Image Navigation for Wildland Fire Location Mapping

Loey Knapp, Patricia Andrews, and John Turek

Image Navigation for Wildland Fire Location Mapping

Paper for the NCGIA Conference on GIS and Environmental Modeling, January 21-24, 1996, Santa Fe, New Mexico.



ABSTRACT


Efficient response to wildland fires is of critical concern to various state and federal agencies, and requires a substantial data archive from which to draw essential information; examples are maps of existing regional fires, fuel type or vegetative greenness, and populated areas that might be threatened. For effective response to the fire, these data must be up to date and easily accessible for prediction, planning, and resource allocation but typically this is not the case. While spacecraft imagery has long held the promise of global data acquisition and real time spatial information, the analysis process has proved to be a bottleneck due to the vast amounts of data and the size of the datasets. The challenge in the area of emergency response is to provide the technology to extract crucial information from images quickly enough to influence the decision-making process.

This paper will describe new technology in the area of content-based search of images which can be applied to emergency preparedness and emergency response. This technology will be discussed in light of a specific emergency, that of wildland fire.


INTRODUCTION


Wildland fire management in the United States is the responsibility of various agencies ranging from federal and state to rural and private. Fire, however, doesn't recognize administrative boundaries. This has led to interagency and international cooperation in detection and response to wildfires. In some cases, dispatchers and coordinators from various agencies are co-located to facilitate cooperation and information sharing. Nevertheless there is room for improvement in tracking fire activity on a national or regional level across all land ownerships as well as in archiving historical fire data. National or regional mapping of fire locations from satellite data would be useful additional information if it were easily accessible across multiple agencies and up to date; if it can be combined with other ancillary data such as topography, urban areas, and administrative boundaries; and if it can be used in conjunction with fire potential models.

Efforts to provide regional and/or national information on fires have taken different forms. The first, official collection and distribution mechanism is through the National Interagency Fire Center (NIFC) in Boise, Idaho. Local agency reports are fed to NIFC for consolidation, analysis, and reporting purposes. Daily fire situation reports from NIFC include a narrative on fires and areas of interest plus summaries of number of fires and acres burned by agency and region of the U.S. and also for Canadian Provinces. (The reports can be found on the Forest Service home page, http://www.fs.fed.us under "Caring for the Land."). The reports are compiled from intelligence information collected from geographic area coordination centers and are dependent on agency-specific input and cooperation.

Another source of wildland fire information is satellite imagery, a source of data which crosses political boundaries. The National Geographic Data Center (NGDC) within the National Oceanographic and Atmospheric Administration (NOAA) has developed classification algorithms for fires using the Defense Mapping Satellite Program (DMSP) operational linescan (OLS) sensor. These algorithms are used on the visible band of the night orbit and are adjusted for clouds, city lights, and lightning (Elvidge and others, in press). This technology could result in an Internet product of fire location maps for the U.S. on a daily basis.

The use of satellite imagery provides national data and avoids problems of reporting delays and consolidation, but a daily map product also has its limitations. A map product is by definition a predefined display. As in any GIS work, the capability to interactively search the imagery in conjuntion with ancillary data sets, would provide the analyst with more information on spatial relationships. In the case of wildland fire, the search must include temporal functionality as time is an essential element of wildland fire analysis. The technology to search images by color, texture, and shape has been developed and is used in multimedia software. An extension of this concept can be used to search multispectral images based on classification algorithms, color, or spectral signature.

Using a wildland fire classification algorithm an analyst could search a sequence of images for current fire locations, examine the temporal extent of a fire, or compare different years/seasons. Including additional classification algorithms the analyst could pose additional queries to determine wildland fires near urban areas or to find other spatial areas exhibiting similar characteristics to those currently burning, e.g. very dry with heavy fuel load. The resulting map of fire locations, combined with ancillary spatial data sets, could provide supplemental information such as fire on all land ownerships or fires per land cover type. Use of satellite-based fire location maps, like the NIFC fire situation reports, would be in monitoring the broad view of fire activity, and not in the management of individual fires.

Provision of an image search capability for wildland fires via the Internet would address the interagency aspect of the wildland fire problem. The primary advantage of Internet over private networks is that it provides general accessibility to information. This accessibility makes Internet useful for individuals or groups which are organizationally unrelated but have a common concern relative to a particular issue, such as the USDA Forest Service, the Bureau of Land Management, and the State forestry department's common interest in wildland fires. Providing remotely-sensed wildland fire information via the internet would solve two problems; 1) inclusion of regional or national in addition to agency data, and 2) accessibility to that data to state and federal, or public and private, analysts.

