Land cover classification with coarse spatial resolution data to derive continuous and discrete maps for complex regions

Regularly updated land cover information at continental or national scales is a requirement for various land management applications as well as biogeochemical and climate modeling exercises. However, monitoring or updating of map products with sufficient spatial detail is currently not widely practiced due to inadequate time-series coverage for most regions of the Earth. Classifications of coarser spatial resolution data can be automatically generated on an annual or finer time scale. However, discrete land cover classifications of such data cannot sufficiently quantify land surface heterogeneity or change. This study presents a methodology for continuous and discrete land cover mapping using moderate spatial resolution time series data sets. The method automatically selects sample data from higher spatial resolution maps and generates multiple decision trees. The leaves of decision trees are interpreted considering the sample distribution of all classes yielding class membership maps, which can be used as estimates for the diversity of classes in a coarse resolution cell. Results are demonstrated for the heterogeneous, small-patch landscape of Germany and the bio-climatically varying landscape of South Africa. Results have overall classification accuracies of 80%. A sensitivity analysis of individual modules of the classification process indicates the importance of appropriately chosen features, sample data balanced among classes, and an appropriate method to combine individual classifications. The comparison of classification results over several years not only indicates the method's consistency, but also its potential to detect land cover changes.     

Land cover mapping in support of LAI and FPAR retrievals from EOS-MODIS and MISR: Classification methods and sensitivities to errors

Land cover maps are used widely to parameterize the biophysical properties of plant canopies in models that describe terrestrial biogeochemical processes. In this paper, we describe the use of supervised classification algorithms to generate land cover maps that characterize the vegetation types required for Leaf Area Index (LAI) and Fraction of Photosynthetically Active Radiation (FPAR) retrievals from MODIS and MISR. As part of this analysis, we examine the sensitivity of remote sensing-based retrievals of LAI and FPAR to land cover information used to parameterize vegetation canopy radiative transfer models. Specifically, a decision tree classification algorithm is used to generate a land cover map of North America from Advanced Very High Resolution Radiometer (AVHRR) data with 1 km spatial resolution using a six-biome classification scheme. To do this, a time series of normalized difference vegetation index data from the AVHRR is used in association with extensive site-based training data compiled using Landsat Thematic Mapper (TM) and ancillary map sources. Accuracy assessment of the map produced via decision tree classification yields a cross-validated map accuracy of 73%. Results comparing LAI and FPAR retrievals using maps from different sources show that disagreement in land cover labels generally do not translate into strong disagreement in LAI and FPAR maps. Further, the main source of disagreement in LAI and FPAR maps can be attributed to specific biome classes that are characterized by a continuum of fractional cover and canopy structure.     

Exploring point-of-interest data from social media for artificial surface validation with decision trees

Artificial surfaces represent one of the key land cover types, and validation is an indispensable component of land cover mapping that ensures data quality. Traditionally, validation has been carried out by confronting the produced land cover map with reference data, which is collected through field surveys or image interpretation. However, this approach has limitations, including high costs in terms of money and time. Recently, geo-tagged photos from social media have been used as reference data. This procedure has lower costs, but the process of interpreting geo-tagged photos is still time-consuming. In fact, social media point of interest (POI) data, including geo-tagged photos, may contain useful textual information for land cover validation. However, this kind of special textual data has seldom been analysed or used to support land cover validation. This paper examines the potential of textual information from social media POIs as a new reference source to assist in artificial surface validation without photo recognition and proposes a validation framework using modified decision trees. First, POI datasets are classified semantically to divide POIs into the standard taxonomy of land cover maps. Then, a decision tree model is built and trained to classify POIs automatically. To eliminate the effects of spatial heterogeneity on POI classification, the shortest distances between each POI and both roads and villages serve as two factors in the modified decision tree model. Finally, a data transformation based on a majority vote algorithm is then performed to convert the classified points into raster form for the purposes of applying confusion matrix methods to the land cover map. Using Beijing as a study area, social media POIs from Sina Weibo were collected to validate artificial surfaces in GlobeLand30 in 2010. A classification accuracy of 80.68% was achieved through our modified decision tree method. Compared with a classification method without spatial heterogeneity, the accuracy is 10% greater. This result indicates that our modified decision tree method displays considerable skill in classifying POIs with high spatial heterogeneity. In addition, a high validation accuracy of 92.76% was achieved, which is relatively close to the official result of 86.7%. These preliminary results indicate that social media POI datasets are valuable ancillary data for land cover validation, and our proposed validation framework provides opportunities for land cover validation with low costs in terms of money and time.     

