Use of new generation geospatial data and technology for low cost drought monitoring and SDG reporting solution : a thesis presented in partial fulfillment of the requirement for the degree of Master of Science in Computer Science at Massey University, Manawatū, New Zealand
Loading...
Date
2018
DOI
Open Access Location
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Massey University
Rights
The Author
Abstract
Food security is dependent on ecosystems including forests, lakes and wetlands,
which in turn depend on water availability and quality. The importance of water
availability and monitoring drought has been highlighted in the Sustainable Development
Goals (SDGs) within the 2030 agenda under indicator 15.3. In this context
the UN member countries, which agreed to the SDGs, have an obligation to report
their information to the UN. The objective of this research is to develop a methodology
to monitor drought and help countries to report their ndings to UN in a
cost-e ective manner.
The Standard Precipitation Index (SPI) is a drought indicator which requires longterm
precipitation data collected from weather stations as per World Meteorological
Organization recommendation. However, weather stations cannot monitor large areas
and many developing countries currently struggling with drought do not have
access to a large number of weather-stations due to lack of funds and expertise.
Therefore, alternative methodologies should be adopted to monitor SPI.
In this research SPI values were calculated from available weather stations in Iran
and New Zealand. By using Google Earth Engine (GEE), Sentinel-1 and Sentinel-
2 imagery and other complementary data to estimate SPI values. Two genetic
algorithms were created, one which constructed additional features using indices
calculated from Sentinel-2 imagery and the other data which was used for feature
selection of the Sentinel-2 indices including the constructed features. Followed by
the feature selection process two datasets were created which contained the Sentinel-
1 and Sentinel-2 data and other complementary information such as seasonal data
and Shuttle Radar Topography Mission (SRTM) derived information.
The Automated Machine Learning tool known as TPOT was used to create optimized
machine learning pipelines using genetic programming. The resulting models yielded an average of 90 percent accuracy in 10-fold cross validation for the Sentinel-
1 dataset and an average of approximately 70 percent for the Sentinel-2 dataset. The
nal model achieved a test accuracy of 80 percent in classifying short-term SPI (SPI-
1 and SPI-3) and an accuracy of 65 percent of SPI-6 by using the Sentinel-1 test
dataset. However, the results generated by using Sentinel-2 dataset was lower than
Sentinel-1 (45 percent for SPI-1 and 65 percent for SPI-6) with the exception of
SPI-3 which had an accuracy of 85 percent.
The research shows that it is possible to monitor short-term SPI adequately using
cost free satellite imagery in particular Sentinel-1 imagery and machine learning. In
addition, this methodology reduces the workload on statistical o ces of countries
in reporting information to the SDG framework for SDG indicator 15.3. It emerged
that Sentinel-1 imagery alone cannot be used to monitor SPI and therefore complementary
data are required for the monitoring process.
In addition the use of Sentinel-2 imagery did not result in accurate results for SPI-1
and SPI-6 but adequate results for SPI-3. Further research is required to investigate
how the use of Sentinel-2 imagery with Sentinel-1 imagery impact the accuracy of
the models.
Description
Keywords
Droughts, Mathematical models, Droughts, Remote sensing, Geospatial data, Sustainable Development Goals