LULC - Special topic

Land Use and Land Cover III

6607.1 - Land Use/Cover Mapping in an Urban Area from Satellite Imagery Using Random Forests: A Comparative Evaluation

Thursday, July 6
2:50 PM - 3:10 PM
Location: Maryland B

Random forests as a novel ensemble learning algorithm have significant potential for land cover mapping in complex environments but have not yet been sufficiently tested by the remote sensing and mapping science community relative to some more popular pattern classifiers. In this research, we implement random forests as a classifier for land cover mapping from a satellite image covering a complex urban area, and evaluate the performance relative to several popular classifiers including Gaussian maximum likelihood, multi-layer-perceptron networks, and support vector machines. Each classifier is carefully configured with the parameter settings recommended by recent literature, and identical training data are used in each classification. The accuracy of each classified map is further evaluated using identical reference data. Then, a statistical rigorous procedure using the z-test is conducted to determine if random forests and each of the other classifiers are significantly different in terms of classification accuracy. Random forests are found to outperform the Gaussian maximum likelihood classifier and multi-layer-perceptron networks. Random forests and support vector machines produce identical overall map accuracy but the former has the smallest standard deviation of categorical accuracies, suggesting their better overall capability in classifying both homogenous and heterogeneous land cover classes. Although random forests perform moderately better, the classifier and each of the other three are shown to be not significantly different in terms of their map accuracies, as revealed by the z-test results. Nevertheless, random forests have shown their robustness due to their most accurate classification, relatively balanced performance across all land cover categories, and relatively easier to implement. These findings should help promote the use of random forests for land cover mapping in complex environments.

Xiaojun Yang

Professor
Florida State University

Dr. Xiaojun Yang is a tenured Full Professor of Geography at Florida State University. He completed his higher education in China, Holland, and USA. Dr. Yang’s research interests include the development of geospatial information science and technology with applications to support geographic inquiries in urban and environmental domains. His research was funded by NASA, NSF, and EPA. Dr. Yang authored more than 100 English publications. He served as a CAGIS Director (2012-2016) and Chair of the ICA Commission on Mapping from Remote Sensor Imagery (2007-2015). Dr. Yang has been a member of the 28th ICC Local Organizing Committee.

Presentation(s):

Send Email for Xiaojun Yang

Di Shi

Director of Cartographic Services Lab and Lecturer
University of Kansas

Dr. Di Shi is a Lecturer and Director of Cartographic Services Lab at the University of Kansas, Lawrence. She obtained his PhD from Florida State University in Summer 2016. Dr. Shi's research interests include the development of geospatial information theories and technologies with urban and environmental applications. She has published several peer-reviewed articles in premier remote sensing and mapping science journals.

Presentation(s):

Send Email for Di Shi

William a. Mackaness

Senior Lecturer in GIS
The University of Edinburgh

William Mackaness is a senior lecturer at The University of Edinburgh in the School of GeoSciences. His research is in map generalisation - developing automated techniques for display of geographic information at multiple levels of detail. His interest extends into mapping via smartphone technologies. His work in Malawi and with aboriginal communities in Australia has led to an interest in the socio technical dimensions of GIS.

Presentation(s):

Send Email for William Mackaness


Assets

6607.1 - Land Use/Cover Mapping in an Urban Area from Satellite Imagery Using Random Forests: A Comparative Evaluation



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Land Use/Cover Mapping in an Urban Area from Satellite Imagery Using Random Forests: A Comparative Evaluation