BACKPROPAGATION NEURAL NETWORK FOR SOIL MOISTURE RETRIEVAL 
USING NAFE’05 DATA : A COMPARISON OF DIFFERENT TRAINING ALGORITHMS 
Soo-See Chai 1 , Bert Veenendaal 1 , Geoff West 2 , Jeffrey Philip Walker 3 
'Department of Spatial Sciences, Curtin University of Technology - (soosee.chai, b.veenendaal)@curtin.edu.au 
2 Department of Computing, Curtin University of Technology - g.west@curtin.edu.au 
department of Civil and Environmental Engineering, The University of Melbourne -j.walker@unimelb.edu.au 
WGs WG IV/9 - Mapping from High Resolution Data 
KEY WORDS: Neural Network, Backpropagation, Soil Moisture, Airborne, Algorithms, Microwave 
ABSTRACT: 
The backpropagation artificial neural network (ANN) is a well-known and widely applied mathematical model for remote sensing 
applications for pattern recognition, approximation and mapping of non-linear functions and time-series prediction. The 
backpropagation ANN algorithm is underpinned by a gradient descent algorithm that is used to modify the network weights to 
maximise performance, using some criterion function. There are a number of variations from this general algorithm and it is 
necessary to explore these to find the best method for any particular application. The application considered in this paper is the 
determination of volumetric soil moisture content given airborne microwave measurements of the H- and V-polarized brightness 
temperature obtained during the National Airborne Field Experiment 2005 (NAFE’05). In this paper, a number of backpropagation 
ANN methods are investigated. Some produce the globally acceptable accuracy of less than or equal to 4%v/v of Root Mean Square 
Error (RMSE). However, the standard deviation among the 11 different variations of backpropagation training algorithms (0.55) is 
significant compared to the accuracy. Hence, there is a need for a full analysis of the backpropagation ANN and careful selection of 
the best backpropagation ANN model to be used. 
1.INTRODUCTION 
Soil moisture in the top few centimetres is an important variable 
which governs the partitioning of rainfall into run-off or infiltra 
tion. Obtaining this value accurately is important in the predic 
tion of erosion, flood or drought. The most accurate way of ob 
taining this value is through actual in-situ field measurements 
but this technique is very time and cost consuming especially 
for large scale soil moisture retrieval. Microwave remote sens 
ing, either active or passive, has been utilized and has shown 
the greatest success in estimating soil moisture in a temporally 
and spatially consistent manner [1]. 
Passive microwave radiation in the L-band for soil moisture re 
trieval is actively being explored because the effect of the at 
mosphere is small and can be modelled and predicted. More 
over the vegetation attenuation at L-band frequencies is 
minimal. Passive microwave radiometer measures the thermal 
emission from the surface. The intensity observed is propor 
tional to the product of soil temperature and the surface emis- 
sivity. This product is called the brightness temperature (Tb) 
[2]. The soil emissivity at microwave wavelengths is a strong 
function of its moisture content because of the large dielectric 
contrast between dry soil and water. The relationship between 
the soil moisture and the received radiation is non-linear, ill- 
posed and complex [3]. 
Artificial Neural Network (ANN) is a model-free estimator as it 
does not rely on an assumed form of the underlying data [4]. 
Using some of the observed data, the ANN will try to obtain an 
approximation to the underlying system that generates the ob 
served data through a process called learning by example. This 
is a method which is based on very sound mathematical princi 
ples and it has proved very successful in developing computa 
tionally efficient algorithms for geophysical applications like 
satellite remote sensing, meteorology, oceanography, numerical 
weather prediction, and climate studies. A type of ANN, the 
multilayer perceptron trained by a backpropagation algorithm 
has been utilized successfully for soil moisture retrieval. Ex 
amples include using an Error Propagation Learning Back 
Propagation (EPLBP) neural network to retrieve soil moisture 
from simulated brightness temperature [5], the use of multi 
frequency microwave radiometer for retrieving soil moisture 
profile using backpropagation neural network to retrieve soil 
moisture [6], and using a backpropagation neural network and 
Levenberg-Marquardt training algorithm to classify and retrieve 
soil moisture and soil temperature profile using remotely sensed 
data [7]. 
Standard backpropagation refers to the gradient descent algo 
rithm based on the Widrow-Hoff learning rule, in which the 
network weights are moved along the negative of the gradient 
of the performance function[8]. The term backpropagation re 
fers to the manner in which the gradient is computed for nonlin 
ear multilayer networks. There are a number of variations on 
the basic algorithm that are based on other standard optimiza 
tion techniques. A review of the literature reveals that there is a 
need for comparison of these algorithms to determine the best 
backpropagation training algorithm in terms of the accuracy of 
the final result. This paper presents a comparison of the accu 
racy of soil moisture retrieval using the different variations on 
the standard backpropagation algorithm provided by MATLAB 
Neural Network Toolbox [8].