The rest of this paper will discuss the potential application of image search in the area of wildland fire location mapping, the application-specific challenges involved in implementing such a system, and the technology challenges which must be overcome for Internet-based image search to meet the implementation needs. This discussion will reference a system currently being architected by IBM that will be tested in a wildland fire application as a USDA Forest Service research project. The purpose of the collaboration is to test new technology in an implementation scenario and, by doing so, determine operational roadblocks.


WILDLAND FIRE LOCATION MAPPING


Primary users of a broad area fire location map product are expected to be in interagency coordination centers. Image queries and the resulting displays can be used as a focus of discussion during morning briefings to agency administrators. The information can be used in assessing the overall fire situation as well as in setting priorities for distribution of limited resources including aircraft, crews, and supplies. Managers in regional centers can view the fire activity in their area of responsibility compared to other areas of the country. The maps and related information can also be provided to the media, the public, and special interest groups, aiding coordination center personnel in carrying out an important function during times of critical fire activity.

In addition to timely fire intelligence information, fire history data can also be queried. Satellite data could be used to determine correct lat/long locations for fires, overcoming errors that often occur in determining and recording fire location. Wildfire data are archived by each U.S. federal agency and by States in separate data bases. Satellite data, of course, can't provide information on cause, cost, etc. so it is not a subsititute for those data bases. But an interagency fire location data base derived from the satellite data would aid post season analysis and planning activities.

Fire location mapping on a broad area will be especially useful when it is combined with GIS ancillary data such as political boundaries (State, county, etc.), land ownership (Forest Service, National Park Service, State, etc.), cities and roads, lakes and rivers, lat/long grid, and vegetation and fuel type. Besides giving a reference for location of fire activity, this information can be used in data searches. A fire manager might want to identify and display fires on Bureau of Land Management land or those in wilderness areas, fires within a specified distance of major cities, or fires that are on forested land.

Links to fire models will add information on fire potential to that on current fire activity. Fire locations can be overlayed on data layers generated by Wildland Fire Assessment System (WFAS), the next generation U.S. fire danger/behavior system (Andrews and others, in press). Products that are currently available to fire managers include a weekly greenness map derived from AVHRR satellite data indicating the state of live vegetation (Burgan and Hartford 1993, Hartford and Burgan in press), a daily fire danger map derived from weather taken at fire weather stations through the U.S., and a daily Haines Index (stability and dryness) map derived from upper air soundings from all North American weather service stations. The capability to "zoom in" to access data on topography and fuel type, would allow a link to site-specific fire behavior models (Andrews 1986) and fire growth simulation models (Finney in press, Finney and Andrews in press).


APPLICATION CHALLENGES

Implementation of an Internet-based image search capability for wildland fires poses several challenges. The first of these is in the area of data. Various sensors can provide information on wildland fires but no single sensor has the spatial resolution, temporal resolution, and spectral range required to provide all the information required by fire analysts. It may be that a data archive specific to wildland fire search will need to include the output from multiple sensors and will also necessitate cross-sensor query capability.

Prevedel (1994) reported on the use of AVHRR satellite data to monitor wildfire activity in several western States for 45 days in 1994. Maps were provided to Multi-Agency Coordinating (MAC) groups that were formed to deal with the severe fire season. In addition to a description of the development and use of the map product, he pointed out some limitations of the satellite data. Sun reflection, lightning, and rocks heated by solar radiation, for example, can be difficult to distinguish from fires. The 1.1 kilometer resolution is also a drawback to identifying and locating small fires.

The NGDC project has made significant progress with the Defense Meteorological Satellite Program (DMSP) operational linescan (OLS) data which uses night-time visual sensors. This sensor eliminates some of the problems with the use of AVHRR, but suffers from some of its own. This data has a spatial resolution of .5 kilometer, an improvement over AVHRR, and comparative studies between the sensors indicate that more fires were located with DMSP OLS data (Elvidge and others, in press). However, fires in urban areas can be missed and thick smoke can keep the satellite from locating fires. Also, the use of the visible band from the night orbit limits the temporal resolution to daily observations.

