Chengquan Huang

Development of a 2001 National Land Cover Database for the United States

Multi-Resolution Land Characterization 2001 (MRLC 2001) is a second-generation Federal consortium designed to create an updated pool of nation-wide Landsat 5 and 7 imagery and derive a second-generation National Land Cover Database (NLCD 2001). The objectives of this multi-layer, multi-source database are two fold: first, to provide consistent land cover for all 50 States, and second, to provide a data framework which allows flexibility in developing and applying each independent data component to a wide variety of other applications. Components in the database include the following: (1) normalized imagery for three time periods per path/row, (2) ancillary data, including a 30 m Digital Elevation Model (DEM) derived into slope, aspect and slope position, (3) perpixel estimates of percent imperviousness and percent tree canopy (4) 29 classes of land cover data derived from the imagery, ancillary data, and derivatives, (5) classification rules, confidence estimates, and metadata from the land cover classification. This database is now being developed using a Mapping Zone approach, with 66 Zones in the continental United States and 23 Zones in Alaska. Results from three initial mapping Zones show single-pixel land cover accuracies ranging from 73 to 77 percent, imperviousness accuracies ranging from 83 to 91 percent, tree canopy accuracies ranging from 78 to 93 percent, and an estimated 50 percent increase in mapping efficiency over previous methods. The database has now entered the production phase and is being created using extensive partnering in the Federal government with planned completion by 2006.

An assessment of support vector machines for land cover classification

The support vector machine (SVM) is a group of theoretically superior machine learning algorithms. It was found competitive with the best available machine learning algorithms in classifying high-dimensional data sets. This paper gives an introduction to the theoretical development of the SVM and an experimental evaluation of its accuracy, stability and training speed in deriving land cover classifications from satellite images. The SVM was compared to three other popular classifiers, including the maximum likelihood classifier (MLC), neural network classifiers (NNC) and decision tree classifiers (DTC). The impacts of kernel configuration on the performance of the SVM and of the selection of training data and input variables on the four classifiers were also evaluated in this experiment.

Derivation of a tasselled cap transformation based on Landsat 7 at-satellite reflectance

A new tasselled cap transformation based on Landsat 7 at-satellite reflectance was developed. This transformation is most appropriate for regional applications where atmospheric correction is not feasible. The brightness, greenness and wetness of the derived transformation collectively explained over 97% of the spectral variance of the individual scenes used in this study.

An approach for mapping large-area impervious surfaces: synergistic use of Landsat-7 ETM+ and high spatial resolution imagery

A wide range of urban ecosystem studies, including urban hydrology, urban climate, land use planning, and resource management, require current and accurate geospatial data of urban impervious surfaces. We developed an approach to quantify urban impervious surfaces as a continuous variable by using multisensor and multisource datasets. Subpixel percent impervious surfaces at 30-m resolution were mapped using a regression tree model. The utility, practicality, and affordability of the proposed method for large-area imperviousness mapping were tested over three spatial scales (Sioux Falls, South Dakota, Richmond, Virginia, and the Chesapeake Bay areas of the United States). Average error of predicted versus actual percent impervious surface ranged from 8.8 to 11.4%, with correlation coefficients from 0.82 to 0.91. The approach is being implemented to map impervious surfaces for the entire United States as one of the major components of the circa 2000 national land cover database.

Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error

Abstract We developed a global, 30-m resolution dataset of percent tree cover by rescaling the 250-m MOderate-resolution Imaging Spectroradiometer (MODIS) Vegetation Continuous Fields (VCF) Tree Cover layer using circa- 2000 and 2005 Landsat images, incorporating the MODIS Cropland Layer to improve accuracy in agricultural areas. Resulting Landsat-based estimates maintained consistency with the MODIS VCF in both epochs (RMSE =8.6% in 2000 and 11.9% in 2005), but showed improved accuracy in agricultural areas and increased discrimination of small forest patches. Against lidar measurements, the Landsat-based estimates exhibited accuracy slightly less than that of the MODIS VCF (RMSE=16.8% for MODIS-based vs. 17.4% for Landsat-based estimates), but RMSE of Landsat estimates was 3.3 percentage points lower than that of the MODIS data in an agricultural region. The Landsat data retained the saturation artifact of the MODIS VCF at greater than or equal to 80% tree cover but showed greater potential for removal of errors through calibration to lidar, with post-calibration RMSE of 9.4% compared to 13.5% in MODIS estimates. Provided for free download at the Global Land Cover Facility (GLCF) website (www.landcover.org), the 30-m resolution GLCF tree cover dataset is the highest-resolution multi-temporal depiction of Earths tree cover available to the Earth science community.

