A comparison of classification techniques for monitoring and mapping land cover and land use changes in the subtropical region of Thai Nguyen, Vietnam : a thesis presented in partial fulfilment of the requirements for the degree of Master of Environmental Management at Massey University, Palmerston North, New Zealand
Deriving land cover/land-use information from earth observation satellite data is one of the
most common applications for environmental monitoring, evaluation and management. Many
parametric and non-parametric classification algorithms have been developed and applied to
such applications. This study looks at the classification accuracies of three algorithms for
different spatial and spectral resolution data. The performance of Random Forest (RF) was
compared to Maximum Likelihood (MLC) and Artificial Neural Network (ANN) algorithms
for the separation of subtropical land cover/land-use categories using Sentinel-2 and Landsat 8
data. The overall, producers’ and users’ accuracies were derived from the confusion matrix,
while local land use statistics were also collected to evaluate the accuracy of classified images.
The accuracy assessment showed the RF algorithm regularly outperformed the MLC and ANN
in both types of imagery data (>90%). This approach also exhibited potential in dealing with
the challenge of separating similar man-made features such as urban/built-up and mining
extraction classes. The ANN algorithm had the lowest accuracy among the three classification
algorithms, while Landsat 8 imagery was most suitable for the classification of subtropical
mixed and complex landscapes.
As the RF algorithm demonstrated a robustness and potential for mapping subtropical land
cover/land-use, this study chose it to monitor and map temporal land cover/land-use changes
in Thai Nguyen, Vietnam between 2000 and 2016. The results of this temporal monitoring
revealed that there were substantial changes in land cover/land use over the course of 16 years.
Agricultural and forest land decreased, while urban and mining extraction land expanded
significantly, and water increased slightly. Changes in land cover/land-use are strongly
associated with geographic locations. The conversion of agriculture and forest into urban/builtup
and mining extraction land was detected largely in the Thai Nguyen central city and southern
regions. In addition, further GIS analysis revealed that approximately 69.6% (100.2km2) of new built-up areas had occurred within 2km of primary roads, and nearly 96% (137.6km2) of new built-up expansion was detected within a 5-km buffer of the main roads. This study also demonstrates the potential of multi-temporal Landsat data and the combination of remote sensing, GIS and R programming to provide a timely, accurate and economical means to map and analyse temporal changes for long-term local land use development planning.
Keywords: Random forest; Land cover mapping; Remote Sensing; Vietnam