Resolution Land Cover Map Research Articles

Next generation of global land cover characterization, mapping, and monitoring

Land cover change is increasingly affecting the biophysics, biogeochemistry, and biogeography of the Earth's surface and the atmosphere, with far-reaching consequences to human well-being. However, our scientific understanding of the distribution and dynamics of land cover and land cover change (LCLCC) is limited. Previous global land cover assessments performed using coarse spatial resolution (300m–1km) satellite data did not provide enough thematic detail or change information for global change studies and for resource management. High resolution (∼30m) land cover characterization and monitoring is needed that permits detection of land change at the scale of most human activity and offers the increased flexibility of environmental model parameterization needed for global change studies. However, there are a number of challenges to overcome before producing such data sets including unavailability of consistent global coverage of satellite data, sheer volume of data, unavailability of timely and accurate training and validation data, difficulties in preparing image mosaics, and high performance computing requirements. Integration of remote sensing and information technology is needed for process automation and high-performance computing needs. Recent developments in these areas have created an opportunity for operational high resolution land cover mapping, and monitoring of the world. Here, we report and discuss these advancements and opportunities in producing the next generations of global land cover characterization, mapping, and monitoring at 30-m spatial resolution primarily in the context of United States, Group on Earth Observations Global 30m land cover initiative (UGLC).

2010 land cover map of insular Southeast Asia in 250-m spatial resolution

In this letter, we present the methodology and accuracy assessment of a new regional 250-m spatial resolution land cover map of insular Southeast Asia. Nearly 500 daily Moderate Resolution Imaging Spectroradiometer images (acquired 2 January–3 July 2010) were used in the production of the map. Additionally, peatland maps, elevation information and Daichi-Advanced Land Observing Satellite mosaic data were utilized in the mapping process. The map covers the Malaysian Peninsula and the major islands of Sumatra, Java, Borneo, Sulawesi and Mindanao, in addition to the western part of New Guinea and numerous smaller islands. The classification scheme of 13 classes reflects the special characteristics of land cover of insular Southeast Asia, including, for example, peatswamp forests and large-scale palm plantations. The map was assessed to have 85% overall accuracy using high-resolution (10–30 m) satellite data as reference material. It can therefore be considered to provide reliable information on the current status of land cover in this region experiencing rapid land cover changes.

Possibilities and limitations of artificial neural networks for subpixel mapping of land cover

Although developments in remote sensing have greatly improved land cover mapping, the mixed pixel problem has not yet been fully addressed. Soft classification techniques have been introduced to address the problem, but they do not show the spatial location of the class proportions in a pixel. Subpixel mapping has been introduced to address the drawbacks of soft classifications. In this work, the feedforward backpropagating neural network (FFBPNN) was used for subpixel mapping. A set of class proportion images, which are to be treated as soft classification results, were created from a high spatial resolution (25 m) land cover dataset. For this purpose, the land cover dataset was aggregated both thematically (into two, four or eight land cover classes) and spatially (into proportion images with pixel sizes of 75, 150 and 300 m). This resulted in nine different combinations that were considered here as study cases. Several FFBPNNs were trained using these proportion images and the original land cover dataset (which was used as a target). Subsequently, the best networks were used to reconstruct high spatial resolution land cover maps of two heterogeneous areas in the south of The Netherlands. The overall accuracies obtained revealed that the networks were influenced by the spatial frequency, shape and size of the different land cover types. Moreover, it was revealed that most of the errors were on the class boundaries where highly mixed pixels are to be expected. The accuracies spanned a wide range of values depending on the complexity of the cases. Although it was not possible to exhaustively explore all network architectures, the results demonstrate the potential of the FFBPNN for subpixel mapping.

