2335 Irving Hill Road

University of Kansas

Lawrence, KS 66045

High spatial and temporal resolution imagery that can be captured at user specified time periods is often critical to effective resource management planning. In this poster paper, the authors describe a new low cost (less than $30,000) multispectral digital imaging camera system. The MS3100 imaging camera is a three band, 1392 x 1040 pixel imaging system that was custom designed by DuncanTech. This paper describes how scientists at the University of Kansas integrated the camera into an airborne digital imaging system that can capture multispectral imagery from a small aircraft. System development is described and sample data are presented.

High spatial and temporal resolution imagery that can be captured at user specified time periods is often critical to effective resource management planning. In this paper, we describe a multispectral digital imaging camera system that we experimentally flew over several study areas on November 14, 2000. The experiment goals were threefold: 1) to assess camera functionality by capturing imagery over a wide variety of terrain including grasslands, woodlands, wetlands, portions of the Kansas River, the University of Kansas (KU) campus, and the Kansas Ecological Reserves, an area maintained by KU to conduct ecological studies; 2) to calibrate the DuncanTech MS3100 digital camera and its computer system for future multitemporal imagery acquisition; and, 3) to provide supplemental information for interpreting images acquired over the same area by the HyMap sensor three weeks prior.

System development comprised four stages: data needs assessment, acquisition of system components, system integration, and system testing.

To accomplish current and future research objectives, we needed a way to get timely high-resolution (both spatially and temporally) images of vegetation over the Kansas Ecological Reserves and portions of the surrounding area. A survey of existing multispectral high-resolution imaging systems showed that the MS3100 digital camera from DuncanTech was would provide high-resolution multispectral imagery and be cost effective (Figure 1).

Computer, camera, inverter, and monitor were mounted to a wooden pallet and connected together. The batteries were mounted on a separate pallet and automotive battery cables used to connect the batteries to the inverter. After determining available mount points in the single engine Cessna 181 airplane, holes were drilled in the pallet and the pallets secured to the rear seat rails on the floor of the aircraft (Figure 3). The camera, mounted on a standard camera tripod, was position over a hole cut in the bottom of the airplane and secured with nylon straps. The person sitting in the passenger seat controls the entire imagery acquisition process using a mini keyboard with touch pad mouse.

Initial ground tests of the system showed that all hardware and software components of the system worked very well. These tests also showed that the batteries could power operation of the system for more than eight hours.

Prior to the test flight, field data were collected at nearly 250 locations throughout the study areas to provide a ground truth data set. The test flight itself was conducted on November 14, 2000 between 10:00am and 2:00pm. Although this date is extremely late in the growing season for the Central Plains, we did not want to wait until the spring to flight test the system and risk losing flight time to "iron out" any problems. Six flight lines were flown covering the Kansas Ecological Reserves as well as the grassland and forested areas north of the Reserves. These flights were flown at an altitude of 10,000 feet, resulting in a spatial resolution of approximately 1 meters. Three other flight lines were flown over the Kansas Ecological Reserves at altitudes of 5,000, 3,000, and 1,000 feet. However, due to the lower altitude and the speed of the aircraft (approximately 180 ft. per second), the imagery from 3,000 and 1,000 feet was quite blurry and deemed unusable. Additional imagery was obtained at 5,000 feet (0.5-meter resolution) over the Haskell-Baker wetlands, the campus of the University of Kansas, and portions of the Kansas River. A total of over 1,600 images (4.2Mb each) were collected over the various study areas.

Images obtained during the test flight were downloaded from the flight computer to a ground computer running image processing and graphics software. The images were viewed individually to identify sequential images and areas of overlap. Sequential images were manually rotated and positioned so that they appeared as a continuous chain of images. The geographic locations of the image chains were visually identified and plotted on a Landsat Thematic Mapper image to serve as an index and reference map, thereby, making image chains over specific regions of interest readily accessible for further analysis. The imagery examples in this paper were selected from a site just north of Lawrence, KS, near the Kansas Ecological Reserves. This location contains a mixture of warm season and cool season grasses along with patches of deciduous and coniferous trees. (Figure 5a).

