Background: Vegetation change occurs as expansion/contraction of existing communities, gradual change of community structure, and/or loss or gain in standing biomass. Sea level rise is expected to induce change in extent and composition of plant communities in coastal wetlands. With rising salinity, mangroves and other salinity-tolerant plants are predicted to expand into freshwater marsh communities, creating a mosaic of salt-tolerant and freshwater woody and herbaceous communities. Remote sensing facilitates detecting these patchy, fine-scale changes with high spatial precision when using fine spatial resolution satellite data, but these data do not have long temporal coverage and are acquired aperiodically. Data available for long temporal extents exist only at lower spatial resolutions.
Question: How do vegetation patterns detected at a fine resolution scale to coarser resolutions, and is it possible to accurately predict change at that resolution from archived remotely sensed data?
Methods: We used multi-spectral WorldView-2 satellite data in combination with airborne LIDAR-derived vegetation heights to map vegetation along a salinity gradient in the southeastern coastal Everglades with high spatial precision (2x2m resolution) using random forest classifiers. In order to detect change in this area over the past 30 years, we developed a framework that allows scaling of vegetation detected at the high spatial resolution to match the coarser resolution of the 30+ year historical Landsat data archive. The re-sampling framework applies k-means clustering algorithms to random samples of the fine-scale map to construct a community classification system that reflects common mixes at the coarser grain (30x30m resolution). Change for scaled vegetation classes are then assessed by comparing change in magnitude and direction of Kauth-Thomas transformed reflectance values.
Results/Conclusions
Results: When applying this method of vegetation scaling and change detection in a freshwater marsh where we knew change had occurred, we were able to detect communities at the fine scale with an accuracy of 89.2%, re-scaled vegetation classes aggregated to 30x30m pixels with 86.3% accuracy, and vegetation change between 1999 and 2009 with 81.8% accuracy (95% confidence). Some changes indicated a shift from one community to another, while large areas changed density within the same vegetation class, and 76.9% of the area did not change.
Conclusion: These results indicate that we can scale vegetation patterns at fine resolutions to detect community change at coarser scales. This is significant because it allows for more precise and accurate prediction of not only changes in vegetation pattern but also of changes in biomass.