Dis-aggregation of airborne flux measurements using footprint analysis

Aircraft measurements of turbulent fluxes are generally being made with the objective to obtain an estimate of regional exchanges between land surface and atmosphere, to investigate the spatial variability of these fluxes, but also to learn something about the fluxes from some or all of the land cover types that make up the landscape. In this study we develop a method addressing this last objective, an approach to disentangle blended fluxes from a landscape into the component fluxes emanating from the various land cover classes making up that landscape. The method relies on using a footprint model to determine which part of the landscape the airborne flux observation refers to, using a high resolution land cover map to determine the fractional covers of the various land cover classes within that footprint, and finally using multiple linear regression on many such flux/fractional cover data records to estimate the component fluxes. The method is developed in the context of three case studies of increasing complexity and the analysis covers three scalar fluxes: sensible and latent heat fluxes and carbon dioxide flux, as well as the momentum flux. A basic assumption under the dis-aggregation method is that the composite flux, i.e. the landscape flux, is a linear average of the component fluxes, i.e. the fluxes from the various land elements. We test and justify this assumption by comparing linear averages of component fluxes in simple ‘binary landscapes’, weighted by their relative area, with directly aircraft observed fluxes. In all case studies dis-aggregation of mixed values for fluxes from heterogeneous areas into component land cover class specific fluxes is feasible using robust least squares regression, both in simple binary ‘landscapes’ and in more complex cases. Both the differences between land cover classes and the differences between synoptic conditions can be resolved, for those land cover classes that make up sufficiently large fractions of the landscape. The regression F-statistic and the closely associated p-values are good indicators for this latter prerequisite and for other sources of uncertainty in the dis-aggregated flux estimates that render it meaningful or not. An analysis of the effect of various sources of errors in input data, footprint estimates and of skewed land cover class distributions is presented. A validation of flux estimates obtained through the dis-aggregation method against independent ground data proved satisfactorily. Recommendations for the use of the method are given as are suggestions for further development.

Assessment of vegetation dynamics and their response to variations in precipitation and temperature in the Tibetan Plateau

The Tibetan Plateau is a region sensitive to climate change, due to its high altitude and large terrain. This sensitivity can be measured through the response of vegetation patterns to climate variability in this region. Time series analysis of Normalized Difference Vegetation Index (NDVI) imagery and correlation analyses are effective tools to study land cover changes and their response to climatic variations. This is especially important for regions like the Tibetan Plateau, which has a complex ecosystem but lacks a lot of detailed in-situ observation data due to its remoteness, vastness and the severity of its climatic conditions. In this research a time series of 315 SPOT VEGETATION scenes, covering the period between 1998 and 2006, has been processed with the Harmonic ANalysis of Time Series (HANTS) algorithm in order to reveal the governing spatiotemporal pattern of variability. Results show that the spatial distribution of NDVI values is in agreement with the general climate pattern in the Tibetan Plateau. The seasonal variation is greatly influenced by the Asian monsoon. Interannual analysis shows that vegetation density (recorded here by the NDVI values) in the entire Tibetan Plateau has generally increased. Using a 1 km resolution land cover map from GLC2000, seven meteorological stations, presenting monthly data on near surface air temperature and precipitation, were selected for correlation analysis between NDVI and climate conditions in this research. A time lag response has also been found between NDVI and climate variables. Except in desert grassland (Shiquanhe station), the NDVI of all selected sites showed strong correlation with air temperature and precipitation, with variations in correlation according to the different land cover types at different locations. The strongest relationship was found in alpine and subalpine plain grass, the weakest in desert grassland.

Forest fragmentation, structure, and age characteristics as a legacy of forest management

The combination of forest inventory and satellite-derived landscape composition and structure provides otherwise unavailable information on regional forest conditions and enables investigation of the cumulative effects of forest management over time. Forest pattern results from a range of both natural and anthropogenic factors. This study characterizes the forest pattern of Vancouver Island, British Columbia (>32,000 km 2) at the watershed scale, using new national datasets of Canadian forest composition and fragmentation, and relating these to forest inventory-derived age structure. Vancouver Island is extensively forested, possessing highly valuable and productive forests, and is managed to meet a range of stakeholder interests. Forest fragmentation metrics were derived from a satellite-derived, 25-m spatial resolution land cover map (i.e., grain) to represent patterns within 1 km analysis units (i.e., extent) for the entire forested area of Canada. We summarized forest fragmentation island-wide and compared trends between forest-dominated (>85% forest) and less-forested (<85% forest) watersheds. We also explored these patterns with partial canonical correspondence analyses to determine the independent and shared relationships of landscape composition, forest fragmentation, and spatial variables with forest age structure. Less-forested watersheds are more fragmented than forest-dominated watersheds, as indicated by more (5.6 versus 2.7 per km 2) and smaller (36 versus 63 ha) forest patches, and a greater edge density (82 versus 55 m/ha). Of the 1283 watersheds examined, 91% were forest-dominated; island-wide trends are thus similar to those of the forest-dominated watersheds. Forest age is related to landscape composition and forest fragmentation, which collectively explain 27% and 53% of the age structure of all and less-forested watersheds, respectively. In both sets of watersheds, young forest stands (1–120 years) are associated with broadleaf forests, agreeing with successional expectations, and patchy forest distributions. For all watersheds, old growth forest stands (>240 years) are associated with dense coniferous forest, but also with early successional communities, smaller patch sizes, and greater edge densities, indicating fragmentation of these forests and reflecting strategies for managing old growth on the landscape. In contrast, watersheds with an abundance of mature forest (121–240 years) are compositionally similar to watersheds with old growth, but are much less fragmented. Our results indicate that although old growth forest stands on Vancouver Island have been retained, they are not typically found within a continuous forest matrix. Medium spatial resolution Earth observation products are now available for the entire forested area of Canada, providing valuable insight into spatial and temporal forest dynamics (i.e., succession, harvesting).

