Survey methods for assessing land cover map accuracy

The increasing availability of digital photographic materials has fueled efforts by agencies and organizations to generate land cover maps for states, regions, and the United States as a whole. Regardless of the information sources and classification methods used, land cover maps are subject to numerous sources of error. In order to understand the quality of the information contained in these maps, it is desirable to generate statistically valid estimates of accuracy rates describing misclassification errors. We explored a full sample survey framework for creating accuracy assessment study designs that balance statistical and operational considerations in relation to study objectives for a regional assessment of GAP land cover maps. We focused not only on appropriate sample designs and estimation approaches, but on aspects of the data collection process, such as gaining cooperation of land owners and using pixel clusters as an observation unit. The approach was tested in a pilot study to assess the accuracy of Iowa GAP land cover maps. A stratified two-stage cluster sampling design addressed sample size requirements for land covers and the need for geographic spread while minimizing operational effort. Recruitment methods used for private land owners yielded high response rates, minimizing a source of nonresponse error. Collecting data for a 9-pixel cluster centered on the sampled pixel was simple to implement, and provided better information on rarer vegetation classes as well as substantial gains in precision relative to observing data at a single-pixel.     

A geostatistical approach for mapping thematic classification accuracy and evaluating the impact of inaccurate spatial data on ecological model predictions

Spatial information in the form of geographical information system coverages and remotely sensed imagery is increasingly used in ecological modeling. Examples include maps of land cover type from which ecologically relevant properties, such as biomass or leaf area index, are derived. Spatial information, however, is not error-free: acquisition and processing errors, as well as the complexity of the physical processes involved, make remotely sensed data imperfect measurements of ecological attributes. It is therefore important to first assess the accuracy of the spatial information being used and then evaluate the impact of such inaccurate information on ecological model predictions. In this paper, the role of geostatistics for mapping thematic classification accuracy through integration of abundant image-derived (soft) and sparse higher accuracy (hard) class labels is presented. Such assessment leads to local indices of map quality, which can be used for guiding additional ground surveys. Stochastic simulation is proposed for generating multiple alternative realizations (maps) of the spatial distribution of the higher accuracy class labels over the study area. All simulated realizations are consistent with the available pieces of information (hard and soft labels) up to their validated level of accuracy. The simulated alternative class label representations can be used for assessing joint spatial accuracy, i.e., classification accuracy regarding entire spatial features read from the thematic map. Such realizations can also serve as input parameters to spatially explicit ecological models; the resulting distribution of ecological responses provides a model of uncertainty regarding the ecological model prediction. A case study illustrates the generation of alternative land cover maps for a Landsat Thematic Mapper (TM) subscene, and the subsequent construction of local map quality indices. Simulated land cover maps are then input into a biogeochemical model for assessing uncertainty regarding net primary production (NPP).     

Accounting for monogastric livestock as a driver in global land use and cover change assessments

Land systems are described based on various characteristics, including land cover composition and agricultural production. However, it is uncertain to what extent livestock, particularly monogastric livestock, determines land systems. We included monogastrics in a land system classification, and statistically analyzed the land cover composition and agricultural production of otherwise similar land systems with and without monogastric livestock. The results indicate that land systems with monogastrics are statistically different from their counterparts in the classification without monogastrics in terms of grassland area and crop yields, but are less different in terms of tree area, crop area, and ruminant livestock production. We then used a land systems map that includes monogastrics in the classification and a similar map that does not include monogastrics to project future changes in a novel manner that integrates livestock as a determinant of land systems. The results show that including monogastrics in otherwise similar projections yields less cropland intensification and more cropland expansion in several world regions, including Northern Africa and the Middle East. Other regions, such as Europe and Australia, were characterized by less decrease or more increase in tree area in the application with monogastrics, mainly due to the occurrence of open forests with monogastrics. This study prompts a call for improved characterization of land systems for land use and cover change (LUCC) assessments in order to better represent LUCC driven by monogastric livestock.     

