Landcover mapping using Ikonos and LiDAR

ABSTRACT:  Landcover mapping using Ikonos and LiDAR remotely sensed imagery for Ft. Lewis, WA.

Eli Rodemaker,  Awu Graduate, Research Scientist Pacific NorthWest National Lab, Sequim Jerry Tagestad,  Research Scientist,   Richland

Ft. Lewis (US Army) a military reserve near Tacoma, WA is in the heart of the increasingly urbanized Puget Sound Lowlands. Due to the Fort having been removed from development it represents a unique ecological setting, containing some of the best-preserved remnants of rare Puget Sound Region ecotypes including, prairies and Oak woodlands.  US Department of Defense reservations while used for intensive purposes such as tracked vehicle maneuvering and ballistics training are mandated by Federal law to maintain the ecosystem integrity of their installations.  Land use and invasive species are two key components Ft. Lewis staff consider when creating management plans. Landcover mapping derived from high spatial resolution remotely sensed images provides spatially contiguous data ideal for complementing the Integrated Training Area Management (ITAM) programs.  A land cover classification for Fort Lewis, Washington was developed to provide current vegetation distribution information and test the phenomenological sensitivity of newly available remote sensing data.  This proof-of-concept project used IKONOS satellite multispectral image data and airborne LiDAR (Light Detecting and Ranging) altimetry data to derive a landcover map of the 13^th Division Prairie area of the installation. Ikonos and LiDAR data were collected at a 4m spatial resolution for this 49km² area.  Using an expert decision rule methodology and gradient analysis techniques, the multispectral imagery, LiDAR derived vegetation heights and topography, and ancillary data of soils were combined into a model of land types.   Components of this investigation, included field sampling needs and precision of land type definitions possible with Ikonos and LiDAR.  Land cover detail includes urban, herbaceous and shrub dominated prairies, deciduous, coniferous and mixed forests, water, and wetland classes.   Special attention was paid to the noxious weed shrub Scot’s Broom and the historical extent of prairie and oak communities.

Methods:

1. Ecological literature, meetings with ITAM and Environmental Directorate
2. Grand tour: determine initial cover type classification, further development of ecological understanding
3. Field work:  collect training data
4. GIS data preparation
5. Statistical summaries of field data to model layers
6. Modeling: cover type classification finalization (ordination)
7. Presentation of project at ITAM
8. Model refinement and accuracy assessment: field verification
9. Deliverables

LandCover Mapping using IKONOS and LIDAR PowerPoint Presentation

Fort Lewis Land cover Classification

Introduction

Pacific Northwest National Laboratory was tasked to develop a high-resolution land cover map of the 13^th Division Prairie.  This proof-of-concept project utilized high-resolution data from two different sources.  On December 12, 1999 LIDAR sensor was flown over the 13^th Division prairie by Spencer Gross of Portland, OR.  The second source of data was from the IKONOS multi-spectral sensor, with images acquired July 27, 2000 and October 15, 2000.

Methods

High-resolution satellite data were combined with LIDAR-derived elevation data to create a detailed land cover map of the 13^th Division Prairie using the expert system classification method.  This method involves the creation of rules based on observed ecological relationships, and relies heavily on field observations.

Data

IKONOS Multispectral - July 2000 and October 2000, July NDVI, October NDVI and NDVI texture.

LIDAR - Vegetation Height, Ground DEM, Topographic Moisture Gradient

The July 2000 IKONOS multispectral sensor data was chosen to be the primary data layer to cluster and from which the vegetation classes would be assigned via expert rule definition.  This date was chosen because the phenology was better during this period than in the October 2000 imagery.

