluri (1999) has modeled through ANNs the humidity and soil 
ST with the Levemberg-Marquardt algorithm and after many 
tests he obtained an accuracy estimation of 98.7% with these 
variables. The same author describes the importance to establish 
an efficient system to extrapolate such information, once that 
these variables are required by distinct geosciences areas. 
George (2003) describes the importance of usage ANNs to 
estimate soil surface temperature using easy access data. 
Thinking on it Veronez et al. (2006) proposed the ANN usage to 
model such variable using only positional information (East and 
North UTM coordinates), altitude and air average temperature. 
The network was established through a supervised training, 
where ST information was extracted from NOAA satellite 
images. 
The results found by Veronez et al. (2006) show that is possible 
to extrapolate ST values on distinct periods from NOAA image 
processing date using as ANN input a variable that changes with 
the time. For the specific case of this research the used variable 
was air average temperature. The ST processing was based on 
NOAA satellite image with a surface coverage from 6/12/2003. 
ST values were extrapolated on 10/8/2005 and the model 
validation was accomplished in a municipality urban area 
located on Rio dos Sinos Hydrographic Basin (BHRS). The 
authors collected ST information with a laser thermometer and 
compared them with values provided by ANN, having an 
average error less than 2.3°C. 
Although some researches aim to simplify the input data on ST 
estimation process, is understood that exist another options to 
be learned using climate data associated with thermal images. 
So the purpose of this research was propose an ANN aiming to