IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
INVERSION OF PHYSICAL RADIATIVE TRANSFER MODELS USING 
MULTISPECTRAL REMOTE SENSING DATA AND GROUND CONTROL 
INFORMATION FOR PRECISION FARMING 
Franz Kurz 
Technical University Munich, Chair for Photogrammetry and Remote Sensing 
ArcisstraBe 21, 80333 München, Germany, email: franz.kurz@bv.tum.de 
Commission VII, WG VII/2 
Keywords: Agriculture, Multispectral, Estimation, Accuracy, Radiation Model 
ABSTRACT: 
We propose a general framework to estimate vegetation parameters from multispectral remote sensing data by inversion of combined 
physical radiative transfer models and by using a moderate amount of ground control information. This framework has been 
exemplarily demonstrated for different winter wheat fields imaged by a Daedalus ATM multispectral scanner in the last two years. 
The focus lies on the variations of vegetation parameters within single fields, which are used to derive information about soil 
heterogeneities for precision farming. For the estimation of vegetation parameters, we use physical radiative transfer models, e.g. 
SAIL and PROSPECT, combined with a linear empirical model. Results show the invertibility of the models for leaf area index, 
chlorophyll content, specific dry matter, and specific water content. À strategy for the use of ground control data is proposed to 
receive high accuracies of the estimated vegetation parameters with a minimum of necessary ground measurements. 
1. INTRODUCTION 
‘ 
Remote sensing techniques play an important role in precision 
farming by providing continuous and contactlessly acquired 
data of agricultural crops. Remote sensors image vegetation, 
which is growing on different soil types with different water 
availability, substrate, impact of cultivation, and relief. These 
differences influence the state of the plants and cause 
heterogeneous regions within single fields. Hence, the 
heterogeneous vegetation acts as an interface between soil and 
remote sensing information, because vegetation parameters 
describing the state of the plants can be deduced from remote 
sensing imagery. 
In this context, a framework for the estimation of vegetation 
parameters from multispectral imagery is proposed. The focus 
of our approach lies on the variation of the vegetation 
parameters within single fields assuming that field borders and 
vegetation type are given. This framework applies both a 
physical and an empirical model to derive the functional 
relationship between vegetation parameters and measured image 
grey values. The physical model is used to estimate selected 
vegetation parameters by an inversion process, whereas the 
empirical model fits the physical model to local characteristics 
and sensor specifics. 
This technique has been exemplarily tested for several sites with 
winter wheat imaged by a Daedalus ATM multispectral scanner 
from DLR (German Aerospace Center) Results show the 
attained accuracies for the estimated vegetation parameters with 
respect to the amount of ground control points. 
2. RELATED WORK 
The estimation of vegetation parameters using physical models 
is based on the description of radiative transfer in the canopy by 
means of an analytical reflectance model. In the last 30 years, 
various models describing radiative transfer in canopy, soil and 
leaves have been published. These models provide the 
relationship between the radiation incoming from the sun and — 
according to the bidirectional reflectance distribution function 
(BRDF) - to the observer scattered radiation. Inputs of these 
models are the structural and spectral parameters of the 
vegetation/soil medium. Models describing the complete 
vegetation/soil medium are called canopy transfer models, e.g. 
the SAIL model (VERHOEF 1984), the Nilson-Kuusk canopy 
reflectance model (NILsON and Kuusk 1989), and the LCM2 
model (GANAPOL et al. 1999). In these models, the leaves are 
considered as the only components of the vegetation canopy 
characterised by their reflectance and transmittance. The 
spectral properties of the leaves are mainly influenced by the 
chemical consistency of the leaves, which can be modelled by 
so called leaf optical physical models, e.g. PROSPECT 
(JACQUEMOUD and BARET 1990), LEAFMOD (GANAPOL et al. 
1998), and SLOPE (MAIER 2000). 
Generally, these models were set up in the forward mode. This 
means output parameters are the reflectance on top of the 
canopy for given parameters of the vegetation/soil medium. The 
solution of the resulting inverse problem was subject of many 
investigations during the last years. Depending on the applied 
sensors two main methods can be distinguished, inversion with 
multidirectional and with multispectral data. Independent of the 
method the inversion of the physical model is conducted using 
different mathematical algorithms such as look up table 
(KNYAZIKHIN et al. 1998), iterative optimisation (JACQUEMOUD 
et al. 1995), and neural networks (BUELGASIM et al. 1998). 
These algorithms adjust the model input parameters in such a 
way that the model-predicted values closely match the measured 
values. A comparison of these methods (PRAGNERE et al. 1999) 
gives slight advantages to the neural networks technique that is 
most robust for different sensors and canopy types. Up to now, 
the inversion studies are performed with simulated reflectance 
or field spectrometer measurements. In practical applications 
with airborne or space-borne sensor data, a variety of empirical 
tools, such as vegetation indices and spectral mixture models, 
are widely used to derive biophysical parameters of the 
vegetation. Our approach combines strict inversion of physical 
models with empirical elements to estimate biophysical 
parameters from airborne sensor data. In previous work (Kurz 
and HELLWICH 2000), we describe our inversion method, the 
investigation of invertibility and the selection of relevant 
biophysical parameters in more detail. We chose four 
parameters for a partial inversion of the applied physical 
models: leaf area index, chlorophyll content, specific dry 
matter, and specific water content. The inversion is conducted 
using simulated annealing followed by a least squares 
adjustment. 
    
mo =k =n = Ww 
- 
fr^ (S CR 40 co 0 0 57m 7 > AYO 
es 
— 
FAR Ap ten ta 
m