High Spatial Resolution Land Cover Mapping Using Remotely Sensed Image

Land cover classification from remotely sensed data is an important topic in remote sensing applications. This paper present an economical analysis of land covers in Mebok Estuary, Penang, Malaysia. This study investigated the potential of using digital camera for land cover mapping Mebok Estuary, Penang, Malaysia. Airborne digital imagery has proved to be an effective tool for land cover studies. A Kodak camera, model DC290, was used to capture images from an elevation of 2438.4 meter on board Cessna 172Q. The use of digital camera as a sensor to capture digital images is cheaper and economical compared to the use of other airborne sensor. This technique overcomes the problem of the difficulty in obtaining cloud-free scenes in the Equatorial region from a satellite platform. Supervised classification techniques were used in the classification analysis. Supervised classification technique (Parallelepiped with Maximum Likelihood as tie breaker classifier, PML, Maximum Likelihood, ML, Minimum Distance-to-Mean, MDM, and Parallelepiped, P) was applied to the digital camera spectral bands (red, green and blue) to extract the thematic information from the acquired scenes. The best supervised classifier was chosen based on the highest overall accuracy and Kappa statistic. The accuracy of the classified images was validated using a reference data set. The study revealed that the Parallelepiped with Maximum Likelihood as tie breaker classifier produced superior result and achieved a high degree of accuracy. The classified land cover map was geometrically corrected to provide a geocode map. The results produced by this study indicated that land cover features could be clearly identified and classified into a land cover map. This study suggested that the land cover types of Mebok Estuary, Penang, Malaysia can be accurately mapped.

Estimation of biomass distribution in Peninsular Malaysia and in the islands of Sumatra, Java and Borneo based on multi-resolution remote sensing land cover analysis

In this study we used high resolution (20 m) land cover maps to derive detailed information on land cover structure within the classes of a regional medium resolution (500 m) land cover map. This enabled improved biomass estimation for the medium resolution land cover classes. Although our results suggested that land cover maps based merely on medium resolution remote sensing data can be used to monitor the extent of forest cover, they also showed that these maps alone are not sufficient to produce reliable regional estimates on above ground biomass in insular Southeast Asia. A quarter of the study area was covered by sub-pixel size (500 m) mosaic of various land cover types containing 14% of biomass. In total, non-forest areas covered over 60% of the study area and included 43% of biomass. In these areas, highly fragmented within class land cover structure was shown to significantly affect biomass estimates. Therefore we conclude that forest/land cover monitoring based merely on medium resolution remote sensing data can no longer be used to sufficiently quantify carbon fluxes connected to land cover changes in insular Southeast Asia, but multi-resolution approaches are needed to perform this task.

Impact of land cover uncertainties on estimates of biospheric carbon fluxes

Large‐scale bottom‐up estimates of terrestrial carbon fluxes, whether based on models or inventory, are highly dependent on the assumed land cover. Most current land cover and land cover change maps are based on satellite data and are likely to be so for the foreseeable future. However, these maps show large differences, both at the class level and when transformed into Plant Functional Types (PFTs), and these can lead to large differences in terrestrial CO2 fluxes estimated by Dynamic Vegetation Models. In this study the Sheffield Dynamic Global Vegetation Model is used. We compare PFT maps and the resulting fluxes arising from the use of widely available moderate (1 km) resolution satellite‐derived land cover maps (the Global Land Cover 2000 and several MODIS classification schemes), with fluxes calculated using a reference high (25 m) resolution land cover map specific to Great Britain (the Land Cover Map 2000). We demonstrate that uncertainty is introduced into carbon flux calculations by (1) incorrect or uncertain assignment of land cover classes to PFTs; (2) information loss at coarser resolutions; (3) difficulty in discriminating some vegetation types from satellite data. When averaged over Great Britain, modeled CO2 fluxes derived using the different 1 km resolution maps differ from estimates made using the reference map. The ranges of these differences are 254 gC m−2 a−1 in Gross Primary Production (GPP); 133 gC m−2 a−1 in Net Primary Production (NPP); and 43 gC m−2 a−1 in Net Ecosystem Production (NEP). In GPP this accounts for differences of −15.8% to 8.8%. Results for living biomass exhibit a range of 1109 gC m−2. The types of uncertainties due to land cover confusion are likely to be representative of many parts of the world, especially heterogeneous landscapes such as those found in western Europe.

