อิทธิพล เอกะหิตานนท์. Improving the Estimated Area of Rice Cultivation by Classification Techniques. Master's Degree(Computer Engineering). King Mongkut's University of Technology Thonburi. KMUTT Library.. : King Mongkut's University of Technology Thonburi, 2552-10-31.
Improving the Estimated Area of Rice Cultivation by Classification Techniques
Abstract:
The area of rice cultivation is a critical data item for predicting rice yield. If the
calculated area is close to the real area, forecasts will be more accurate. This thesis
focuses on improving the area estimation using several methodologies. The Office of
Agricultural Economics (OAE) currently uses visual interpretation of a single
LANDSAT TM epoch to classify rice area. However the process is time consuming and
requires highly trained people. It is also error prone because it considers only a single.
point in time; rice fields that have not yet been planted will be omitted.
The objectives of this study are to improved initial estimates using novel classification
techniques based on low resolution remote sensing images. Rice is difficult to classify
correctly because the rice fields have different spectral signatures during different
periods of growth (transplanting, growing, reproducing, mellowing, harvesting) but
multiple periods may be present in a single scene. We examine multitemporal data in
two methods, NDVI difference classification and fuzzy logic classification. NDVI
difference classification defines the rice class by using the difference of NDVI between
two dates of image data. Fuzzy logic classification is one technique which can deal with
multitemporal data and can have multiple rules per class for different signatures. To find
appropriate fuzzy logic rules, we used an optimization technique modeled on
evolutionary biology called "genetic algorithm" (GA).
NDVI difference classification produced good results when the selected dates matched
the period of rice changing state and the selected difference threshold range was
appropriate to the class. The images should have maximum difference between two
dates. The results of fuzzy logic demonstrated with appropriate parameters that GA
produced a stable set of fuzzy rules based on the training data. Using the fuzzy rules
from GA process to classify rice areas resulted in class assignment that was close to the
visual interpretation and as accurate as maximum likelihood classification for some
image epochs.