Rapid Land Cover Classification Using Landsat Time Series based on the Google Earth Engine

Accurate maps of land cover at high spatial resolution are fundamental to many researchs on carbon cycle, climate change monitoring and soil degradation. Google Earth Engine is a cloud-based platform that makes it easy to access high-performance computing resources for processing very large geospatial datasets. It offer opportunities for generating land cover maps designed to meet the increasingly detailed information needs for science,monitoring, and reporting.In this study, we classified the land cover types in Shanxi using Landsat time series data based on the Google Earth Engine Platform. We selected 1 580 sample points be visual interpretation of the original fine spatial resolution images along with Google Earth historical images over six different cover types. We defined training data by randomly sampling 60% of the sample points. The remaining 40% was used for validation. We generated two diffirent types of Landsat composite: (1) one based on median values which is used as the input image for single-date classification; (2)one based on percentile values which is used as input images for time series classification. Random forest classification was performed with two different types of Landsat composites. Random forest classification was performed with two different types of Landsat composites.We visually compared the single-date based to the time series based cover maps of 1990, 2000, 2010 and 2017 in five local areas, and we future compared the results of time series to other products. We aslo performed an accuracy assessment on the land cover classification products. The results shown: (1) The results of time series classification had an overall accuracy of 84%~94%. The time series results improved overall accuracy by 5%~10% compared to single-date results; (2) The result of time series achieves the classification accuracy of products such as CNLUCC, GlobeLand30 and FROM-GLC.The following conclusions were drawn: (1) Cloud computing and archived Landsat data in the GEE has many advantages for land cover classification at a large geographic scale, such as s strong timeliness, short time cycle and low cost; (2) The statistics metrics from Landsat time series is a viable means for discrimination of land cover types, which is particularly useful for the time series classification.     

影像的土地覆被快速分类

精确的土地覆盖信息是进行碳循环、气候变化监测、土壤退化等相关科学研究的基础。随着云计算技术的不断成熟,一些高效算法与平台被不断提出,用来充分挖掘遥感数据所包含的海量信息。基于Google Earth Engine(GEE)云平台,利用随机森林监督分类法对1990、2000、2010、2017年的山西省土地覆被进行了分类。参考Google Earth高清影像选择的1580个样本点,对分类结果进行了验证;同时将分类结果与CNLUCC、GlobeLand30、FROM-GLC等现有土地覆被分类产品进行比较。验证和对比发现时间序列分类结果的总体精度达到86%~94%,比同期单时相分类总体精度提高了5%~10%;本文时间序列结果达到了CNLUCC、GlobeLand30、FROM-GLC等产品的分类精度。结果表明:①在快速准确土地覆被分类方面,时间序列影像与云平台结合,显示出时效性强、时间周期短、成本低等优势;②时间序列百分位数指标能有效地区分不同土地覆被类型的物候差别,在进行土地覆被分类方面显示出简单、易用、高效等特点。该方法对于深入研究大区域尺度的土地覆被变化过程具有重要的参考价值。     

Land surface phenology from MODIS: Characterization of the Collection 5 global land cover dynamics product

Information related to land surface phenology is important for a variety of applications. For example, phenology is widely used as a diagnostic of ecosystem response to global change. In addition, phenology influences seasonal scale fluxes of water, energy, and carbon between the land surface and atmosphere. Increasingly, the importance of phenology for studies of habitat and biodiversity is also being recognized. While many data sets related to plant phenology have been collected at specific sites or in networks focused on individual plants or plant species, remote sensing provides the only way to observe and monitor phenology over large scales and at regular intervals. The MODIS Global Land Cover Dynamics Product was developed to support investigations that require regional to global scale information related to spatio-temporal dynamics in land surface phenology. Here we describe the Collection 5 version of this product, which represents a substantial refinement relative to the Collection 4 product. This new version provides information related to land surface phenology at higher spatial resolution than Collection 4 (500-m vs. 1-km), and is based on 8-day instead of 16-day input data. The paper presents a brief overview of the algorithm, followed by an assessment of the product. To this end, we present (1) a comparison of results from Collection 5 versus Collection 4 for selected MODIS tiles that span a range of climate and ecological conditions, (2) a characterization of interannual variation in Collections 4 and 5 data for North America from 2001 to 2006, and (3) a comparison of Collection 5 results against ground observations for two forest sites in the northeastern United States. Results show that the Collection 5 product is qualitatively similar to Collection 4. However, Collection 5 has fewer missing values outside of regions with persistent cloud cover and atmospheric aerosols. Interannual variability in Collection 5 is consistent with expected ranges of variance suggesting that the algorithm is reliable and robust, except in the tropics where some systematic differences are observed. Finally, comparisons with ground data suggest that the algorithm is performing well, but that end of season metrics associated with vegetation senescence and dormancy have higher uncertainties than start of season metrics.     