Synergy of SMOS Microwave Radiometer and Optical Sensors to Retrieve Soil Moisture at Global Scale

Methods to retrieve surface soil moisture were assessed at the global scale for one entire year by using simulated Soil Moisture and Ocean Salinity brightness temperatures (T B) and vegetation coverage information which can be derived from optical sensors. The global T B database consists of half-degree continental pixels and accounts for within-pixel heterogeneity, based on 1-km resolution land cover maps. The retrievals were performed by using a three-parameter inversion method applied to the L-band Microwave Emission of the Biosphere model. By using a Bayesian approach, vegetation data were injected as a priori information. Two options were investigated to profit from normalized difference vegetation index products: providing an a priori knowledge either on vegetation optical depth or on the vegetation cover fraction (f cover). The latter option allows for a better description of the surface heterogeneity by considering a bare soil fraction. When an error of 1 K is applied to the T B, both synergistic schemes significantly improved the soil moisture accuracy compared with methods using microwave data only. Using the vegetation a priori information, about 80% of the pixels present soil moisture retrieval accuracy less than 0.04 m3middotm-3 in terms of root-mean-square error, whereas methods based only on the microwave data provide 63% of pixels of the studied area with this accuracy. If the error in T B is larger (2 or 3 K), the soil moisture retrieval accuracy decreases significantly for both methods. The use of optical data to give a priori value of vegetation optical option is then the best for these cases.

Effects of cutline density and land-cover heterogeneity on landscape metrics in western Alberta

Forest cutlines are narrow, linear features created in geophysical surveys. In many areas of Canada, forest cutlines are not consistently detected using relatively coarse spatial resolution land-cover maps, such as those produced by classification of Landsat Thematic Mapper (TM) imagery. However, such features may be important in certain wildlife management applications, including those which require an assessment of landscape structure, or forest fragmentation, at various scales. Higher spatial resolution satellite imagery obtained from sensors on platforms such as Satellite Pour l’Observation de la Terre (SPOT) and the Indian Remote Sensing (IRS) system may be used to map forest cutlines for these applications. In this study, a TM-based land-cover map of western Alberta is analyzed with forest cutlines mapped from a TM-IRS fusion image, and the effect of increasing cutline density is quantified on five commonly used landscape metrics used to characterize landscape structure in grizzly bear habitat assessment. The accuracy of the fusion image interpretation was determined to be 88%. Simulated landscapes were tested first, and the study area was divided into 104 hexagon-shaped sample landscapes of about 6 km diameter each. Across these sample landscapes, cutline density and initial landscape heterogeneity were significant parameters in explaining change in three metrics, namely edge density, mean patch size, and patch context (expressed as mean nearest-neighbour distance). Patch size variability (expressed as the coefficient of variation of mean patch size) and patch dispersion (expressed as the coefficient of variation of mean nearest-neighbour distance) required additional information on cutline positioning. Overall, the density of the introduced cutline network and the pre-cutline metric value reliably predicted and quantified the response of landscape metrics of interest to grizzly bear biologists. This study shows the importance of mapping forest cutlines regarding their role in changing landscape structure quantification and points out the necessity of using additional remotely sensed data when the feature responsible for the landscape transformation is of too small a size to appear reliably in common TM-based classified imagery.

Adding uncertainty to forest inventory plot locations: effects on analyses using geospatial data

The Forest Inventory and Analysis (FIA) program of the USDA Forest Service alters plot locations before releasing data to the public to ensure landowner confidentiality and sample integrity, but using data with altered plot locations in conjunction with other spatially explicit data layers produces analytical results with unknown amounts of error. We calculated the potential error from using altered location data in combination with other data layers that varied in mean map unit size. The incidence of errors associated with the use of altered plot locations exhibited a strong inverse relationship to the mean map unit size of the other data sets used in the analyses. For a 30 m × 30 m resolution land cover map, plot misclassification rates ranged from 32% to 66%, whereas only 1%–10% of plots were misclassified for ecological subsection data (mean polygon size 9067 km2). Housing density data derived from the US Decennial Census (mean polygon size = 5.7 km2) represented an intermediate condition, with 5%–70% of data points misclassified when altered plot locations were used. These analyses demonstrate the impacts of altering FIA plot locations and represent an important step toward making the FIA database more helpful to a broad variety of end users.