In addition to simply testing the system’s flight worthiness, assessing the quality of the camera’s multispectral data was a high priority. One exercise in this assessment was to calculate the Normalized Difference Vegetation Index (NDVI) ((band 1 – band 2)/(band 1 + band 2) (Figure 5b). Calculation of the NDVI allowed us to quickly identify areas with similar greenness values and, in conjunction with field data, make qualitative assessment of specific vegetation type and condition (e.g., cool vs. warm season grass and senescent leaf-on vs. senescent leaf-off,). After calculating the NDVI, another exercise looked at using the data to perform land cover classification. An unsupervised classification was performed on three different images: the original three-band image, the NDVI image, and a four-band image (the original three-band image plus NDVI).

The three and four-band classified images were very similar to one another and both did a good job of separating cool and warm season grasses as well as deciduous and coniferous trees. The classified NDVI image also separated cool and warm season grasses, however, there was considerable confusion between deciduous forests and cool season grasses. This confusion was likely due to the fact that both cover types were dormant at the time of imaging, therefore, both cover types had relatively low NDVI values. In all three of the image classifications, the influence of slope, aspect, and shadows had an impact on the classification, and increased the frequency of confusion between cool season grasses and deciduous forests. Visual comparison between the three classified images, the raw imagery, and field data found that the four-band image produced the best initial classification results.

Post-classification processing of the four-band image was conducted to improve the original classification. Cluster busting techniques were used to further separate grasslands from deciduous forests, and water from heavily shaded deciduous forests. This process greatly increased the accuracy of water bodies, however, only moderate improvements were achieved separating the grassland/forest mixed classes. The final step in post processing was the application of a 3x3 moving median filter to remove some of the "salt and pepper" from the image to produce a smoother appearing image.

The final classified image (Figure 5c) is a decent representation of the major land cover patterns found within the study area. The major patches of warm-season grasses (Andropogon gerardii and Andropogon scoparius), cool-season grasses (Bromus inermis and Festuca arundinacia), and forested lands, are clearly visible and closely follow the land cover patterns observed in the field as well as those apparent in the raw imagery. Despite the post classification processing, some confused land-cover classes were still present. North facing slopes of deciduous forests were often confused with warm-season grasses, and conversely, south facing slopes of warm-season grasses were confused with deciduous forest. These misclassifications may be due to the fact that the trees, despite being dormant, still retained most of their leaves. This condition, unusual for mid November, may have produced a spectral reflectance pattern similar to dormant grass litter that was further confused by varying light intensity as a result of slope aspect.

The results of this research show that investing financial and personnel resources into an airborne multispectral digital imaging system can be a cost-effective and efficient method for conducting research requiring such data. Despite the late date for acquiring the imagery, initial processing and classification of the data indicates that the camera will be useful for a wide variety of research objectives requiring on-demand, high-resolution data. Some potential applications of the imagery are: 1) to identify localized indications of stress in the development of both natural and agricultural vegetation; 2) to identify small-scale changes in natural composition due to succession or anthropogenic stress; and, 3) to monitor urban development patterns and assist in the planning of efficient road and utility routes. Another promising use of this imagery, though not previously addressed in this paper, is the use of data from the blue band for water quality studies. Because water is relatively transparent around the blue wavelength (450-520nm), this system may prove to be very useful for measuring submerged aquatic vegetation, turbidity, and channel structure in water bodies such as streams, rivers, and lakes. The most useful feature of this imaging system, however, is its low overall cost and its flexibility to meet user needs. Because it is mounted in a small airplane, imagery can be flown over specific regions at varying altitudes to produce whatever spatial and temporal resolution are determined to be important. Researchers can now obtain multispectral data when and where they want it instead of being limited by a satellite’s relatively rigid imaging schedule.

Richardson, A. J., J.H. Everitt, and D.E. Escobar. (1993). Reflectance calibration of aerial video imagery with automatic gain compensation on and off. International Journal of Remote Sensing. 14(15):2791-2801.