Application of Remote Sensing Technologies for the Delineation and Assessment of Coastal Marshes and their Constituent Species

About | Data | Maps | Study Sites | Image Comparison | Recommendations | Links and Publications

Delineation and monitoring of coastal marshes | Identification of vegetative species within marshes

We have developed semi-automated techniques, based on pixel and object-oriented classification, for the classification of Landsat imagery into water, upland, and marsh categories.The techniques have a high classification accuracy and are repeatable on images collected at different times. Using these techniques, we have produced a map that inventories the location of marshes around the whole of Long Island Sound in September of 2002. These techniques could be applied to other Landsat images to track the distribution and extent of marshes over time although the spatial resolution of the Landsat image and the size of change occurring within the coastal marshes could become problematic.

A benefit of Landsat imagery is that a single path captures almost the entire Long Island Sound region (excluding the extreme eastern part of the Sound) thereby eliminating any temporal or radiometric resolution issues that are typical of aerial surveys. With a revisit time of every 16 days, acquisition of useable imagery is dependent on the atmospheric conditions of the area. If the timeframe is limited to just the July through September growing season, there is only the possibility of six or seven scenes that could be collected.

Landsat imagery, which has 30 meter spatial resolution, is useful for identifying the general location of coastal marshes but is too coarse to capture accurately their full extent and detail. ASTER imagery, while having higher spatial resolution in the visible to near-infrared portion of the electromagnetic spectrum did not perform any better in terms of classifying coastal marshes. A significant drawback to the ASTER imagery is the inconsistent temporal resolution (lack of quality growing season imagery) and the footprint size of the ASTER imagery (several scenes are required to cover the entire Sound area). QuickBird satellite imagery has a higher spatial resolution (2.44-meters multispectral) that is capable of capturing the full extent of the coastal marshes within the Long Island Sound area. Another advantage is the ability to point the sensor off nadir providing for additional opportunities for collection and avoiding poor atmospheric conditions. The drawbacks to the sensor are the north-south orientation of the path of the sensor thus requiring numerous scenes to be collected to cover the full extent of the Sound. It is extremely difficult to obtain a temporally consistent set of image data because of this.

The classification methodologies developed here for Landsat were combined with ancillary data sets such as elevation, impervious surface and other land cover classification to analyze a subset of marshes in the context of their surrounding land cover. Elevation data provide an advantage to the classification process as long as the data are of sufficiently small spatial resolution to have several data pixels fall within the upland boundaries of the coastal marsh. Additionally, the elevation accuracy is very important. This experiment was able to identify the relative magnitude of environmental pressures (e.g., sea level rise, urban and suburban development) on each marsh. The combination of factors was used to derive an environmental risk assessment for select marshes. This could be used to 1) identify marshes that are in need of attention, 2) identify the possible causes for increased risk in each marsh, 3) monitor the changes in marsh health through time and in response to mitigation strategies. The analyses employed here are flexible, allowing the addition of other georeferenced data if such data were to become available (e.g., pollutants).

The classification approaches used in this project can be adapted by other researchers in other GIS-based applications to address other environmental and management objectives.

We have developed a semi-automated classification methodology using QuickBird data to identify and differentiate the dominant species P. australis, S. patens and Typha spp. in Ragged Rock Creek Marsh. Critical to the success of this task was 1) the use of image data of high enough resolution to be comparable to the stands of plants under study (meter scale), and 2) documentation of the phenology of plants throughout the growing season to predict the spectral response of each species through time.

Our field surveys find that for Connecticut:
1. P. australis is best distinguished by its high NIR response late in the growing season due to its high biomass especially with respect to the other species.
2. Typha spp. is best distinguished by NIR response in June and high red/green response in August which corresponds to peak biomass and senescence, respectively.
3. S. patens is best distinguished by pigment differences that result in a unique green/blue ratio throughout the growing season, peaking in July.

The phenological variability seen in the field spectra was repeatable at Ragged Rock Creek over two field seasons, appears to be representative of the behavior of these species at other marshes in CT based on our more limited field surveys of 2004, and appears to coincide with the behavior of P. australis at NY and NJ marshes based on comparison to the literature. To our knowledge, these are the only field spectra collected of marsh vegetation in CT, and as such should serve as a baseline for future remote sensing and ecological research. The spectra are inventoried and available on our web site for use by other researchers.

We have shown that the behavior the high spectral resolution field data is recognizable in the coarser spectral resolution QuickBird data and that high spatial resolution, moderate-resolution multispectral data such as QuickBird are adequate to measure and distinguish remotely marsh vegetation. This should also apply to other four-band data sets including NIR aerial photography, where, for example, P. australis could be recognized in the early fall as having the highest NIR response in a single image. The benefits of remotely sensed data over aerial photographs in this example are clear: 1) the remote sensing data are already radiometrically and geometrically corrected allowing mosaicking of scenes over larger areas than a single aerial photo, and 2) the remote sensing data allow for procedures such as ratioing that eliminate noise due to atmospheric variability.

The radiometry rules make clear the necessity of multi-temporal imagery for mapping multiple species on a complex tidal marsh due to the growth habit of the numerous marsh species. When the objective is to map and isolate a single species, a single date of imagery could be adequate. This is likely the case for the recognition of P. australis, where acquisition of a single date of imagery late in the growing season would be adequate for inventory and monitoring. For P. australis monitoring, we suggest that current aerial surveys be flown late in the growing season and/or that these surveys be replaced by or augmented with high resolution satellite or aerial remote sensing data such as QuickBird, ADS40 or John Deere data. The advantages to the ADS40 and John Deere data over QuickBird are the higher spatial resolution and the temporal flexibility afforded by flying aerial surveys.

Top of canopy information derived from LiDAR or other data source capable of capturing plant canopies, such as ADS40 Digital Surface Model (DSM) data, is an added advantage for identifying various plant species. The success of using such information is dependent on the stem and crown density of the plant community being measured (areas with low plant density are not as easily identified) and there being sufficient height differences among the plants being classified.

Back to Top