Engineering topics
Publications and source records attributed to Davis, Frank W..
We explore the idea that key information about plant and soil diversity can largely be hidden in low-variance features, comparing to extensive ground truth at a highly diverse and well-studied California field site.
The biodiversity and high productivity of coastal terrestrial and aquatic habitats are the foundation for important benefits to human societies around the world. These globally distributed habitats need frequent and broad systematic assessments, but field surveys only cover a small fraction of these areas. Satellite-based sensors can repeatedly record the visible and near-infrared reflectance spectra that contain the absorption, scattering, and fluorescence signatures of functional phytoplankton groups, colored dissolved matter, and particulate matter near the surface ocean, and of biologically structured habitats (floating and emergent vegetation, benthic habitats like coral, seagrass, and algae). These measures can be incorporated into Essential Biodiversity Variables (EBVs), including the distribution, abundance, and traits of groups of species populations, and used to evaluate habitat fragmentation. However, current and planned satellites are not designed to observe the EBVs that change rapidly with extreme tides, salinity, temperatures, storms, pollution, or physical habitat destruction over scales relevant to human activity. Making these observations requires a new generation of satellite sensors able to sample with these combined characteristics: (1) spatial resolution on the order of 30 to 100-m pixels or smaller; (2) spectral resolution on the order of 5 nm in the visible and 10 nm in the short-wave infrared spectrum (or at least two or more bands at 1,030, 1,240, 1,630, 2,125, and/or 2,260 nm) for atmospheric correction and aquatic and vegetation assessments; (3) radiometric quality with signal to noise ratios (SNR) above 800 (relative to signal levels typical of the open ocean), 14-bit digitization, absolute radiometric calibration less than 2%, relative calibration of 0.2%, polarization sensitivity less than 1%, high radiometric stability and linearity, and operations designed to minimize sunglint; and (4) temporal resolution of hours to days. We refer to these combined specifications as H4 imaging. Enabling H4 imaging is vital for the conservation and management of global biodiversity and ecosystem services, including food provisioning and water security. An agile satellite in a 3-d repeat low-Earth orbit could sample 30-km swath images of several hundred coastal habitats daily. Nine H4 satellites would provide weekly coverage of global coastal zones. Such satellite constellations are now feasible and are used in various applications.
As part of research on forest ecosystems, the Jet Propulsion Laboratory (JPL) and collaborating research teams have conducted multi-season airborne synthetic aperture radar (AIRSAR) experiments in three forest ecosystems including temperate pine forest (Duke, Forest, North Carolina), boreal forest (Bonanza Creek Experimental Forest, Alaska), and northern mixed hardwood-conifer forest (Michigan Biological Station, Michigan). The major research goals were to improve understanding of the relationships between radar backscatter and phenological variables (e.g. stand density, tree size, etc.), to improve radar backscatter models of tree canopy properties, and to develop a radar-based scheme for monitoring forest phenological changes. In September 1989, AIRSAR backscatter data were acquired over the Duke Forest. As the aboveground biomass of the loblolly pine forest stands at Duke Forest increased, the SAR backscatter at C-, L-, and P-bands increased and saturated at different biomass levels for the C-band, L-band, and P-band data. We only use the P-band backscatter data and ground measurements here to study the relationships between the backscatter and stand density, the backscatter and mean trunk dbh (diameter at breast height) of trees in the stands, and the backscatter and stand basal area.
Decision makers need to know the reliability of output products from GIS analysis. For many GIS applications, it is not possible to compare these products to an independent measure of 'truth'. Sensitivity analysis offers an alternative means of estimating reliability. In this paper, we present a CIS-based statistical procedure for estimating the sensitivity of wildlife habitat models to uncertainties in input data and model assumptions. The approach is demonstrated in an analysis of habitat associations derived from a GIS database for the endangered California condor. Alternative data sets were generated to compare results over a reasonable range of assumptions about several sources of uncertainty. Sensitivity analysis indicated that condor habitat associations are relatively robust, and the results have increased our confidence in our initial findings. Uncertainties and methods described in the paper have general relevance for many GIS applications.
In recent years several analytical models were developed to predict microwave scattering by trees and forest canopies. These models contribute to the understanding of radar backscatter over forested regions to the extent that they capture the basic interactions between microwave radiation and tree canopies, understories, and ground layers as functions of incidence angle, wavelength, and polarization. The Santa Barbara microwave model backscatter model for woodland (i.e. with discontinuous tree canopies) combines a single-tree backscatter model and a gap probability model. Comparison of model predictions with synthetic aperture radar (SAR) data and L-band (lambda = 0.235 m) is promising, but much work is still needed to test the validity of model predictions at other wavelengths. The validity of the model predictions at P-band (lambda = 0.68 m) for woodland stands at our Mt. Shasta test site was tested.
