Sub-pixel land cover mapping for per-field classification

A method was developed to transform a soft land cover classification into hard land cover classes at the sub-pixel scale for subsequent per-field classification. First, image pixels were segmented using vector boundaries. Second, the pixel segments (ranked by area) were labelled with a land cover class (ranked by class typicality). Third, a hard per-field classification was generated by examining each polygon (representing a land cover parcel, or field) in its entirety (by grouping the fragments of the polygon contained within different image pixels) and assigning to it the modal land cover class. The accuracy of this technique was considerably higher than that of both a corresponding hard per-pixel classification and a perfield classification based on hard per-pixel classified imagery.     

Boosting a fast neural network for supervised land cover classification

It is demonstrated that the use of an ensemble of neural networks for routine land cover classification of multispectral satellite data can lead to a significant improvement in classification accuracy. Specifically, the AdaBoost.M1 algorithm is applied to a sequence of three-layer, feed-forward neural networks. In order to overcome the drawback of long training time for each network in the ensemble, the networks are trained with an efficient Kalman filter algorithm. On the basis of statistical hypothesis tests, classification performance on multispectral imagery is compared with that of maximum likelihood and support vector machine classifiers. Good generalization accuracies are obtained with computation times of the order of 1 h or less. The algorithms involved are described in detail and a software implementation in the ENVI/IDL image analysis environment is provided.     

Contribution of multispectral and multitemporal information from MODIS images to land cover classification

The goal of this study is to evaluate the relative usefulness of high spectral and temporal resolutions of MODIS imagery data for land cover classification. In particular, we highlight the individual and combinatorial influence of spectral and temporal components of MODIS reflectance data in land cover classification. Our study relies on an annual time series of twelve MODIS 8-days composited images (MOD09A1) monthly acquired during the year 2000, at a 500 m nominal resolution. As our aim is not to propose an operational classifier directed at thematic mapping based on the most efficient combination of reflectance inputs — which will probably change across geographical regions and with different land cover nomenclatures — we intentionally restrict our experimental framework to continental Portugal. Because our observation data stream contains highly correlated components, we need to rank the temporal and the spectral features according not only to their individual ability at separating the land cover classes, but also to their differential contribution to the existing information. To proceed, we resort to the median Mahalanobis distance as a statistical separability criterion. Once achieved this arrangement, we strive to evaluate, in a classification perspective, the gain obtained when the dimensionality of the input feature space grows. We then successively embedded the prior ranked measures into the multitemporal and multispectral training data set of a Support Vector Machines (SVM) classifier. In this way, we show that, only the inclusion of the approximately first three dates substantially increases the classification accuracy. Moreover, this multitemporal factor has a significant effect when coupled with combinations of few spectral bands, but it turns negligible as soon as the full spectral information is exploited. Regarding the multispectral factor, its beneficence on classification accuracy remains more constant, regardless of the number of dates being used.     

Hierarchical mapping of Northern Eurasian land cover using MODIS data

The Northern Eurasian land mass encompasses a diverse array of land cover types including tundra, boreal forest, wetlands, semi-arid steppe, and agricultural land use. Despite the well-established importance of Northern Eurasia in the global carbon and climate system, the distribution and properties of land cover in this region are not well characterized. To address this knowledge and data gap, a hierarchical mapping approach was developed that encompasses the study area for the Northern Eurasia Earth System Partnership Initiative (NEESPI). The Northern Eurasia Land Cover (NELC) database developed in this study follows the FAO-Land Cover Classification System and provides nested groupings of land cover characteristics, with separate layers for land use, wetlands, and tundra. The database implementation is substantially different from other large-scale land cover datasets that provide maps based on a single set of discrete classes. By providing a database consisting of nested maps and complementary layers, the NELC database provides a flexible framework that allows users to tailor maps to suit their needs. The methods used to create the database combine empirically derived climate–vegetation relationships with results from supervised classifications based on Moderate Resolution Imaging Spectroradiometer (MODIS) data. The hierarchical approach provides an effective framework for integrating climate–vegetation relationships with remote sensing-based classifications, and also allows sources of error to be characterized and attributed to specific levels in the hierarchy. The cross-validated accuracy was 73% for the land cover map and 73% and 91% for the agriculture and wetland classifications, respectively. These results support the use of hierarchical classification and climate–vegetation relationships for mapping land cover at continental scales.     