Spectral Clustering

The July 2000 IKONOS data were clustered using the ISODATA (Tou and Gonzalez, 1977; Sabins, 1987; Jain, 1989) unsupervised clustering routine. The ISODATA clustering method uses the minimum spectral distance formula to form clusters. It begins with arbitrary cluster means and each time the clustering repeats; the means of these clusters are shifted. The new cluster means are used for the next iteration. The ISODATA utility repeats the clustering of the image until either: a maximum number of iterations have been performed, or a maximum percentage of unchanged pixels have been reached between two iterations.  The result of this procedure is an image file with n number of classes.  These classes represent the spectrally similar pixels that may or may not be of the same cover type. This operation was performed several times on the data, varying the number of classes was on each successive run until we had clustered the data to 25, 35 and 50 spectral clusters.  The statistics of the clusters were then analyzed to determine where the variability within clusters leveled off.  Based on this analysis, we chose 35 spectral clusters to model.

Field Data Collection

During May of 2001, we collected ground truth data in the region of the 13th Division Prairie.  The training site collection consisted of ground observation, field notes on plant species observed, photographs and "polygon" collection.  Extensive notes were collected on plant species, vegetation height, component percentages and confidence that the observed cover type exemplified that class.  The polygon collection required that we determine an area on the ground that represented a particular cover type (oak hardwoods, fescue prairie etc.).  The polygon was then digitized directly on the screen of the computer using the multispectral image as a backdrop.  Using this method, training sites can be rapidly collected and accurately placed over relatively homogeneous cover types.  One hundred and four training site polygons were collected over the course of 3 days for the following cover types:

• Douglas Fir

• Oak

• Oak/Douglas Fir

• Ash

• Cedar

• Poplar

• Riparian Vegetation

• Fescue

• Meadow

• Scot’s Broom

• Berry

• Snowberry

• Disturbed Soils

Classification

Once the training site data was collected and checked, we revisited the optimal number of clusters.  We intersected the training site data (vegetation cover type) with the 20, 35 and 50 spectral cluster data layers.   The results of this intersection showed that 35 spectral clusters was a sufficient number to model the cover types of interest.  This was determined by analyzing the distribution of vegetation type by spectral cluster.  This analysis showed that 20 spectral clusters were insufficient to classify all the spectral variants within the cover types.  35 spectral clusters were confirmed to be a sufficient number with training sites (cover types) occurring in all but one spectral cluster.

The spectral data were then summarized by all the ancillary data layers and vise versa.  This process counts and summarizes all the occurrences of a particular spectral cluster within a vegetation type, or height class etc.  For example, the results of intersecting the “35 spectral clusters” file with the “Vegetation Height” file, showed the distribution of Vegetation height values for every spectral cluster.  This information is necessary to begin building modeling statements that allow us to assign meaning to the spectral cluster. Each spectral cluster was summarized by all the other data layers and all other data layers were summarized by spectral clusters, as well as one another.  Below is table showing the summarizations we performed:

Building Modeling Statements

Once all the data were summarized, we used the ERDAS Imagine Spatial Modeler program to apply our logical statements and assign land cover type to spectral clusters.  Meaning was assigned to clusters by analyzing all the data for a particular spectral cluster, recognizing the ecological variables and observed occurrences of the variable, and applying a logical statement to assign a land cover class.  For example, the figure below plots the distribution of vegetation height in spectral cluster 1.

Figure 1.  Distribution of Vegetation Heights within spectral cluster number 1

The summarization data were analyzed to determine modeling cut off values and apply ecologically significant rules to model the correct cover types.  The final model is included in Appendix A.

Post Classification

The preliminary land cover classes were intersected back with the training site polygons.  The results of this intersection allowed us to determine which classes to combine into coarser land cover categories (ash, poplar>deciduous trees).  For example, if the statistics showed that our training site data for snowberry didn't intersect any snowberry pixels in the classification, this indicates that we were unable to model that cover type correctly.  This information allowed us to intelligently aggregate the data classes to the final legend:

• Dry Conifer Forest

• Wet Conifer Forest

• Deciduous Trees

• Oak

• Oak/Douglas Fir

• Berry

• Riparian Woody Vegetation

• Mat/Floating Vegetation

• Fescue-Dominated Prairie

• Scot’s Broom

• Mixed Fescue/Scot’s Broom

• Water

Two different spatial aggregations were then performed to

A.    output to the national maps standards and

B.     modify the clump size of a cover type based on ecologically significance. 

A.  The national standard for Minimum Mapping Unit is roughly 20 pixels.  Which is to say that communities less than 20 pixels shall be aggregated into the surrounding class.  For IKONOS data, the result is a map of 1:1500 scale. 