Validation and comparison of 1 km global land cover products in China

Four 1 km global land cover products are currently available to the scientific community: the University of Maryland (UMD) global land cover product, the International Geosphere–Biosphere Programme Data and Information System Cover (IGBP‐DISCover), the MODerate resolution Imaging Spectrometer (MODIS) global land cover product and Global Land Cover 2000 (GLC2000). Because of differences in data sources, temporal scales, classification systems and methodologies, it is important to compare and validate these global maps before using them for a variety of studies at regional to global scales. This study aimed to perform the validation and comparison of the four global land cover datasets, and to examine the suitability and accuracy of different coarse spatial resolution datasets in mapping and monitoring cropland across China. To meet this objective, we compared the four global land cover products with the National Land Cover Dataset 2000 (NLCD‐2000) at three scales to evaluate the accuracy of estimates of aggregated cropland areas in China. This was followed by a spatial comparison to assess the accuracies of the four products in estimating the spatial distribution of cropland across China. A comparative analysis showed that there are varying levels of apparent discrepancies in estimating the cropland of China between these four global land cover datasets, and that both area totals and spatial (dis)agreement between them vary from region to region. Among these, the MODIS dataset has the best fit in depicting China's croplands. The coarse spatial resolution and the per pixel classification approach, as well as landscape heterogeneity, are the main reasons for the large discrepancies between the global land cover datasets tested and the reference data.     

Comparison and assessment of coarse resolution land cover maps for Northern Eurasia

Information on land cover at global and continental scales is critical for addressing a range of ecological, socioeconomic and policy questions. Global land cover maps have evolved rapidly in the last decade, but efforts to evaluate map uncertainties have been limited, especially in remote areas like Northern Eurasia. Northern Eurasia comprises a particularly diverse region covering a wide range of climate zones and ecosystems: from arctic deserts, tundra, boreal forest, and wetlands, to semi-arid steppes and the deserts of Central Asia. In this study, we assessed four of the most recent global land cover datasets: GLC-2000, GLOBCOVER, and the MODIS Collection 4 and Collection 5 Land Cover Product using cross-comparison analyses and Landsat-based reference maps distributed throughout the region. A consistent comparison of these maps was challenging because of disparities in class definitions, thematic detail, and spatial resolution. We found that the choice of sampling unit significantly influenced accuracy estimates, which indicates that comparisons of reported global map accuracies might be misleading. To minimize classification ambiguities, we devised a generalized legend based on dominant life form types (LFT) (tree, shrub, and herbaceous vegetation, barren land and water). LFT served as a necessary common denominator in the analyzed map legends, but significantly decreased the thematic detail. We found significant differences in the spatial representation of LFT's between global maps with high spatial agreement (above 0.8) concentrated in the forest belt of Northern Eurasia and low agreement (below 0.5) concentrated in the northern taiga-tundra zone, and the southern dry lands. Total pixel-level agreement between global maps and six test sites was moderate to fair (overall agreement: 0.67-0.74, Kappa: 0.41-0.52) and increased by 0.09-0.45 when only homogenous land cover types were analyzed. Low map accuracies at our tundra test site confirmed regional disagreements and difficulties of current global maps in accurately mapping shrub and herbaceous vegetation types at the biome borders of Northern Eurasia. In comparison, tree dominated vegetation classes in the forest belt of the region were accurately mapped, but were slightly overestimated (10%-20%), in all maps. Low agreement of global maps in the northern and southern vegetation transition zones of Northern Eurasia is likely to have important implications for global change research, as those areas are vulnerable to both climate and socio-economic changes.     