Comparison of pixel- and object-based classification in land cover change mapping

Land use/land cover (LULC) change occurs when humans alter the landscape, and this leads to increasing loss, fragmentation and spatial simplification of habitat. Many fields of study require monitoring of LULC change at a variety of scales. LULC change assessment is dependent upon high-quality input data, most often from remote sensing-derived products such as thematic maps. This research compares pixel- and object-based classifications of Landsat Thematic Mapper (TM) data for mapping and analysis of LULC change in the mixed land use region of eastern Ontario for the period 1995–2005. For single date thematic maps of 10 LULC classes, quantitative and visual analyses showed no significant accuracy difference between the two methods. The object-based method produced thematic maps with more uniform and meaningful LULC objects, but it suffered from absorption of small rare classes into larger objects and the incapability of spatial parameters (e.g. object shape) to contribute to class discrimination. Despite the similar map accuracies produced by the two methods, temporal change maps produced using post-classification comparison (PCC) and analysed using intensive visual analysis of errors of omission and commission revealed that the object-based maps depicted change more accurately than maximum likelihood classification (MLC)-derived change maps.     

Comparing MODIS and ETM+ data for regional and global land classification

Nearly simultaneous reflectance data sets from the Landsat 7 Enhanced Thematic Mapper Plus (ETM+), at 30-m resolution, and the Terra satellite instrument MODIS, at 500-m resolution, are compared for their ability to map fractional coverage of surface types over large areas. Lower spatial resolution MODIS classification results are generally comparable those of ETM+, with discrepancies for some regions with mixed surface types. Analysis of laboratory and field spectra suggests an ambiguity, the “brightness ambiguity”, which can prevent accurate area estimation of pixels having two or more surface types. This ambiguity, plus general mathematical inversion issues, can account for the discrepancy. Thus, occasional high-resolution measurements, as from Landsat 7, are necessary to refine estimations of large area surface types from MODIS and similar lower spatial resolution instruments.     

Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data

Automatic land cover classification from satellite images is an important topic in many remote sensing applications. In this paper, we consider three different statistical approaches to tackle this problem: two of them, namely the well-known maximum likelihood classification (ML) and the support vector machine (SVM), are noncontextual methods. The third one, iterated conditional modes (ICM), exploits spatial context by using a Markov random field. We apply these methods to Landsat 5 Thematic Mapper (TM) data from Tenerife, the largest of the Canary Islands. Due to the size and the strong relief of the island, ground truth data could be collected only sparsely by examination of test areas for previously defined land cover classes.We show that after application of an unsupervised clustering method to identify subclasses, all classification algorithms give satisfactory results (with statistical overall accuracy of about 90%) if the model parameters are selected appropriately. Although being superior to ML theoretically, both SVM and ICM have to be used carefully: ICM is able to improve ML, but when applied for too many iterations, spatially small sample areas are smoothed away, leading to statistically slightly worse classification results. SVM yields better statistical results than ML, but when investigated visually, the classification result is not completely satisfying. This is due to the fact that no a priori information on the frequency of occurrence of a class was used in this context, which helps ML to limit the unlikely classes.     

MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets

Information related to land cover is immensely important to global change science. In the past decade, data sources and methodologies for creating global land cover maps from remote sensing have evolved rapidly. Here we describe the datasets and algorithms used to create the Collection 5 MODIS Global Land Cover Type product, which is substantially changed relative to Collection 4. In addition to using updated input data, the algorithm and ancillary datasets used to produce the product have been refined. Most importantly, the Collection 5 product is generated at 500-m spatial resolution, providing a four-fold increase in spatial resolution relative to the previous version. In addition, many components of the classification algorithm have been changed. The training site database has been revised, land surface temperature is now included as an input feature, and ancillary datasets used in post-processing of ensemble decision tree results have been updated. Further, methods used to correct classifier results for bias imposed by training data properties have been refined, techniques used to fuse ancillary data based on spatially varying prior probabilities have been revised, and a variety of methods have been developed to address limitations of the algorithm for the urban, wetland, and deciduous needleleaf classes. Finally, techniques used to stabilize classification results across years have been developed and implemented to reduce year-to-year variation in land cover labels not associated with land cover change. Results from a cross-validation analysis indicate that the overall accuracy of the product is about 75% correctly classified, but that the range in class-specific accuracies is large. Comparison of Collection 5 maps with Collection 4 results show substantial differences arising from increased spatial resolution and changes in the input data and classification algorithm.     

M5 model tree for land cover classification

Tree based regression models like a M5 algorithm represent a promising development in machine learning research. A recent study suggests that a M5 model tree algorithm can be used for classification problems after some modification. This letter explores the usefulness of a M5 model tree for classification problems using multispectral (Landsat‐7 Enhanced Thematic Mapper Plus (ETM+)) for a test area in eastern England. Classification accuracy achieved by using a M5 model tree is compared with a univariate decision tree with and without using boosting. Results show that the M5 model tree achieves a significantly higher level of classification accuracy than a decision tree and works equally well to a boosted decision tree. Further, a model tree based classification algorithm works well with small as well as noisy datasets.     