Modeling and interpreting the accuracy assessment error matrix for a doubly classified map

This paper considers two maps having the same spatial extent and the same mapping categories but where each map is subject to classification error. An overlay of the maps yields a (dis)similarity matrix whose (i, j)-entry is the areal proportion placed into category i by the first map and into category j by the second map. A parametric model, called the latent truth model, is proposed which specifies the dissimilarity matrix in terms of the true (but unknown) proportions for the mapping categories as well as the unknown error rates for the two maps. The number of parameters in the model exceeds the degrees of freedom in the dissimilarity matrix. However, a method of regularization is applied to effectively reduce the dimension of the parameter space and to permit model fitting. From the fitted model, one obtains estimates for the true mapping proportions as well as estimated error matrices for each of the maps. Accuracy assessment characteristics for each map (such as user's accuracy, producer's accuracy, overall accuracy, and the kappa coefficient) can be computed from the estimated error matrices. Methods are illustrated with two landcover maps of Wicomico County, Maryland.     

Multiscale advanced raster map analysis system: Definition, design and development

This paper brings together a multidisciplinary initiative to develop advanced statistical and computational techniques for analyzing, assessing, and extracting information from raster maps. This information will provide a rigorous foundation to address a wide range of applications including disease mapping, emerging infectious diseases, landscape ecological assessment, land cover trends and change detection, watershed assessment, and map accuracy assessment. It will develop an advanced map analysis system that integrates these techniques with an advanced visualization toolbox, and use the system to conduct large case studies using rich sets of raster data, primarily from remotely sensed imagery. As a result, it will be possible to study and evaluate raster maps of societal, ecological, and environmental variables to facilitate quantitative characterization and comparative analysis of geospatial trends, patterns, and phenomena. In addition to environmental and ecological studies, these techniques and tools can be used for policy decisions at national, state, and local levels, crisis management, and protection of infrastructure. Geospatial data form the foundation of an information-based society. Remote sensing has been a vastly under-utilized resource involving a multi-million dollar investment at the national levels. Even when utilized, the credibility has been at stake, largely because of lack of tools that can assess, visualize, and communicate accuracy and reliability in timely manner and at desired confidence levels. Consider an imminent 21st century scenario: What message does a multi-categorical map have about the large landscape it represents? And at what scale, and at what level of detail? Does the spatial pattern of the map reveal any societal, ecological, environmental condition of the landscape? And therefore can it be an indicator of change? How do you automate the assessment of the spatial structure and behavior of change to discover critical areas, hot spots, and their corridors? Is the map accurate? How accurate is it? How do you assess the accuracy of the map? How do we evaluate a temporal change map for change detection? What are the implications of the kind and amount of change and accuracy on what matters, whether climate change, carbon emission, water resources, urban sprawl, biodiversity, indicator species, human health, or early warning? And with what confidence? The proposed research initiative is expected to find answers to these questions and a few more that involve multi-categorical raster maps based on remote sensing and other geospatial data. It includes the development of techniques for map modeling and analysis using Markov Random Fields, geospatial statistics, accuracy assessment and change detection, upper echelons of surfaces, advanced computational techniques for geospatial data mining, and advanced visualization techniques.     

Effect of Land Cover and Ecosystem Mapping on Ecosystem‐Risk Assessment in the Little Karoo,South Africa

Extinction‐risk assessments aim to identify biological diversity features threatened with extinction. Although largely developed at the species level, these assessments have recently been applied at the ecosystem level. In South Africa, national legislation provides for the listing and protection of threatened ecosystems. We assessed how land‐cover mapping and the detail of ecosystem classification affected the results of risk assessments that were based on extent of habitat loss. We tested 3 ecosystem classifications and 4 land‐cover data sets of the Little Karoo region, South Africa. Degraded land (in particular, overgrazed areas) was successfully mapped in just one of the land‐cover data sets. From <3% to 25% of the Little Karoo was classified as threatened, depending on the land‐cover data set and ecosystem classification applied. The full suite of threatened ecosystems on a fine‐scale map was never completely represented within the spatial boundaries of a coarse‐scale map of threatened ecosystems. Our assessments highlight the importance of land‐degradation mapping for the listing of threatened ecosystems. On the basis of our results, we recommend that when budgets are constrained priority be given to generating more‐detailed land‐cover data sets rather than more‐detailed ecosystem classifications for the assessment of threatened ecosystems. El Efecto de la Cobertura Terrestre y el Mapeo de Ecosistemas en la Valoración de Riesgos en los Ecosistemas en Little Karoo, Sudáfrica     