A study is in progress comparing Airborne Synthetic Aperture Radar (AIRSAR) backscatter from coniferous forest plots containing gaps to backscatter from adjacent gap-free plots. Issues discussed are how do gaps in the range of 400 to 1600 sq m (approximately 4-14 pixels at intermediate incidence angles) affect forest backscatter statistics and what incidence angles, wavelengths, and polarizations are most sensitive to forest gaps. In order to visualize the slant-range imaging of forest and gaps, a simple conceptual model is used. This strictly qualitative model has led us to hypothesize that forest radar returns at short wavelengths (eg., C-band) and large incidence angles (e.g., 50 deg) should be most affected by the presence of gaps, whereas returns at long wavelengths and small angles should be least affected. Preliminary analysis of 1989 AIRSAR data from forest near Mt. Shasta supports the hypothesis. Current forest backscatter models such as MIMICS and Santa Barbara Discontinuous Canopy Backscatter Model have in several cases correctly predicted backscatter from forest stands based on inputs of measured or estimated forest parameters. These models do not, however, predict within-stand SAR scene texture, or 'intrinsic scene variability' as Ulaby et al. has referred to it. For instance, the Santa Barbara model, which may be the most spatially coupled of the existing models, is not truly spatial. Tree locations within a simulated pixel are distributed according to a Poisson process, as they are in many natural forests, but tree size is unrelated to location, which is not the case in nature. Furthermore, since pixels of a simulated stand are generated independently in the Santa Barbara model, spatial processes larger than one pixel are not modeled. Using a different approach, Oliver modeled scene texture based on an hypothetical forest geometry. His simulated scenes do not agree well with SAR data, perhaps due to the simple geometric model used. Insofar as texture is the expression of biological forest processes, such as succession and disease, and physical ones, such as fire and wind-throw, it contains useful information about the forest, and has value in image interpretation and classification. Forest gaps are undoubtedly important contributors to scene variance. By studying the localized effects of gaps on forest backscatter, guided by our qualitative model, we hope to understand more clearly the manner in which spatial heterogeneities in forests produce variations in backscatter, which collectively give rise to scene texture.
This paper discusses some basic scientific issues and research needs in the joint processing of remotely sensed and GIS data for environmental analysis. Two general topics are treated in detail: (1) scale dependence of geographic data and the analysis of multiscale remotely sensed and GIS data, and (2) data transformations and information flow during data processing. The discussion of scale dependence focuses on the theory and applications of spatial autocorrelation, geostatistics, and fractals for characterizing and modeling spatial variation. Data transformations during processing are described within the larger framework of geographical analysis, encompassing sampling, cartography, remote sensing, and GIS. Development of better user interfaces between image processing, GIS, database management, and statistical software is needed to expedite research on these and other impediments to integrated analysis of remotely sensed and GIS data.
The time-dependent correspondence between maps of surface brightness temperature derived from Landsat Thematic Mapper data and mapped terrain variables over a tallgrass prairie in northeastern Kansas is examined. Individual terrain variables including burning treatment, vegetation cover type (agriculture, prairie, woody vegetation), hillslope position, and greenness exhibit varying degrees of association with surface temperature. Burning treatment is most strongly associated with mid-morning surface temperature. Examination of terrain strata based on combinations of terrain variables, notably burning treatment and hillslope position, suggest that terrain variables interact in affecting surface temperature. Interaction between hillslope position, burning treatment, and surface temperature is more important in August than in May.
An ecological land classification was developed for a complex region in southern California using geographic information system techniques of map overlay and contingency table analysis. Land classes were identified by mutual information analysis of vegetation pattern in relation to other mapped environmental variables. The analysis was weakened by map errors, especially errors in the digital elevation data. Nevertheless, the resulting land classification was ecologically reasonable and performed well when tested with higher quality data from the region.
A method for exploring the topographic modulation of clear-sky incoming radiation using a two-stream atmospheric model together with digital elevation data is presented to obtain the temporal and spatial distribution of this energy. Utilizing this methodology for the Konza Prairie, it is observed that the variance and spatial autocorrelation of the simulated radiation data changed with sun angle and elevation grid spacing. An analytical formula that describes the behavior of the variance as a function of optical depth, sun angle, and mean terrain slope can be developed by considering direct solar radiation variability on a constant-slope terrain which has a uniform albedo and a uniform distribution of slopes. The decrease in mean slope and the increase in slope and autocorrelation is attributed to a loss of some of the shorter-wavelength terrain features at coarser grid spacings.
An analysis is made of time-dependent covariance of the greenness vegetation index with mapped terrain variables over the Konza Prarie (Kansas) during the 1987 growing season. The analysis was part of an ongoing project to establish appopriate ground-sampling and data-integration strategies for satellite-based monitoring of land surface climate conditions. Greenness images for six dates between May and October were derived from atmospherically corrected thematic mapper (TM) data and coregistered with maps of woody vegetation, fire, and soils. Local variance in greenness peaked in mid-June, falling rapidly until mid-August, and declining gradually thereafter. Greenness images exhibited positive autocorrelation up to distances of 180-210 m, but the dominant scale of pattern occurred at a block size of 60 m by 60 m throughout the growing season. 40-44 percent of total scene variance in July and August was accounted for by the effects of woody vegetation (8.9 percent of the area), prairie burning, and soil type. The effect of these terrain variables was fairly consistent between June and late August and was manifested as additional high-frequency spatial variation in imagery from that period.