The Geostationary Operational Environmental Satellite Visible Infrared Spin Scan Radiometer Atmospheric Sounder (GOES VAS) sensor, which provides measurements in the visible and infrared regions, is another potential fire data source. The temporal resolution of 30 minutes is attractive but the spatial resoluton of .9 kilometers in the visible region and 6.9-13.8 kilometers in the infrared region limits the sensors use to location of very large fires or the monitoring of diurnal cycles of fire activity.

Research on the use of satellites for detecting and locating wildland fires is ongoing. When better techniques have been developed, tested, and become available, they will be used. Investments in two-meter resolution sensors are likely to bring significant payoff in this area towards the end of the century.

The second challenge is in the area of data representation. Prevedel noted that distribution of raw uninterpreted satellite data can lead to misinterpretation of the fire situation. For example, if a pixel is saturated it will indicate the presence of fire throughout the pixel regardless of the true spatial extent of the fire, thus overestimating the area of the fire. On the other hand, the inability of this technology to detect some fires due to size, clouds, smoke or other sensor problems will lead to an underestimation of both number of fires and area burning. Finally, 'counting' fires may cause problems as more than one small fire might be alight within a pixel or a single fire may have a gap, visually indicating two fires instead of one.

Due to these problems it is essential to consider the representation of the information to fire analysts and response teams. For instance, glyphs might be used to indicate a fire location rather than number of pixels beyond the threshold value. Color selection and provision of contextual information such as political boundaries must also be considered. These factors indicate the need for visualization capabilities beyond predefined image display which allow the user to specify the nature of the data representation. Such features are common in desktop software but not yet provided on the Internet.


TECHNOLOGY CHALLENGES

From a technology perspective, the application outlined above raises serious challenges. An interactive Internet-based image search system comprises several key elements including user-specified queries and displays, image search through classification algorithms or user defined parameters such as color, and data retrieval (download). In the latter case it is important to be able to reconstruct the image in a lossless manner as many applications cannot afford the loss of data.

These system elements cannot be implemented without the following technical capabilities:

  1. compression with associated decompression to a lossless image
  2. a spatial query language
  3. classification algorithms which can be applied to compressed data
  4. simple GIS functionality including georegistration, resampling, and aggregation against compressed data,
  5. visualization operations, also for use with compressed data.

System Structure

The overall organization of the system can be seen in Figure 1. (not yet available) Clients connect to servers using standard internet protocols (i.e., http). This approach takes advantage of and builds on the tremendous infrastructure that is in the process of being developed.

Client Side Interface

The internet provides a common interface for access to information through standard Web Browsers like Mosaic or Netscape. Due to the availability of this interface on multiple platforms it would be desirable to base a system approach entirely on standard Web browsers. Unfortunately, even though new capabilities are constantly being added (e.g., JAVA) it is unlikely that browsers will provide the basis required to fully support the image manipulation required by a generic emergency management system. In particular, visualization tools remain particularly weak. The system under development takes a two-tiered approach. Limited functionality is available through a standard Web Browser while users who need to run complex queries can download an enhanced browser.

The enhanced browser supports a language that allows visualization and parameter specification to be integrated with requests to the server. One of the goals of the language is to preserve the illusion of a single session even if the session spans multiple requests to the server. More importantly, the language allows the user to dynamically define semantics of the queries and tailor the system to the specific needs of the environment with minimal effort.

Server Organization

The system server is based on the use of a standard http daemon. The server initiates a session by invoking the query parser described along with the rest of the major system components below: \begin{itemize} \item {\em Query Parser}: Queries that come into the system need to provide a mechanism that can freely specify new models for extracting features and relate them back to content. The query parser offers a general purpose language based on image set manipulations that allows for this kind of interaction. \item {\em Data Manager}: The data manager provides support for all ancillary data and indices. The current version of the manager uses IBM's DB2 engine. We will be providing extensions to experiment with the effectiveness of different index structures. \item {\em Image Manager}: Satellite and GIS data typically come in raster format that can be viewed as an "image" or a 2-dimensional lattice. The Image manager stores N-dimensional lattice data in transformed/compressed format that can be used to facilitate image analysis and feature extraction for content based search. This is described in further detail below. \item {\em Image Processing Engine}: The image processing engine is used to provide the ability to filter, process and do feature extraction on images during the search and retrieval process. These operations need to be closely tied to the representation of the data on disk as described below. \item {\em Visualization Engine}: Although it is necessary to provide visualization tools on the client side, in some cases it will be more effective to provide the final product directly from the server side. The current version of the visualization engine uses IBM's Data Explorer product. \end{itemize}

Challenges

The management of the large and diverse kinds of data required by an emergency management system pose significant technical challenges. Some of these are outlined in the remainder of this section.