Accuracy assessment using sub-pixel fractional error matrices of global land cover products derived from satellite data

Information on land cover distribution at regional and global scales has become fundamental for studying global changes affecting ecological and climatic systems. The remote sensing community has responded to this increased interest by improving data quality and methodologies for extracting land cover information. However, in addition to the advantages provided by satellite products, certain limitations exist that need to be objectively quantified and clearly communicated to users so that they can make informed decisions on whether and how land cover products should be used. Accuracy assessment is the procedure used to quantify product quality. Some aspects of accuracy assessment for evaluating four global land cover maps over Canada are discussed in this paper. Attempts are made to quantify limiting factors resulting from the coarse spatial resolution of data used for generating land cover information at regional and global levels. Sub-pixel fractional error matrices are introduced as a more appropriate way for assessing the accuracy of mixed pixels. For classification with coarse spatial resolution data, limitations of the classification method produce a maximum achievable accuracy defined as the average percent fraction of dominant land cover of all pixels in the mapped area. Relationships among spatial resolution, landscape heterogeneity and thematic resolution were studied and reported. Other factors that can affect accuracy, such as misregistration and legend conversion, are also discussed.     

Comparisons of land cover and LAI estimates derived from ETM+ and MODIS for four sites in North America: a quality assessment of 2000/2001 provisional MODIS products

The MODIS land science team produces a number of standard products, including land cover and leaf area index (LAI). Critical to the success of MODIS and other sensor products is an independent evaluation of product quality. In that context, we describe a study using field data and Landsat ETM+ to map land cover and LAI at four 49-km² sites in North America containing agricultural cropland (AGRO), prairie grassland (KONZ), boreal needleleaf forest, and temperate mixed forest. The purpose was to: (1) develop accurate maps of land cover, based on the MODIS IGBP (International Geosphere-Biosphere Programme) land cover classification scheme; (2) derive continuous surfaces of LAI that capture the mean and variability of the LAI field measurements; and (3) conduct initial MODIS validation exercises to assess the quality of early (i.e., provisional) MODIS products. ETM+ land cover maps varied in overall accuracy from 81% to 95%. The boreal forest was the most spatially complex, had the greatest number of classes, and the lowest accuracy. The intensive agricultural cropland had the simplest spatial structure, the least number of classes, and the highest overall accuracy. At each site, mapped LAI patterns generally followed patterns of land cover across the site. Predicted versus observed LAI indicated a high degree of correspondence between field-based measures and ETM+ predictions of LAI. Direct comparisons of ETM+ land cover maps with Collection 3 MODIS cover maps revealed several important distinctions and similarities. One obvious difference was associated with image/map resolution. ETM+ captured much of the spatial complexity of land cover at the sites. In contrast, the relatively coarse resolution of MODIS did not allow for that level of spatial detail. Over the extent of all sites, the greatest difference was an overprediction by MODIS of evergreen needleleaf forest cover at the boreal forest site, which consisted largely of open shrubland, woody savanna, and savanna. At the agricultural, temperate mixed forest, and prairie grassland sites, ETM+ and MODIS cover estimates were similar. Collection 3 MODIS-based LAI estimates were considerably higher (up to 4 m² m⁻²) than those based on ETM+ LAI at each site. There are numerous probable reasons for this, the most important being the algorithms' sensitivity to MODIS reflectance calibration, its use of a prelaunch AVHRR-based land cover map, and its apparent reliance on mainly red and near-IR reflectance. Samples of Collection 4 LAI products were examined and found to consist of significantly improved LAI predictions for KONZ, and to some extent for AGRO, but not for the other two sites. In this study, we demonstrate that MODIS reflectance data are highly correlated with LAI across three study sites, with relationships increasing in strength from 500 to 1000 m spatial resolution, when shortwave-infrared bands are included.     

Land cover characterization and change detection for environmental monitoring of pan-Europe

Environmental studies need up-to-date and reliable information on land use and land cover. Such databases, which can be characterized by a high spatial accuracy and that can be updated easily, are currently not available for Europe as a whole. We investigated the applicability of satellite data for Pan-European Land Cover Monitoring (PELCOM). The main objective was to develop a method by which to obtain a 1 km spatial resolution pan-European land cover database that can be updated easily using National Oceanic and Atmospheric Administration Advanced Very High Resolution Radiometer (NOAA AVHRR) satellite data. The database will be used as input for environmental impact studies and climate research. The study takes full advantage of both multi-spectral and multi-temporal 1 km AVHRR data. The proposed methodology for land cover mapping has its limitations in monitoring changes due to the spatial resolution and the limited accuracy of AVHRR-derived land cover data. Therefore, a change detection technique based on the use of thematic fraction images highlights those areas where the proportions of the various land cover types have changed.     