Multinomial mixture model with heterogeneous classification probabilities

Royle and Link (Ecology 86(9):2505?C2512, 2005) proposed an analytical method that allowed estimation of multinomial distribution parameters and classification probabilities from categorical data measured with error. While useful, we demonstrate algebraically and by simulations that this method yields biased multinomial parameter estimates when the probabilities of correct category classifications vary among sampling units. We address this shortcoming by treating these probabilities as logit-normal random variables within a Bayesian framework. We use Markov chain Monte Carlo to compute Bayes estimates from a simulated sample from the posterior distribution. Based on simulations, this elaborated Royle-Link model yields nearly unbiased estimates of multinomial and correct classification probability estimates when classification probabilities are allowed to vary according to the normal distribution on the logit scale or according to the Beta distribution. The method is illustrated using categorical submersed aquatic vegetation data.     

Need for a global map of forest naturalness for a sustainable future

There is a growing need to assess and monitor forest cover and its conservation status over global scales to determine human impact on ecosystems and to develop sustainability plans. Recent approaches to measure regional and global forest status and dynamics are based on remotely sensed estimates of tree cover. We argue that tree cover should not be used to assess the area of forest ecosystems because tree cover is an undefined subset of forest cover. For example, tree cover can indicate a positive trend even in the presence of deforestation, as in the case of plantations. We believe a global map of forest naturalness that accounts for the bio-ecological integrity of forest ecosystems, for example, intact forests, old-growth forest patches, rewilding forests (exploited forest landscapes undergoing long-term natural succession), and managed forests is needed for global forest assessment.     

土壤可蚀性（K）值图编制方法的初步研究

利用土壤普查图件和剖面理化分析的成果资料编制了土壤可蚀性（Ｋ）值图，介绍了求取剖面点Ｋ值的查图表法、图斑分并与图斑Ｋ值的确定原则和编制程序等，应用上述方法，完成了我国第一张地区级Ｋ值图，经分析，与公式算法比，查图表法的Ｋ值精度为８５．４％，与径流小区实测值比，用Ｋ值图上相应Ｋ值监测的土壤年流失量的精度为８６．０％。     

Accounting for geophysical information in geostatistical characterization of unexploded ordnance (UXO) sites

Efficient and reliable unexploded ordnance (UXO) site characterization is needed for decisions regarding future land use. There are several types of data available at UXO sites and geophysical signal maps are one of the most valuable sources of information. Incorporation of such information into site characterization requires a flexible and reliable methodology. Geostatistics allows one to account for exhaustive secondary information (i.e.,, known at every location within the field) in many different ways. Kriging and logistic regression were combined to map the probability of occurrence of at least one geophysical anomaly of interest, such as UXO, from a limited number of indicator data. Logistic regression is used to derive the trend from a geophysical signal map, and kriged residuals are added to the trend to estimate the probabilities of the presence of UXO at unsampled locations (simple kriging with varying local means or SKlm). Each location is identified for further remedial action if the estimated probability is greater than a given threshold. The technique is illustrated using a hypothetical UXO site generated by a UXO simulator, and a corresponding geophysical signal map. Indicator data are collected along two transects located within the site. Classification performances are then assessed by computing proportions of correct classification, false positive, false negative, and Kappa statistics. Two common approaches, one of which does not take any secondary information into account (ordinary indicator kriging) and a variant of common cokriging (collocated cokriging), were used for comparison purposes. Results indicate that accounting for exhaustive secondary information improves the overall characterization of UXO sites if an appropriate methodology, SKlm in this case, is used.     