Image Compression for Data Storage

In order to fully understand the nature of the problem at hand consider the following example: Landsat Thematic Mapper scenes are available at a resolution of 30 meters per pixel containing 7 spectral bands. These data could be quite useful in the event of an emergency, aiding in the location of urban/wildlife interfaces, water sources, etc.

Full coverage of the United States with Landsat imagery would require on the order of 100GB. Unfortunately, these images usually have partial cloud cover and so several scenes from each location need to be stored in order to ensure that timely data are available on the area of interest. In the near future, new commercial satellites are expected to have a resolutions down in the 2 meter range thereby increasing the storage requirements by 2 orders of magnitude.

Other sources of data, such as the OLS sensor that is one possible source of fire data, have reduced spatial resolution but increased temporal resolution. Thus, even though the price of storage devices will continue to drop at a dramatic rate, the cost of storing satellite images in an operational system will continue to be the dominant cost of the system. Proper use of compression can dramatically reduce this cost.

Compression techniques can be either {\em lossless} or {\em lossy}. A lossless compression scheme is one which guarantees perfect reconstruction of all of the bits in the original dataset. A lossy compression scheme, on the other hand, does not reconstruct most images exactly but rather allows the loss of some information in order to achieve higher compression ratios. The assumption is that only "unimportant" information is lost. Because it is difficult to determine exactly what information will be relevant in the analysis stage, our compression techniques need to provide a progressive framework that allows users to select exactly the level of loss that they can tolerate. The system needs to provide for lossless retrieval of data. The compression mechanisms under development achieve lossless compression that is comparable to the best lossless compression schemes currently available for image products while still maintaining the ability to do progressive retrieval.

Analysis of Compressed Imagery

Although extraction of relevant features at the time of data ingest into a predefined schema is an important part of any information system, it is by no means sufficient. The needs of each region and office can be quite different. As an example, the algorithm that is used for identifying fires in OLS data relies on detecting sources of visible and near-infrared emission on the earth's surface in the absence of solar illumination (i.e., at night). While this algorithm can be readily applied in places such as Brazil and parts of the continental U.S., it can not be applied in places such as Northern Canada and Alaska.

In essence the feature extraction process is by its very nature lossy and cannot represent all of the content contained in the imagery products that feed into the analysis required by the different regions. Furthermore, the processing involved in deriving the attributes can be quite expensive and techniques to make this processing more efficient are required. Thus, although useful and necessary, the use of a predefined schema will be insufficient to adequately support the search mechanisms required by an emergency management system. It is necessary to provide the functionality that will allow the user to visualize, define and extract features dynamically thereby performing content-based search in real-time on the image products rather than assuming that all information will be available at the time of the emergency.

Although it would seem that the cost associated with this processing is higher when images are stored using compression, this need not be the case; by laying out the data in an appropriate fashion, the search process actually becomes more efficient and since it is not necessary to look at all the bits associated with the original image. The primary technical idea behind our project is that it is possible to increase the speed of searching through images stored in a digital library while simultaneously reducing the storage requirements.

Search, Manipulation, and Distribution via Internet

Another problem that arises is based on the limited bandwidth imposed by our transmission medium. In the foreseeable future it is expected that most offices will be connecting through telephone at relatively slow transmission rates. As a result, it is essential that only the relevant information be returned to the user. The system provides this functionality via the integration of several mechanisms including server based search and content extraction, caching on the client and compression.

As one example, when dealing with complex images, it is important to be able to {\em navigate} through the image. Figure 2 (not yet available) shows a portion of a Landsat TM image of the Indiana/Kentucky border. The figure shows a scenario in which a manager can start with a large area view of a region of interest: i.e, they start with a 128x128 pixel representation of a 2048x2048 portion of the image. Here, one can clearly see the major features of the landscape including the Ohio river but cannot identify other relevant features. At each step of zooming in the resolution is doubled, while the size of the browse product remains unchanged. As the resolution increases, it becomes easier to identify features such as mountains, cities and roads for managing site specific information. By maintaining small viewing windows, the user has managed to isolate the information of interest in an interactive fashion yet the cost of data transmission has been kept low. Even across telephone lines, each iteration of the search will take only a few seconds.