Evaluating Swiss pollen productivity estimates using a simulation approach

We simulated pollen assemblages for 11 sites on the Swiss Plateau (Schweizer Mittelland) based on a low resolution land cover map and pollen productivity estimates (PPE) from this region. A comparison between observed and simulated pollen data demonstrated that the majority of the simulated pollen proportions were within a range of 5% of over- or underestimation, and that at eight of the 11 sites the dominant taxon was correctly estimated. The modelled sum of arboreal pollen was correct within ±11% at nine sites. Our results indicate that the PPE established for the Swiss Plateau can be used to simulate pollen assemblages with reasonable accuracy. Moreover, our results justify the use of the POLLSCAPE simulation using the Prentice-Sugita model and its variations of pollen dispersal and deposition in more complex topographic landscapes than those to which they have hitherto been applied.

Satellite detection of land-use change and effects on regional forest aboveground biomass estimates

We used remote-sensing-driven models to detect land-cover change effects on forest aboveground biomass (AGB) density (Mg.ha(-1), dry weight) and total AGB (Tg) in Minnesota, Wisconsin, and Michigan USA, between the years 1992-2001, and conducted an evaluation of the approach. Inputs included remotely-sensed 1992 reflectance data and land-cover map (University of Maryland) from Advanced Very High Resolution Radiometer (AVHRR) and 2001 products from Moderate Resolution Imaging Spectroradiometer (MODIS) at 1-km resolution for the region; and 30-m resolution land-cover maps from the National Land Cover Data (NLCD) for a subarea to conduct nine simulations to address our questions. Sensitivity analysis showed that (1) AVHRR data tended to underestimate AGB density by 11%, on average, compared to that estimated using MODIS data; (2) regional mean AGB density increased slightly from 124 (1992) to 126 Mg ha(-1) (2001) by 1.6%; (3) a substantial decrease in total forest AGB across the region was detected, from 2,507 (1992) to 1,961 Tg (2001), an annual rate of -2.4%; and (4) in the subarea, while NLCD-based estimates suggested a 26% decrease in total AGB from 1992 to 2001, AVHRR/MODIS-based estimates indicated a 36% increase. The major source of uncertainty in change detection of total forest AGB over large areas was due to area differences from using land-cover maps produced by different sources. Scaling up 30-m land-cover map to 1-km resolution caused a mean difference of 8% (in absolute value) in forest area estimates at the county-level ranging from 0 to 17% within a 95% confidence interval.

Canada-wide maps of dominant tree species from remotely sensed and ground data

Information on the spatial distribution of dominant tree species is important as input or validation data in large scale ecological studies, such as carbon budget estimation, biodiversity assessment, and vegetation dynamic modelling. We used a 1-km resolution land cover map of Canada produced from 1998 SPOT VGT satellite data as the base map. Fractions of coniferous needleleaf or deciduous broadleaf forests within each 1-km pixel were then estimated by developing relationships with climate using land-cover maps from Landsat images selected as a representative sample. The fraction of coniferous/deciduous forest within each 1-km pixel was further divided into dominant species fractions using species composition information at the sub-ecoregion level, derived from statistics in the Canadian Forest Inventory (CanFI 1991/94 version), from narrative descriptions of ecoregions, and from topographical data. A total of 17 Canada-wide maps of dominant tree species groups at 1 km-resolution were produced. The estimates of this study agree well with CanFI data at the ecoregion level where available, and show good correspondence with published range maps of Canadian tree species.

Multitemporal land cover mapping for Canada: methodology and products

A mapping methodology is presented for generating a land cover time series from coarse spatial resolution earth observation data. Historically, this has been a difficult task because of inconsistencies that can arise between maps due to inherent noise present in satellite observations. The new methodology reduces the inconsistency by incorporating several information sources unique to the presented approach of updating an existing land cover map backward and forward in time. It consists of change detection and a local evidence classification decision rule that incorporates the local spectral similarity for each class, local land cover proportions, and expected class changes based on the previous class and change direction. The methodology has been implemented to produce land cover maps of Canada for 1985, 1990, 1995, and 2000 from data acquired by the series of National Oceanic and Atmospheric Administration (NOAA) – advanced high-resolution radiometer (AVHRR) sensors. Accuracy assessment based on medium-resolution (30 m) reference data shows that land cover data produced with this new approach have an overall accuracy similar to that of other 1 km resolution land cover maps of Canada, but this product maintains high consistency between years, with a thematic resolution of 12 classes. An analysis of spatial and temporal patterns of land cover disturbances demonstrates the potential application of the multitemporal land cover time series.