Composite likelihood approach to the regression analysis of spatial multivariate ordinal data and spatial compositional data with exact zero values

In many environmental and ecological studies, it is of interest to model compositional data. One approach is to consider positive random vectors that are subject to a unit-sum constraint. In landscape ecological studies, it is common that compositional data are also sampled in space with some elements of the composition absent at certain sampling sites. In this paper, we first propose a practical spatial multivariate ordered probit model for multivariate ordinal data, where the response variables can be viewed as the discretized non-negative compositions without the unit-sum constraint. We then propose a novel two-stage spatial mixture Dirichlet regression model. The first stage models the spatial dependence and the presence of exact zero values, and the second stage models all the non-zero compositional data. A maximum composite likelihood approach is developed for parameter estimation and inference in both the spatial multivariate ordered probit model and the two-stage spatial mixture Dirichlet regression model. The standard errors of the parameter estimates are computed by an estimate of the Godambe information matrix. A simulation study is conducted to evaluate the performance of the proposed models and methods. A land cover data example in landscape ecology further illustrates that accounting for spatial dependence can improve the accuracy in the prediction of presence/absence of different land covers as well as the magnitude of land cover compositions.     

Application of analytical hierarchy process and geographic information systems in land-use suitability evaluation: a case study of Dümrek village (Çanakkale,Turkey)

This study used both analytical hierarchy process (AHP) and geographic information systems (GIS) to make a land-use suitability analysis for the village of Dümrek, NW Turkey. Primarily, alternative land uses for agriculture, meadow–pasture and forest were determined along with criteria for these alternatives and a hierarchical structure was produced and used to determine the weights of the criteria. Spatial data were identified by means of GIS and calculations were made using the suitability values specified and weights obtained from AHP. Suitability maps were then produced for the above land use alternatives. Subsequently, a synthesized suitability map was formed from these maps. According to the weights specified by AHP, the order of land use preferences among the alternatives for rural development of Dümrek was agriculture, forest and meadow. The synthesized suitability map showed that the areas allocated for forest and agriculture were close to the present ratios of use; however, meadow land, which does not exist at present, should be allocated as a land use to constitute 12.5% of the study area. Achieving sustainability in land use is possible by planners and administrators considering results obtained from land suitability mapping studies at the stage of allocating land uses. The method applied in this research is considered useful when taking policy decisions covering the evaluation of rural land use, particularly for subunits of the state administration.     

Refining Biodiversity Conservation Priorities

Abstract:  Although there is widespread agreement about conservation priorities at large scales (i.e., biodiversity hotspots), their boundaries remain too coarse for setting practical conservation goals. Refining hotspot conservation means identifying specific locations (individual habitat patches) of realistic size and scale for managers to protect and politicians to support. Because hotspots have lost most of their original habitat, species endemic to them rely on what remains. The issue now becomes identifying where this habitat is and these species are. We accomplished this by using straightforward remote sensing and GIS techniques, identifying specific locations in Brazil's Atlantic Forest hotspot important for bird conservation. Our method requires a regional map of current forest cover, so we explored six popular products for mapping and quantifying forest: MODIS continuous fields and a MODIS land cover (preclassified products), AVHRR, SPOT VGT, MODIS (satellite images), and a GeoCover Landsat thematic mapper mosaic (jpg). We compared subsets of these forest covers against a forest map based on a Landsat enhanced thematic mapper. The SPOT VGT forest cover predicted forest area and location well, so we combined it with elevation data to refine coarse distribution maps for forest endemic birds. Stacking these species distribution maps enabled identification of the subregion richest in threatened birds—the lowland forests of Rio de Janeiro State. We highlighted eight priority fragments, focusing on one with finer resolved imagery for detailed study. This method allows prioritization of areas for conservation from a region >1 million km² to forest fragments of tens of square kilometers. To set priorities for biodiversity conservation, coarse biological information is sufficient. Hence, our method is attractive for tropical and biologically rich locations, where species location information is sparse.     

