Intelligent System Based Automatic Prediction of Drought Using Satellite Images
No Thumbnail Available
Date
2011-08
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Early days of technology employed traditional methods like rainfall distribution, duration of
sunshine, and the related meteorological data to forecast a forthcoming event regarding
drought. Recently, satellite remote sensing has been considered as an appropriate tool for
deriving information in spatial and temporal domains by providing multi-spectral reflectance
data at regular intervals. Satellites from centers like National Oceanic and Atmospheric
Administration (NOAA) and Meteosat capture the spectral reflectance from green plant and
provide this frequently to monitor green vegetation conditions on the ground. Using the red and
infrared band reflectances a vegetation index called Normalized Difference Vegetation Index
(NDVI) was derived; which is vital to access the evolution of drought as well as predict crop
yield.
The aims of this study are to analyze series of deviation of NDVI images, extract virtual
drought objects from the series, and investigate for drought patterns from historical image for
growing season. Subsequent to this, appropriate prediction model of these patterns was
developed for early measures while within the same season. And it was applied on the subset of
image data with reported drought occurrences in Ethiopia.
In this study, the virtual drought objects extracted from images over the growing season (June -
September) were found to exhibit a given pattern for the historical drought years. After
producing the descriptors of drought objects in the series using principal component analysis,
combined and separate artificial neural network (ANN) models were used to predict these
patterns. In the combined approach all the descriptors of the object in the next time step were
predicted all at the same time while in the separate approach the prediction was made one by
one. Accordingly, the models designed to forecast the future state of drought object using these
two approaches yielded promising results. Especially, the three and four time lag combined
ANN prediction model produced an overall RMSE of 20.80 and 16.43, respectively, which was
a better result compared to a 36.92 RMSE of separate ANN approach. It is understood that this
work will give new views for ways in drought prediction for early warning and crop condition
monitoring at near real-time.
Keywords: Drought prediction, NDVI images, Virtual drought object, Intelligent system
Description
Keywords
Drought prediction, Ndvi Images, Virtual Drought Object, Intelligent System