B.  A single 4-meter x 4-meter pixel identified as Dry Coniferous Forest is likely an error, and if not an error, as single tree does not constitute a cover type.  Though, a single 4-meter x 4-meter identified as berry may be ecologically significant. By following the national map standards one doesn’t allow for the variability in cover type growth patterns and ecological significance.  Two different data products were output to allow Fort Lewis to determine the usefulness of each of these methods.

Accuracy Assessment

The time period of the project did not allow for extensive field verification of the map. However, PNNL did receive Land Condition Mapping (LCM) data from Fort Lewis ITAM office.  This data was collected by...methods.  Though the LCM data were not collected specifically with accuracy assessment in mind, they were determined to be useful for this purpose.  The accuracy assessment was performed in 2 different operations; using summary and point sampling.

Summary Method

This method involved the creation of a fescue and Scot’s broom image from the LCM points.  These images had 25-meter pixels, with pixel values ranging from 0-100 representing the component percent of the vegetation being represented.  We then ran a 5x5 majority filter over our high-resolution classification (4-meter pixels) to give it a similar resolution to the LCM image.  The high-resolution image data was then summarized by each of the LCM “Zones” e.g. 10 percent zone, 20 percent zone etc.  The results of the summaries are below in figures 1-3.

Figure 1.  Results of summarizing PNNL Fescue data by Fort Lewis ITAM Fescue data.

Figure 2.  Results of summarizing PNNL Scot’s Broom data by Fort Lewis ITAM Scot’s Broom data.

Note that the proportions of fescue and Scot’s Broom coverage measured via PNNL's expert rule classification methods, were in good agreement with the ground data.  As LCM fescue and Scot’s Broom percentage increased, the PNNL-derived coverage also increased.  It  also should be noted that though the number of PNNL disturbance pixels in the LCM disturbance zones 1-10, were in relative agreement, the absolute agreement of the points was not good.  This is likely due to the time difference between when the LCM crew collected their data, and when the satellite acquired the image.  The disturbed areas measured by the LCM crew likely would change (revegetate) more rapidly that the fescue or Scot’s Broom areas.  Also the data were collected in a slightly different manner according to the 1999 Land Condition Mapping Report “No effort was made to differentiate between the type, level or age of disturbance.  Severely disturbed prairie (i.e. areas of major soil disturbance that were bulldozed, excavated, graded or contained numerous ‘tank turns’) were given cover estimates on the same criteria as areas that received less visually damaging tracked vehicle effects.”   For this reason, the LCM disturbance data was left out of the accuracy assessment.

Point Sampling

The Land Condition Mapping data were useful in another way as well.  We prepared the data by selecting the LCM points that were 50% or greater of the cover type of interest (fescue, Scot’s Broom and disturbed).  These points were then compared against the high-resolution land cover map.  The results are summarized in the table below.  It should be noted that since we chose LCM points that contained 50% and greater of fescue, it was deemed appropriate that we should count the “hits” of PNNL fescue pixels as well as mixed Scot’s Broom/Fescue pixels.  The same was done for the LCM Scot’s Broom points.

Table 2.  Results of Accuracy Assessment
Results

Conclusion

We conclude that the high resolution images from satellite platforms are well suited for landcover classification.  When coupled with high-resolution DEM (LIDAR), the classification accuracy is increased for woody vegetation.  The December vegetation height data did not help in the classification of the grasses and forbs and low deciduous shrubs.  If this technique were applied on a site-wide basis, a hierarchical approach may be advantageous.  Wherein inexpensive, lower resolution (Landsat or SPOT) data would be acquired for the entire base and a classification could be developed for all non-prairie areas.  The high-resolution sensor would be tasked to image the prairie areas, and a higher order classification could be developed from this data for these sensitive and important areas.