Spatially-balanced sampling versus unbalanced stratified sampling for assessing forest change: evidences in favour of spatial balance

Large-scale remote sensing-based inventories of forest cover are usually carried out by combining unsupervised classifications of satellite pixels into forest/non forest classes (map data) with subsequent time-consuming visual on-screen imagery classification of a probabilistic sample of pixels taken as the ground truth (reference data). In this paper the estimation of forest change from a sample of reference data is approached by: (i) exploiting map data to construct strata in which changes are occurred, and then adopting the stratified sampling joined with the HT estimator with most sampling effort devoted to strata where changes are occurred irrespective of their size, as suggested in most remote sensing literature regarding land change assessments; (ii) adopting a spatial scheme ensuring spatially balanced samples, as suggested in most recent statistical literature regarding spatial surveys, and exploiting the map data in the difference estimator. The results of a comparison performed on an artificial population of reference data generated from a real population of map data recorded in Sardinia (Italy) discourage the use of unbalanced stratified samples that achieve the worst precision. The best results are obtained by means of spatially balanced samples or stratification with nearly proportional allocation to strata.     

Are existing vegetation maps adequate to predict bird distributions?

Bird species are selective on the vegetation types in which they are found but predictive models of bird distribution based on variables derived from land-use/land-cover maps tend to have limited success. It has been suggested that accuracy of existing maps used to derive predictors is in part responsible for the limited success of bird distribution models. In two areas of 4900 km² of Western Andalusia, Spain, we compared the predictive ability of bird distribution models derived from two existing general-purpose land-use/land-cover maps, which differ in their resolution and accuracy: a coarse scale vegetation map of Europe, the CORINE land-cover map, and a detailed regional map, the 1995 land-use/land-cover map of Andalusia from the SINAMBA (Consejerı́a de Medio Ambiente, Junta de Andalucı́a). We compared the bird distribution models derived from these general-purpose vegetation maps with models derived from two more accurate structural vegetation maps built considering directly variables that influence bird habitat selection, one built from satellite images for this study and another obtained by improving the resolution and accuracy of the SINAMBA map with satellite data. We sampled the presence/absence of bird species at 857 points using 15-min point surveys. Predictive models for 54 bird species were built with generalised additive models (GAMs), using as potential predictors the same set of landscape and vegetation structure variables measured on each map. We compared for each bird species the predictive accuracy of the best model derived from each map. Vegetation structure measured at bird sample points was used as ground-truth for comparing the accuracy of vegetation maps. Although maps differed in their resolution and accuracy, the results show that all of them produced similarly accurate bird distribution models, with a mixed map produced with both thematic and satellite information being the best. The models derived from the more accurate vegetation structure maps obtained from satellite data were not more accurate than those derived directly from the SINAMBA or CORINE maps. Our results suggest that some general-purpose land-use/land-cover maps are accurate enough to derive bird distribution models. There is a certain limit to improve vegetation maps above which there is no effect in their power to predict bird distribution.     

Assessing the risk of soil contamination in the Swiss Jura using indicator geostatistics

The probability that the concentrations of toxic substances in soil or other medium exceed tolerablemaxima at any unsampled place can be estimated by indicator geostatistics. The method is developed and used to estimate and map the risk of contamination by cadmium, copper and lead in the topsoil of a 14.5 km 2 region in the Swiss Jura. It combines both direct measurements of metal concentrations and thecalibration of a geological map, and it shows that the risk of toxicity is least on Argovian rocks. Two approaches are proposed to divide a region into safe' and 'hazardous' zones on the basis of probability maps. The first declares as contaminated all places where the risk of contamination exceeds a given threshold. The second approach first evaluates the financial costs that might result from a wrongdeclaration, after which the site is allocated to a class so as to minimize that cost. The risk of exposure for humans and animals is generally greater for contaminated agricultural land than for forest soil, and so land use is taken into account in both procedures.     