Beware of per-pixel characterization of land cover

A simulation experiment was carried out to analyse the effects of the modulation transfer function on our ability to estimate the proportions of land cover within a pixel by linear mixture modelling. In the simulated landscape the proportion of each land cover type in every pixel was known exactly. The standard error of the estimate (SEE) between percentages derived from mixture modelling and the actual land cover percentages was 11%. Substantial improvements in estimating the percentages can be obtained simply by deriving estimates for pixels of twice the original dimensions, the SEE dropping to 4.16%, though this is with the obvious consequence of a final product with a coarser spatial resolution. Alternatively by deconvolving the input bands using a linear approximation of the point spread function the SEE can be reduced by almost as much, namely to 5.11%. If we combine the two approaches, by first doconvolving the bands, estimating the percentages and then aggregating resultant pixels to twice their original linear dimensions, the SEE drops to 2.24%.

Global characterization and monitoring of forest cover using Landsat data: opportunities and challenges

Abstract The compilation of global Landsat data-sets and the ever-lowering costs of computing now make it feasible to monitor the Earths land cover at Landsat resolutions of 30 m. In this article, we describe the methods to create global products of forest cover and cover change at Landsat resolutions. Nevertheless, there are many challenges in ensuring the creation of high-quality products. And we propose various ways in which the challenges can be overcome. Among the challenges are the need for atmospheric correction, incorrect calibration coefficients in some of the data-sets, the different phenologies between compilations, the need for terrain correction, the lack of consistent reference data for training and accuracy assessment, and the need for highly automated characterization and change detection. We propose and evaluate the creation and use of surface reflectance products, improved selection of scenes to reduce phenological differences, terrain illumination correction, automated training selection, and the use of information extraction procedures robust to errors in training data along with several other issues. At several stages we use Moderate Resolution Spectroradiometer data and products to assist our analysis. A global working prototype product of forest cover and forest cover change is included.

Impact of sensor's point spread function on land cover characterization: Assessment and deconvolution

Measured and modeled point spread functions (PSF) of sensor systems indicate that a significant portion of the recorded signal of each pixel of a satellite image originates from outside the area represented by that pixel. This hinders the ability to derive surface information from satellite images on a per-pixel basis. In this study, the impact of the PSF of the Moderate Resolution Imaging Spectroradiometer (MODIS) 250 m bands was assessed using four images representing different landscapes. Experimental results showed that though differences between pixels derived with and without PSF effects were small on the average, the PSF generally brightened dark objects and darkened bright objects. This impact of the PSF lowered the performance of a support vector machine (SVM) classifier by 5.4% in overall accuracy and increased the overall root mean square error (RMSE) by 2.4% in estimating subpixel percent land cover. An inversion method based on the known PSF model reduced the signals originating from surrounding areas by as much as 53%. This method differs from traditional PSF inversion deconvolution methods in that the PSF was adjusted with lower weighting factors for signals originating from neighboring pixels than those specified by the PSF model. By using this deconvolution method, the lost classification accuracy due to residual impact of PSF effects was reduced to only 1.66% in overall accuracy. The increase in the RMSE of estimated subpixel land cover proportions due to the residual impact of PSF effects was reduced to 0.64%. Spatial aggregation also effectively reduced the errors in estimated land cover proportion images. About 50% of the estimation errors were removed after applying the deconvolution method and aggregating derived proportion images to twice their dimensional pixel size.

Forest Disturbance and North American Carbon Flux

North Americas forests are thought to be a significant sink for atmospheric carbon. Currently, the rate of sequestration by forests on the continent has been estimated at 0.23 petagrams of carbon per year, though the uncertainty about this estimate is nearly 50%. This offsets about 13% of the fossil fuel emissions from the continent [Pacala et al., 2007]. However, the high level of uncertainty in this estimate and the scientific communitys limited ability to predict the future direction of the forest carbon flux reflect a lack of detailed knowledge about the effects of forest disturbance and recovery across the continent. The North American Carbon Program (NACP), an interagency initiative to better understand the distribution, origin, and fate of North American sources and sinks of carbon, has highlighted forest disturbance as a critical factor constraining carbon dynamics [Wofsy and Harris, 2002]. National forest inventory programs in Canada, the United States, and Mexico provide important information, but they lack the needed spatial and temporal detail to support annual estimation of carbon fluxes across the continent. To help with this, the NACP recommends that scientists use detailed remote sensing of the land surface to characterize disturbance.

A stepwise regression tree for nonlinear approximation: applications to estimating subpixel land cover

A stepwise regression tree (SRT) algorithm was developed for approximating complex nonlinear relationships. Based on the regression tree of Breiman et al . (BRT) and a stepwise linear regression (SLR) method, this algorithm represents an improvement over SLR in that it can approximate nonlinear relationships and over BRT in that it gives more realistic predictions. The applicability of this method to estimating subpixel forest was demonstrated using three test data sets, on all of which it gave more accurate predictions than SLR and BRT. SRT also generated more compact trees and performed better than or at least as well as BRT at all 10 equal forest proportion interval ranging from 0 to 100%. This method is appealing to estimating subpixel land cover over large areas.