Advances in classification for land cover mapping using SPOT HRV imagery

Abstract

High-resolution data from the HRV (High Resolution Visible) sensors onboard the SPOT-1 satellite have been utilized for mapping semi-natural and agricultural land cover using automated digital image classification algorithms. Two methods for improving classification performance are discussed. The first technique involves the use of digital terrain information to reduce the effects of topography on spectral information while the second technique involves the classification of land-cover types using training data derived from spectral feature space. Test areas in Snowdonia and the Somerset Levels were used to evaluate the methodology and promising results were achieved. However, the low classification accuracies obtained suggest that spectral classification alone is not a suitable tool to use in the mapping of semi-natural cover types.     

Extraction of training samples from time-series MODIS imagery and its utility for land cover classification

A number of classification techniques to generate land cover maps from satellite imagery have been proposed but supervised classification with manual selection and delineation of training samples (TSs) continues to be the preferred technique. The current practices of field visits and manual delineation of TSs by visual recognition are highly demanding on both resources and time, with limited utility. With an increase in the number of Earth Observation Satellite (EOS) platforms and the enormous data that they generate, there is a need to process the data quickly and efficiently for creating global science products. Towards this goal, an attempt has been made in this article to develop a method for the automatic extraction of the TSs from the time series of Moderate Resolution Imaging Spectroradiometer (MODIS – 250 m) vegetation index (VI), which can then be used for supervised classification to create a land cover map with any classification technique on relevant remotely sensed data. The TSs contained 1.27%, 0.09% and 1.18% of the total pixels for the forest, crop and water classes of the study region. Validation with Advanced Wide Field Sensor (AWiFS – 56 m)-derived national land use/land cover (LULC) map of India shows a complete agreement with the location of what can be considered as pure class pixels. The article also demonstrates and compares the utility of these TSs with an expert choice of TSs on MODIS time-series data using k-nearest neighbour, and support vector machine (SVM) classifiers and on a single-scene Linear Imaging Self-Scanning Sensor-3 (LISS-3 – 24 m) imagery using maximum likelihood (ML) classifier.     

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.     

GLC2000: a new approach to global land cover mapping from Earth observation data

A new global land cover database for the year 2000 (GLC2000) has been produced by an international partnership of 30 research groups coordinated by the European Commission's Joint Research Centre. The database contains two levels of land cover information—detailed, regionally optimized land cover legends for each continent and a less thematically detailed global legend that harmonizes regional legends into one consistent product. The land cover maps are all based on daily data from the VEGETATION sensor on‐board SPOT 4, though mapping of some regions involved use of data from other Earth observing sensors to resolve specific issues. Detailed legend definition, image classification and map quality assurance were carried out region by region. The global product was made through aggregation of these. The database is designed to serve users from science programmes, policy makers, environmental convention secretariats, non‐governmental organizations and development‐aid projects. The regional and global data are available free of charge for all non‐commercial applications from http://www.gvm.jrc.it/glc2000.     

Fully fuzzy supervised classification of land cover from remotely sensed imagery with an artificial neural network

Mixed pixels are a major problem in mapping land cover from remotely sensed imagery. Unfortunately, such imagery may be dominated by mixed pixels, and the conventional hard image classification techniques used in mapping applications are unable to appropriately represent the land cover of mixed pixels. Fuzzy classification techniques can, however, accommodate the partial and multiple class membership of mixed pixels, and be used to derive an appropriate land cover representation. This is, however, only a partial solution to the mixed pixel problem in supervised image classification. It must be reognised that the land cover on the ground is fuzzy, at the scale of the pixel, and so it is inappropriate to use procedures designed for hard data in the training and testing stages of the classification. Here an approach for land cover classification in which fuzziness is accommodated in all three stages of a supervised classification is presented. Attention focuses on the classification of airborne thematic mapper data with an artificial neural network. Mixed pixels could be accommodated in training the artificial neural network, since the desired output for each training pixel can be specified. A fuzzy land cover representation was derived by outputting the activation level of the network's output units. The activation level of each output unit was significantly correlated with the proportion of the area represented by a pixel which was covered with the class associated with the unit (r>0.88, significant at the 99% level of confidence). Finally, the distance between the fuzzy land cover classification derived from the artificial neural network and the fuzzy ground data was used to illustrate the accuracy of the land cover representation derived. The dangers of hardening the classification output and ground data sets to enable a conventional assessment of classification accuracy are also illustrated; the hardened data sets were over three times more distant from each other than the fuzzy data sets.     