Surface soil moisture retrieval from L-band radiometry: a global regression study

Using a global simulation of L-band (1.4 GHz) brightness temperature (T/sub B/) for two years (1987 and 1988), the relationship between L-band brightness temperatures and surface soil moisture was analyzed using simple regression models. The global T/sub B/ dataset describes continental pixels at a half-degree spatial resolution and accounts for within-pixel heterogeneity, based on 1-km resolution land cover maps. Two different statistical methods were investigated. First, a single regression model was obtained using a linear combination of T/sub B/ indexes. This method consisted in retrieving surface soil moisture using the same global regression model for all the pixels. Second, a regression model was calibrated over each pixel using similar linear combinations of the T/sub B/ indexes. In both cases, the influence of the radiometric noise on T/sub B/ was investigated. Applying these two methods, the capability of L-band T/sub B/ observations to monitor surface soil moisture was evaluated at the global scale and during a two-year time period. Global maps of the estimated accuracy of the soil moisture retrievals were produced. These results contribute to better define the potential of the observations from future spaceborne missions such as the Soil Moisture and Ocean Salinity (SMOS) mission.

Global soil moisture retrieval from a synthetic L‐band brightness temperature data set

A technique to retrieve surface soil moisture was assessed at the global scale using a synthetic data set of L‐band (1.4 GHz) brightness temperatures TB for 2 years, 1987 and 1988. The global TB database consists of half‐degree continental pixels and accounts for within‐pixel heterogeneity, on the basis of 1 km resolution land cover maps. The retrievals were performed using a three‐parameter inversion method applied to the L‐band Microwave Emission of Biosphere model (L‐MEB). Three land surface variables were retrieved simultaneously from the multiangular and dual‐polarization TB data: surface soil moisture wg, vegetation optical depth τ, and surface temperature TS. The retrievals were obtained in two TS configurations: TS was either unknown or known with an uncertainty of 2 K. Applying these two assumptions, global maps of the estimated accuracy of the wg retrievals were produced, and the capability of the TB to monitor wg was evaluated. A sensitivity study was carried out in order to analyze the effect of the main parameters that may affect the retrieval accuracy: the fraction cover of open water and forests, frozen soil conditions, and the radiometric noise on TB. These results contribute to the better definition of the potential of the observations from future spaceborne missions such as the Soil Moisture and Ocean Salinity (SMOS) project.

National Park vegetation mapping using multitemporal Landsat 7 data and a decision tree classifier

Decision tree classifiers have received much recent attention, particularly with regards to land cover classifications at continental to global scales. Despite their many benefits and general flexibility, the use of decision trees with high spatial resolution data has not yet been fully explored. In support of the National Park Service (NPS) Vegetation Mapping Program (VMP), we have examined the feasibility of using a commercially available decision tree classifier with multitemporal satellite data from the Enhanced Thematic Mapper-Plus (ETM+) instrument to map 11 land cover types at the Delaware Water Gap National Recreation Area near Milford, PA. Ensemble techniques such as boosting and consensus filtering of the training data were used to improve both the quality of the input training data as well as the final products. Using land cover classes as specified by the National Vegetation Classification Standard at the Formation level, the final land cover map has an overall accuracy of 82% ( κ=0.80) when tested against a validation data set acquired on the ground ( n=195). This same accuracy is 99.5% when considering only forest vs. nonforest classes. Usage of ETM+ scenes acquired at multiple dates improves the accuracy over the use of a single date, particularly for the different forest types. These results demonstrate the potential applicability and usability of such an approach to the entire National Park system, and to high spatial resolution land cover and forest mapping applications in general.

A new SPOT4-VEGETATION derived land cover map of Northern Eurasia

The European Commission's Joint Research Centre and the Russian Academy of Science's Centre for Forest Ecology and Productivity have produced a new 1 km spatial resolution land cover map of Eurasia from 1999 SPOT4-VEGETATION data. The legend is designed to serve users from science programmes, policy makers, environmental convention secretariats, non-governmental organizations, development-aid projects and the national forest service. The 1999 map is also being updated as part of an international exercise to map Global Land Cover for the year 2000. This Letter describes the map legend, the image classification method, the map accuracy assessment process and presents the land cover map.