A Two-stage Machine Learning Framework for Detection and Characterization of Tree Mortality Using High Spatial Resolution Satellite Imagery
Abstract
Physiological response to drought is a great uncertainty in the biospheric feedback to climate variability and change. Detection and classification of tree species, stress, and mortality using high resolution satellite images would significantly improve regional and global vegetation modeling and understanding and predicting the global vegetation response to climate change. In this study, using high resolution satellite imagery of Quickbird and Worldview, we have developed a new two-stage machine learning framework to separate vegetation from non-vegetation in the first stage, and then assess vegetation condition, e.g. healthy, stressed, or dead trees in the second stage. The supervised learning was applied in the first stage classification model. The performance of classification accuracy was compared among several learning algorithms, e.g. minimal distance, maximum likelihood, support vector machines. The minimal distance algorithm yielded the best performance with about 95% overall classification accuracy among the compared algorithms. We then applied the unsupervised learning algorithms to further assess health of the vegetation that was identified in the first stage. Clustering algorithms were applied to classify the trees in an automated fashion. The IsoData algorithm accurately identified live and dead trees. In summary, the two-stage model has demonstrated promising classification accuracy and computing efficiency to detect and characterize tree mortality.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2011
- Bibcode:
- 2011AGUFM.H11L..07C
- Keywords:
-
- 0540 COMPUTATIONAL GEOPHYSICS / Image processing;
- 1640 GLOBAL CHANGE / Remote sensing;
- 1807 HYDROLOGY / Climate impacts;
- 1942 INFORMATICS / Machine learning