Using earth observation data from multiple sources to map rare habitats in a coastal conservation area

Kenfig NNR (National Nature Reserve) is a coastal sand dune system in south Wales, UK. The site is an important location for the conservation of the fen orchidLiparis loeselii, a significant proportion of the UK population is found solely on the site. Approaches to the mapping and monitoring of the habitats at Kenfig NNR using EO (Earth Observation) methods are investigated. Typical airborne EO missions over such sites produce more than a single source of EO data; these may include various optical imaging sensors with different spectral ranges, film cameras and ranging devices to measure topography. Conservation managers are thus presented with the problem of which sources of data to use when producing a land cover map of the site of interest. Using a data set gathered over the Kenfig NNR site, we investigate land cover mapping methods for conservation. The land cover types of interest typically cover small areas within a much larger site so they present a hard problem for the EO data and associated classification methods to solve. Land cover classifications produced from the data sets provide a set of competing hypotheses of land cover type for the site. Methods we use to resolve this competition between the data sets include voting methods, data fusion methods and a method utilising fuzzy logic to aggregate information. This paper is intended to act as an introduction to some of the issues involved in using EO data for habitat mapping in highly heterogeneous coastal dune environments and to present some preliminary results of the performance of each method.     

Prediction of potential areas of species distributions based on presence-only data

We introduce a methodology to infer zones of high potential for the habitat of a species, useful for management of biodiversity, conservation, biogeography, ecology, or sustainable use. Inference is based on a set of sites where the presence of the species has been reported. Each site is associated with covariate values, measured on discrete scales. We compute the predictive probability that the species is present at each node of a regular grid. Possible spatial bias for sites of presence is accounted for. Since the resulting posterior distribution does not have a closed form, a Markov chain Monte Carlo (MCMC) algorithm is implemented. However, we also describe an approximation to the posterior distribution, which avoids MCMC. Relevant features of the approach are that specific notions of data acquisition such as sampling intensity and detectability are accounted for, and that available a priori information regarding areas of distribution of the species is incorporated in a clear-cut way. These concepts, arising in the presence-only context, are not addressed in alternative methods. We also consider an uncertainty map, which measures the variability for the predictive probability at each node on the grid. A simulation study is carried out to test and compare our approach with other standard methods. Two case studies are also presented.     

Satellite detection of bird communities in tropical countryside.

The future of biodiversity hinges partly on realizing the potentially high conservation value of human-dominated countryside. The characteristics of the countryside that promote biodiversity preservation remain poorly understood, however, particularly at the fine scales at which individual farmers tend to make land use decisions. To address this problem, we explored the use of a rapid remote sensing method for estimating bird community composition in tropical countryside, using a two-step process. First, we asked how fine-grained variation in land cover affected community composition. Second, we determined whether the observed changes in community composition correlated with three easily accessible remote sensing metrics (wetness, greenness, and brightness), derived from performing a tasseled-cap transformation on a Landsat Enhanced Thematic Mapper Plus image. As a comparison, we also examined whether the most commonly used remote sensing indicator in ecology, the Normalized Difference Vegetation Index (NDVI), correlated with community composition. We worked within an agricultural landscape in southern Costa Rica, where the land comprised a complex and highly heterogeneous mosaic of remnant native vegetation, pasture, coffee cultivation, and other crops. In this region, we selected 12 study sites (each < 60 ha) that encompassed the range of available land cover possibilities in the countryside. Within each site, we surveyed bird communities within all major land cover types, and we conducted detailed field mapping of land cover. We found that the number of forest-affiliated species increased with forest cover and decreased with residential area across sites. Conversely, the number of agriculture-affiliated species using forest increased with land area devoted to agricultural and residential uses. Interestingly, we found that the wetness and brightness metrics predicted the number of forest- and agriculture-affiliated species within a site as well as did detailed field-generated maps of land cover. In contrast, NDVI and the closely correlated greenness metric did not correlate with land cover or with bird communities. Our study shows the strong potential of the tasseled-cap transformation as a tool for assessing the conservation value of countryside for biodiversity.