The IGBP-DIS global 1km land cover data set,DISCover: First results

The International Geosphere-Biosphere Programme Data and Information System (IGBP-DIS) is co-ordinating the development of global land data sets from Advanced Very High Resolution Radiometer (AVHRR) data. The first is a 1 km spatial resolution land cover product 'DISCover', based on monthly Normalized Difference Vegetation Index composites from 1992 and 1993. DISCover is a 17 class land cover dataset based on the science requirements of IGBP elements. Mapping uses unsupervised classification with post-classification refinement using ancillary data. Draft Africa, North America and South America products are now available for peer review.     

Identification of carbon quantifiable regions in the former Soviet Union using unsupervised classification of AVHRR global vegetation index images

Global Vegetation Index (GVI) data from the Advanced Very High Resolution Radiometer (AVHRR) was used to identify macro-scale vegetation/ land cover regions in the former Soviet Union (FSU). These regions are a better representation of surface vegetation and land cover than can be obtained from existing thematic maps of the FSU. Image classes were identified through cluster analysis using the ISODATA clustering algorithm and a maximum likelihood classifier. Qualitative analysis of the image variants produced with different input parameters indicated that an image with 42 classes best represented significant details in vegetation and land cover patterns without producing uninterpretable levels of details that represent artefacts of the clustering algorithm. Initial identification of image classes has been made by considering the weight of evidence provided by quantitative and qualitative analysis of existing maps, analytical tools from class statistics, ancillary data from a variety of sources and expert assessment by Russian scientists with extensive field experience in the FSU. Overall, this method of image classification using GVI data appears to describe accurately regions with similar vegetation and hind cover across the FSU. Some questions regarding the identification of wetlands and potential problems with classification in the Russian high arctic are discussed. The products of this research will help improve carbon budget estimates of the FSU by providing accurate delineation and definition of carbon quantifiable regions.     

Automating land cover mapping of Scotland using expert system and knowledge integration methods

Effective land cover mapping often requires the use of multiple data sources and data interpretation methods, particularly when no one data source or interpretation method provides sufficiently good results. Method-oriented approaches are often only effective for specific land cover class/data source combinations, and cannot be applied when different classification systems or data sources are required or available. Here we present a method, based on Endorsement Theory, of pooling evidence from multiple expert systems and spatial datasets to produce land cover maps. Individual ‘experts’ are trained to produce evidence for or against a class, with this evidence being categorised according to strength. An evidence integration rule set is applied to evidence lists to produce conclusions of different strength regarding individual classes, and the most likely class identified. The only expert system design implemented currently within the methodology is a neural network model, although the system has been designed to accept information from decision trees, fuzzy k-means and Bayesian statistics as well. We have used the technique to produce land cover maps of Scotland using three classification systems of varying complexity. Mapping accuracy varied between 52.6% for a map with 96 classes to 88.8% for a map with eight classes. The accuracy of the maps generated is higher than when individual datasets are used, showing that the evidence integration method applied is suitable for improving land cover mapping accuracy. We showed that imagery was not necessarily the most important data source for mapping where a large number of classes are used, and also showed that even data sources that produce low accuracy scores when used for mapping by themselves do improve the accuracy of maps produced using this integrative approach. Future work in developing the method is identified, including the inclusion of additional expert systems and improvement of the evidence integration, and evaluation carried out of the overall effectiveness of the approach.     