A scalable approach to mapping annual land cover at 250 m using MODIS time series data: A case study in the Dry Chaco ecoregion of South America

Land use and land cover (LULC) maps from remote sensing are vital for monitoring, understanding and predicting the effects of complex human-nature interactions that span local, regional and global scales. We present a method to map annual LULC at a regional spatial scale with source data and processing techniques that permit scaling to broader spatial and temporal scales, while maintaining a consistent classification scheme and accuracy. Using the Dry Chaco ecoregion in Argentina, Bolivia and Paraguay as a test site, we derived a suite of predictor variables from 2001 to 2007 from the MODIS 250 m vegetation index product (MOD13Q1). These variables included: annual statistics of red, near infrared, and enhanced vegetation index (EVI), phenological metrics derived from EVI time series data, and slope and elevation. For reference data, we visually interpreted percent cover of eight classes at locations with high-resolution QuickBird imagery in Google Earth. An adjustable majority cover threshold was used to assign samples to a dominant class. When compared to field data, we found this imagery to have georeferencing error < 5% the length of a MODIS pixel, while most class interpretation error was related to confusion between agriculture and herbaceous vegetation. We used the Random Forests classifier to identify the best sets of predictor variables and percent cover thresholds for discriminating our LULC classes. The best variable set included all predictor variables and a cover threshold of 80%. This optimal Random Forests was used to map LULC for each year between 2001 and 2007, followed by a per-pixel, 3-year temporal filter to remove disallowed LULC transitions. Our sequence of maps had an overall accuracy of 79.3%, producer accuracy from 51.4% (plantation) to 95.8% (woody vegetation), and user accuracy from 58.9% (herbaceous vegetation) to 100.0% (water). We attributed map class confusion to limited spectral information, sub-pixel spectral mixing, georeferencing error and human error in interpreting reference samples. We used our maps to assess woody vegetation change in the Dry Chaco from 2002 to 2006, which was characterized by rapid deforestation related to soybean and planted pasture expansion. This method can be easily applied to other regions or continents to produce spatially and temporally consistent information on annual LULC.     

Fully-fuzzy supervised classification of sub-urban land cover from remotely sensed imagery: Statistical and artificial neural network approaches

Fully-fuzzy classification approaches have attracted increasing interest recently. These approaches allow for multiple and partial class memberships at the level of individual pixels and accommodate fuzziness in all three stages of a supervised classification of remotely sensed imagery. A fully-fuzzy classification strategy may be deemed more objective and correct than partially-fuzzy approaches where fuzziness is only accommodated in one or two of the three classification stages. This paper describes two approaches to the fully-fuzzy classification of remotely sensed imagery: a statistical approach based on a modified fuzzy c-means clustering algorithm performed in a supervised mode and an artificial neural network based approach. This is followed by the documentation of a case study using Landsat Thematic Mapper (TM) data of an Edinburgh suburb. Both approaches were applied to derive fully-fuzzy classifications of land cover, with fuzzy ground data, critical for training and testing the classifications, derived from indicator kriging. Results confirmed the superiority of fully-fuzzy over their respective partially-fuzzy classification counterparts, which is beneficial given their more relaxed requirements for training pixels (i.e. training pixels need not be pure). Similar accuracies were obtained with the artificial neural network and statistical approaches to classification. It is suggested that due emphasis must be placed on derivation and analysis of fuzzy ground data as well as fuzzy classified data in order to further improve fully-fuzzy classifications.     

Characterizing vegetation cover in global savannas with an annual foliage clumping index derived from the MODIS BRDF product

The global savanna biome is characterized by enormous diversity in the physiognomy and spatial structure of the vegetation. The foliage clumping index can be calculated from bidirectional reflectance distribution function (BRDF) data. It measures the response of the darkspot reflectance to increased shadow associated with clumped vegetation and is related to leaf area index. Clumping index theoretically declines with increasing woody cover until the tree canopy begins to become uniform. In this study, clumping index is calculated for Moderate Resolution Imaging Spectroradiometer BRDF data for the Australian tropical savanna, the tropical savannas of South America, and the tropical savannas of east, west and southern Africa and compared with site-based measurements of tree canopy cover, and with area-based classifications of land cover. There were differences in sensitivity of clumping index between red and near-infrared reflectance channels, and between savanna systems with markedly different woody vegetation physiognomy. Clumping index was broadly related to foliage cover from historical site data in Australia and in West Africa and Kenya, but not in Southern Africa nor with detailed site-based demographic data in the cerrado of Brazil. However, clumping index decreased with proportion of woody cover in land cover datasets for east Africa, Australia and the Colombian Llanos. There was overlap in the range of clumping index values for forest, cerrado and campo land covers in Brazil. Clumping index was generally negatively correlated with percentage tree cover from the MODIS Vegetation Continuous Fields product, but regional differences in the relationship were evident. There were large differences in the frequency distributions of clumping index from savanna, woody savanna and grassland land cover classes between global ecoregions. The clumping index shows differing sensitivity to savanna woody cover for red and NIR reflectance, and requires regional calibration for application as a universal indicator.     

