IAPRS & SIS. Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad. India,2002
ARTIFICIAL NEURAL NETWORKS & MAXIMUM LIKELIHOOD FOR
CLASSIFICATION OF PADDY WITH MULTI-TEMPORAL REFLECTANCE IMAGES
C. S. Murthy*, P. V. Raju, K. Abdul Hakeem and K.V.S. Badrinath
National Remote Sensing Agency, Hyderabad 500 037, India — (murthy. cs, raju pv, abdulhakeem_k, badrinath_kvs) @nrsa.gov.in
COMMISSION VII, WG VII / 1.2
KEY WORDS: Artificial Neural Networks, Multi-temporal, Classification, Reflectance, Accuracy
ABSTRACT:
The objective of the present study is to evaluate the performance of Artificial Neural Networks (ANN) and Maximum Likelihood (ML)
for the classification of paddy with multi temporal reflectance images. The study was carried out in Sriramsagar command area, one ofthe
largest surface irrigation systems in Andhra Pradesh state, India. The spatial information on the extent of paddy crop is very important for
the command area authorities, to assess the irrigation utilization pattern. The classification of paddy in the study area was constrained by
two factors; (1) non-availability of overpass data on optimum date and (2) staggered crop calendar. Two approaches were adopted for
classification, namely, (1) sequential classification and (2) composite data classification. Sequential classification and masking was
performed with February and March data, separately with ANN and MLC. The results indicate that the total estimated paddy area using
ANN is 53% of reference estimate, where as that of MLC is only 38%. However, the accuracy of the estimate is 89% for ANN and 92%
for MLC. Thus, the performance of ANN is relatively better than MLC, though in absolute sense, both the classifiers have failed to
achieve complete classification through sequential approach. Combining the reflectance images of the two overpass dates i.e with multi-
temporal data did the composite data classification. The estimated paddy area with MLC is about of 57% of reference estimate, where as
that with ANN is 74%. The accuracy of the estimate is 89.14% for MLC and 89.84% with ANN. Thus, ANN classification has delineated
more paddy area, compared to MLC without compromising with the accuracy of classification. The study concludes that Artificial Neural
Networks has the potential for improving crop classifications using multi-temporal remote sensing images.
“ 1. INTRODUCTION
The differences in spectral reflectance pattern as a function of
physiological properties of different vegetation types form the
basic logic to identify the agricultural crops with remote sensing
images. One of the major findings from various studies on remote
sensing based crop land inventory is that maximum discrimination
between different crop types occurs at different stages in the
growth cycle and therefore, it is not possible to capture all these
differences using a single date image. Consequently, the
application of multi date or multi-temporal image data has been
recommended to improve the discrimination ability (Hlavaka et al.
(1980), Conese and Maseli (1991), Pax-Lenny and Wood cock
(1997, Viera and Mather (2000). Although, Maximum Likelihood
classification is the most common classification algorithm, studies
have shown that other classification techniques such as Principal
Components Analysis by Richards (1984), Change vector analysis
by Lambin and Strahler (1994), greenness profile analysis by
Badhwar (1984), spectra-temporal response surfaces analysis by
Viera et al (1998, 2000), unsupervised classification and temporal
profiling by Lo (1986) have greatly improved the classification.
Non parametric classifiers like Artificial Neural Networks (ANN)
have shown promise in the areas of pattern recognition and
classification (Lipman 1987, King 1989, Benediktsson et al, 1990
and Liu et al., 2000). ANN are parallel calculation architectures
whose structures are based on the human brain. If suitably trained
using a set of examples, they can learn and extract the link
between the input data and the corresponding out data (Lippman,
1987).
Sriramsagar project command area, one of the largest surface
irrigation systems in Andhra Pradesh. During rabi season (post
* Corresponding author
monsoon period i.e., December - May), paddy is the predominant
crop followed by cotton and maize in the command area. The
spatial information on the extent of paddy crop is very important
for the command area authorities, to assess the irrigation
utilization and to evaluate the irrigation system performance.
Therefore, remote sensing data analysis was undertaken by
National Remote Sensing Agency, in the command area during
rabi 1998-99.
The satellite data of January is very crucial for separating paddy
from other crops. This is because, paddy in its early stages, has
good seperability with the other crops which are in good
vegetative phase. During rabi 1998-99, the satellite data of
January is available only for one path, due to cloud cover problem.
Therefore, crucial data for classification is missing for some
portion of the study area falling in the 2" path of the satellite
coverage. Further, the paddy cultivation is characterized by
significant staggering in crop calendar, resulting in large
variability in the crop condition during February and March.
Under this background, the current study is taken up to improve
the classification of paddy under these constraints, through
different classification techniques.
2. OBJECTIVE OF THE STUDY
To evaluate the performance of Maximum Likelihood and
Artificial Neural Networks algorithms for classifying paddy from
non-paddy crops. The constraints for paddy classification in this
study, as discussed in earlier sections are two fold; (1) non-
availability of optimal overpass data and (2) staggered crop
calendar.