Heterogeneous data sets

One substantial difficulty is the need to deal with different data sources and types, for example, vector data for maps, raster data for digital elevation maps and satellite imagery. Even when the data sources are similar, for example all raster products, the management and mapping of different resolutions pose additional difficulties. The prototype in developement provides a framework that allows the integration of additional tools and the coordination of data migration between those tools.

Scalability

Scalability is a serious concern for a project of this type. It is likely that the amount of data required for an operational environment and the user load on the system will pose major challenges. As part of the technology that is being applied to the system server, IBM is exploring the use of parallelism and distribution to support larger user groups and the use of hierarchical storage formats to support large data volumes.


CONCLUSION


Numerous problems involve multiple agencies and require a centralized data archive specific to that problem but accessible by all of those concerned, public and private alike. One such problem is wildland fire, which is of interest to federal, state, and private land management groups. Other such problems are noxious weed infestation, agricultural plant health, and response to natural hazards. Within these problem areas, data collection across political boundaries and accessibility to the data can be major issues. The use of satellite imagery, accessed via the Internet, may address these problems.

In this paper, we have outlined the application of new technology in content-based search of satellite imagery over the Internet to the area of wildland fire. Significant application and technical challenges must be addressed for this technology to become operational. With successful resolution of these challenges, however, content-based search can be applied across a range of emergencies as well as to other pressing cross-agency problems.


ACKNOWLEDGEMENTS

This work is partially supported by NASA contract NCC5-101


REFERENCES


Andrews, Patricia L., (1986) BEHAVE: Fire Behavior Prediction And Fuel Modeling System-- BURN Subsystem, Part 1. General Technical Report INT-194. Ogden, UT: U.S. Department of Agriculture, Forest Service, Intermountain Research Station. 130 p.

Andrews, Patricia L., Bradshaw, Larry S., Burgan, Robert E., Chase, Carolyn H., and Hartford, Roberta A., (in press) WFAS: Wildland Fire Assessment System--Status 1995, Presented at 1995 Interior West Fire Council Meeting, St. George, Utah. Nov. 1-3, 1995.

Burgan, Robert E., and Hartford, Roberta A., (1993) Monitoring Vegetation Greenness With Satellite Data. Gen. Tech. Rep. INT-297. Ogden, UT: U.S. Department of Agriculture, Forest Service, Intermountain Research Station. 13 P.

Castelli, Vittorio, C.S> Li, I. Kontoyiannis, and J.J. Turek, Progressive Classification of Satellite Images in a Wavelet Framework, NJIT, September 22, 1995.

Elvidge, Christopher D., Herbert W. Kroehl, Eric A. Kihn, Kimberley E. Baugh, Ethan R. Davis, and Wei Min Hao, (in press) Algorithm for the Retrieval of Fire Pixels from DMSP Operational Linescan System Data, "Global Biomass Burning", edited by Joel S. Levine.

Finney, Mark A. (in press) FARSITE, A Fire Area Simulator for Fire Managers. Presented at The Biswell Symposium, February 15-17, 1994, Walnut Creek CA.

Finney, Mark A., and Andrews, Patricia L., (in press) The FARSITE fire area simulator: Fire management applications and lessons of summer 1994. Presented at 1994 Interior West Fire Council Meeting and Symposium, Coeur d'Alene, ID, November 1-3, 1994.

Hartford, Roberta A., and Burgan, Robert E., (in press) Vegetation Condition and Fire Occurrence: A Remote Sensing Connection. A Paper presented at the Interior West Fire Council Meeting and Symposium, Coeur d'Alene, ID, November 1-3, 1994.

Prevedel, David A. (1994) Project Sparkey: A Strategic Wildfire Monitoring Package Using AVHRR Satellites and GIS Photogrametric Engineering and Remote Sensing 60:1 271-278.


Loey Knapp, Senior Analyst
IBM Corporation
6300 Diagonal Highway
Boulder, CO 80301
phone: 303-924-0482
fax: 303-924-0518
email: loey@vnet.ibm.com


Patricia Andrews, Project Leader
USDA Forest Service
Intermountain Research Station
P.O. Box 8089
Missoula, MT 59807
phone: 406-329-4827
fax: 406-329-4825


John Turek, Project Leader
IBM T.J. Watson Research Lab
P.O. Box 704
Yorktown Heights, N.Y. 10598
phone: 914-784-7591
fax: 914-784-6031
email: jjt@watson.ibm.com