Relating spatial patterns of fractional land cover to savanna vegetation morphology using multi-scale remote sensing in the Central Kalahari

Understanding, monitoring, and managing savanna ecosystems requires characterizing both functional and structural properties of vegetation. From a functional perspective, in savannas, quantitative estimation of fractional cover of photosynthetic vegetation (f_PV), non-photosynthetic vegetation (f_NPV), and bare soil (f_BS) is important as it relates to carbon dynamics and ecosystem function. On the other hand, vegetation morphology classes describe the structural properties of the ecosystem. Due to high functional diversity and structural heterogeneity in savannas, accurately characterizing both these properties using remote sensing is methodologically challenging. While mapping both fractional cover and vegetation morphology classes are important research themes within savanna remote sensing, very few studies have considered systematic investigation of their spatial association across different spatial resolutions. Focusing on the semi-arid savanna ecosystem in the Central Kalahari, this study utilized f_PV, f_NPV, and f_BS derived in situ and estimated from spectral unmixing of high- (GeoEye-1), medium- (Landsat TM), and coarse- (MODIS) spatial resolution imagery to investigate: (i) the impact of reducing spatial resolution on both magnitude and accuracy of fractional cover; and (ii) how fractional-cover magnitude and accuracy are spatially associated with savanna vegetation morphology classes. Endmembers for Landsat TM and GeoEye-1 were derived from the image based on purity measures; for MODIS (MCD43A4), the challenge of finding spectral endmembers was addressed following an empirical multi-scale hierarchical approach. GeoEye-1-derived fractional estimates showed comparatively closest agreement with in situ measurements and were used to evaluate Landsat TM and MODIS. Overall results indicate that increasing pixel size caused consistent increases in variance of and error in fractional-cover estimates. Even at coarse spatial resolution, f_PV was estimated with higher accuracy compared with f_NPV and f_BS. Assessment considering vegetation morphology of samples revealed both morphology- and cover-specific differences in accuracy. At larger pixel sizes, in areas with dominant woody vegetation, f_PV was overestimated at the cost of mainly underestimating f_BS; in contrast, in areas with dominant herbaceous vegetation, f_NPV was overestimated with a corresponding underestimation of both f_PV and f_BS. These results underscore that structural and functional heterogeneity in semi-arid savanna both impact retrieval of fractional cover, suggesting that comprehensive remote sensing of savannas needs to take both structure and cover into account.     

Applying evidential reasoning methods to agricultural land cover classification

Land cover classification remains an important and complex problem when land observational satellites are involved. Conventional methods of image classification have been shown to provide adequate results; however, when complex surface arrangements are encountered the reliability of conventional approaches can be called into question. This situation is particularly apparent when land cover classes cannot be well discriminated, or when land cover categories are sufficiently broad in their definition that a single spectral response pattern cannot adequately capture the inherent within variation associated with that class. In this paper the problem of agricultural land mapping was examined and the potential for applying soft-classification procedures based on the Dempster-Shafer Theory of Evidence was demonstrated. Results of this study show that applying Dempster-Shafer Theory in image classification can yield thematic maps with accuracies that can support their operational use. In addition, through the application of this technique, a framework can be developed to support and guide the use of subjective judgement during the classification process and permit greater flexibility in the formulation of informational classes.     

NOAA-AVHRR data processing for the mapping of vegetation cover

The NOAA-AVHRR images have been widely used for global studies due to their low cost, suitable wavebands and high temporal resolution. Data from the AVHRR sensor (Bands 1 and 2) transformed to the Normalized Difference Vegetation Index (NDVI) are the most common product used in global land cover studies. The purpose of this Letter is to present the vegetation, soil, and shade fraction images derived from AVHRR, in addition to NDVI, to monitor land cover. Six AVHRR images from the period of 21 to 26 June 1993 were composed and used to obtain the above mentioned products over Sa o Paulo State, in the south-east of Brazil. Vegetation fraction component values were strongly correlated with NDVI values ( r 0.95; n 60). Also, the fraction image presented a good agreement with the available global vegetation map of Sao Paulo State derived from Landsat TM images.