Some challenges in global land cover mapping: An assessment of agreement and accuracy in existing 1 km datasets

Many investigators need and use global land cover maps for a wide variety of purposes. Ironically, after many years of very limited availability, there are now multiple global land cover maps and it is not readily apparent (1) which is most useful for particular applications or (2) how to combine the different maps to provide an improved dataset. The existing global land cover maps at 1 km spatial resolution have arisen from different initiatives and are based on different remote sensing data and employed different methodologies. Perhaps more significantly, they have different legends. As a result, comparison of the different land cover maps is difficult and information about their relative utility is limited. In an attempt to compare the datasets and assess their strengths and weaknesses we harmonized the thematic legends of four available coarse-resolution global land cover maps (IGBP DISCover, UMD, MODIS 1-km, and GLC2000) using the LCCS-based land cover legend translation protocols. Analysis of the agreement among the global land cover maps and existing validation information highlights general patterns of agreement, inconsistencies and uncertainties. The thematic classes of Evergreen broadleaf trees, Snow and Ice, and Barren show high producer and user accuracy and good agreement among the datasets, while classes of mixed tree types show high commission errors. Overall, the results show a limited ability of the four global products to discriminate mixed classes characterized by a mosaic of trees, shrubs, and herbaceous vegetation. There is a strong relationship between class accuracy, spatial agreement among the datasets, and the heterogeneity of landscapes. Suggestions for future mapping projects include careful definition of mixed unit classes, and improvement in mapping heterogeneous landscapes.     

结合地籍数据的高密度城区面向对象遥感分类 

利用高分辨率遥感影像和GIS辅助数据,对高密度城区进行面向对象的土地利用覆被分类研究。使用NAIP高分辨率航空遥感影像,在多尺度影像分割的基础上,针对特定地物选择合适的影像分割参数。采用决策树方法建立高密度城市地区的分类规则,并结合该地区地籍图数据作为辅助数据,逐步进行高密度城市地区地物信息提取。利用辅助数据进行面向对象的遥感分类效果优于单纯依靠遥感影像进行的分类,且有效提取了道路和复杂的房屋等信息,得到了理想的分类结果,其总分类精度从常规面向对象方法的84.08%提高到89.79%。利用辅助数据进行遥感分类提高了高分辨率遥感影像的分类精度,说明了利用辅助数据进行遥感分类方法的有效性。     

Global land cover classification at 1 km spatial resolution using a classification tree approach

This paper on reports the production of a 1 km spatial resolution land cover classification using data for 1992-1993 from the Advanced Very High Resolution Radiometer (AVHRR). This map will be included as an at-launch product of the Moderate Resolution Imaging Spectroradiometer (MODIS) to serve as an input for several algorithms requiring knowledge of land cover type. The methodology was derived from a similar effort to create a product at 8 km spatial resolution, where high resolution data sets were interpreted in order to derive a coarse-resolution training data set. A set of 37 294 x 1 km pixels was used within a hierarchical tree structure to classify the AVHRR data into 12 classes. The approach taken involved a hierarchy of pair-wise class trees where a logic based on vegetation form was applied until all classes were depicted. Multitemporal AVHRR metrics were used to predict class memberships. Minimum annual red reflectance, peak annual Normalized Difference Vegetation Index (NDVI), and minimum channel three brightness temperature were among the most used metrics. Depictions of forests and woodlands, and areas of mechanized agriculture are in general agreement with other sources of information, while classes such as low biomass agriculture and high-latitude broadleaf forest are not. Comparisons of the final product with regional digital land cover maps derived from high-resolution remotely sensed data reveal general agreement, except for apparently poor depictions of temperate pastures within areas of agriculture. Distinguishing between forest and non-forest was achieved with agreements ranging from 81 to 92% for these regional subsets. The agreements for all classes varied from an average of 65% when viewing all pixels to an average of 82% when viewing only those 1 km pixels consisting of greater than 90% one class within the high-resolution data sets.     

NDVI-derived land cover classifications at a global scale

Phenological differences among vegetation types, reflected in temporal variations in the Normalized Difference Vegetation Index (NDVI) derived from satellite data, have been used to classify land cover at continental scales. Extending this technique to global scales raises several issues: identifying land cover types that are spectrally distinct and applicable at the global scale; accounting for phasing of seasons in different parts of the world; validating results in the absence of reliable information on global land cover; and acquiring high quality global data sets of satellite sensor data for input to land cover classifications. For this study, a coarse spatial resolution (one by one degree) data set of monthly NDVI values for 1987 was used to explore these methodological issues. A result of a supervised, maximum likelihood classification of eleven cover types is presented to illustrate the feasibility of using satellite sensor data to increase the accuracy of global land cover information, although the result has not been validated systematically. Satellite sensor data at finer spatial resolutions that include other bands in addition to NDVI, as well as methodologies to better identify and describe gradients between cover types, could increase the accuracy of results of global land cover data sets derived from satellite sensor data.     

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.