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.     

遥感影像超分辨率制图研究进展

尺度问题是土地覆盖分类中的一个核心问题,向下尺度转换又是其中的难点。混合像元分解可以得到亚像元尺度的类别组分百分比,但无法求得亚像元的具体位置。遥感影像超分辨率制图是由粗空间分辨率的影像得到高空间分辨率分类结果图的技术,可用于地表分类向下尺度转换,近年来该技术已成为遥感影像分类和尺度转换领域的研究热点。对超分辨率制图研究进展做了详细论述,从超分辨率制图的发展和研究现状、主要方法、精度评价等几方面进行了详细阐述,并分析了当前超分辨率制图算法存在的主要问题,以及可能的研究重点和发展空间。     

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.     

Estimating the spatial distribution of the population of Riyadh,Saudi Arabia using remotely sensed built land cover and height data

This paper investigates the use of Landsat ETM+, remotely sensed height data, ward-level census population, and dwelling units to downscale population in Riyadh, Saudi Arabia. Regression analysis is used to model the relationship between density of dwelling units and built area proportion at the block level and the coefficients used to downscale density of dwelling units to the parcel level. The population distribution is estimated based on average population per dwelling unit. Seven models were fitted and compared. First, a conventional approach, using ISODATA-classified built land cover alone as a covariate, is used as a benchmark against which to evaluate six more sophisticated models. The conventional model results in low accuracy measured by overall relative error (ORE) (+116%). Approaches for potentially increasing accuracy are explored, incorporating above-surface height data into the downscaling process. These include masking out zero and near-zero height areas when estimating built area; using height to estimate the number of floors; replacing the ISODATA model with a support vector machine; estimating volume-adjusted habitable space; stratifying the study area into different building categories; and preservation of the dependent variable for the best model. These approaches result in large increases in accuracy in the density of dwelling unit estimates. However, while the height data accounts for the vertical dimension (primarily through the number of floors), it is not possible to account for variation in dwelling density which arises due to other factors such as living standards, affluence and other spatially varying factors, without further data.     

Contextual correction: techniques for improving land cover mapping from remotely sensed images

Large area land cover mapping is an important application of remote sensing. A digital land cover map of Great Britain has recently been compiled by supervised classification of Landsat Thematic Mapper data. The work has involved development of a range of post classification procedures to correct contextual errors associated with the use of spectral classification algorithms. This paper describes these procedures and examines their effects upon the map product including a comparison with field survey data.     

Object-based land cover classification of shaded areas in high spatial resolution imagery of urban areas: A comparison study

A significant proportion of high spatial resolution imagery in urban areas can be affected by shadows. Considerable research has been conducted to investigate shadow detection and removal in remotely sensed imagery. Few studies, however, have evaluated how applications of these shadow detection and restoration methods can help eliminate the shadow problem in land cover classification of high spatial resolution images in urban settings. This paper presents a comparison study of three methods for land cover classification of shaded areas from high spatial resolution imagery in an urban environment. Method 1 combines spectral information in shaded areas with spatial information for shadow classification. Method 2 applies a shadow restoration technique, the linear-correlation correction method to create a “shadow-free” image before the classification. Method 3 uses multisource data fusion to aid in classification of shadows. The results indicated that Method 3 achieved the best accuracy, with overall accuracy of 88%. It provides a significantly better means for shadow classification than the other two methods. The overall accuracy for Method 1 was 81.5%, slightly but not significantly higher than the 80.5% from Method 2. All of the three methods applied an object-based classification procedure, which was critical as it provides an effective way to address the problems of radiometric difference and spatial misregistration associated with multisource data fusion (Method 3), and to incorporate thematic spatial information (Method 1).     

基于面向对象方法的遥感影像桥梁提取

面向对象的分析方法是一种有效的高分辨率遥感影像处理技术,提出一种基于图像对象的水上桥梁识别方法。首先采用区域生长方法对影像进行分割,以分割后产生的图像对象为基本处理单元进行分类,提取出水体类别。然后在分析桥梁目标特征的基础上,利用图像对象的形状特征,以及桥梁和水体的上下文关系特征,提取影像中的桥梁目标。最后以实验验证了所提出方法的有效性。     

A cartographic and database approach for land cover/use mapping and generalization from remotely sensed data

Classification of remotely sensed data involves a set of generalization processes, i.e. reality is simplified to a set of a few classes that are relevant to the application under consideration. This article introduces an approach to image classification that uses a class hierarchy structure for mapping unit definition at different generalization levels. This structure is implemented as an operational relational database and allows querying of more detailed land cover/use information from a higher abstraction level, which is that viewed by the map user. Elementary mapping units are defined on the basis of an unsupervised classification process in order to determine the land cover/use classes registered in the remotely sensed data. Mapping unit composition at different generalization levels is defined on the basis of membership values of sampled pixels to land cover/use classes. Unlike fuzzy classifications, membership values are presented to the user at mapping unit level.     

Multiple-point simulation-based method for extraction of objects with spatial structure from remotely sensed imagery

Classification of Combining Spectral information and Spatial information upon Multiple-point statistics (CCSSM) is a method for information extraction that introduces multiple-point simulation (MPS) to increase the classification accuracy of remotely sensed imagery data by incorporating structural information through a training image. This paper focuses on (1) applying CCSSM using a multigrid approach to a Satellite Pour l'Observation de la Terre (SPOT) 5 image, (2) adopting consensus-based fusion to combine two different information sources, the spectral information from supervised classification and spatial structure information from the MPS and (3) analysing the change trend for the accuracy of information extraction and optimizing the proportions in the combination of the two different information sources. We demonstrate that, even if the spectral information from the SPOT 5 image used in the classification results in better classification accuracy, with the introduction of spatial structure information from MPS the accuracy of the information extraction can still be increased significantly.     

Evaluating Hyperion capability for land cover mapping in a fragmented ecosystem: Pollino National Park, Italy

We analyze the capability of Hyperion spaceborne hyperspectral data for discriminating land cover in a complex natural ecosystem according to the structure of the currently used European standard classification system (CORINE Land Cover 2000). For this purpose, we used Hyperion imagery acquired over Pollino National Park (Italy).Hyperion pre-processed data (30 m spatial resolution) were classified at the pixel level using common parametric supervised classification methods. The algorithms' performance and class level accuracy were compared with those obtained for the same area using airborne hyperspectral MIVIS data (7 m spatial resolution).Moreover, in selected test areas characterized by heterogeneous land cover (as mapped by MIVIS classification) a Linear Spectral Unmixing (LSU) technique was applied to Hyperion data to derive the abundance fractions of land cover endmembers. The accuracy of the LSU analysis was evaluated using the Residual Error parameter, by comparing Hyperion LSU results with land cover fractional abundances achieved from reference data (i.e., MIVIS and air-photo classification).The results show the potential of Hyperion spaceborne hyperspectral imagery in mapping land cover and vegetation diversity up to the 4th level of the CORINE legend, even at the sub-pixel level, within a fragmented ecosystem such as that of Pollino National Park. Moreover, we defined a criterion for evaluating the Hyperion accuracy in retrieving land cover abundances at the sub-pixel scale. Sub-pixel analysis allowed us to determine the optimal threshold to select the areas on which consistent fractional land cover monitoring can be achieved using the Hyperion sensor.     

Wetland mapping in the Congo Basin using optical and radar remotely sensed data and derived topographical indices

This study reports results of a classification tree approach to mapping the wetlands of the Congo Basin, focusing on the Cuvette Centrale of the Congo River watershed, an area of 1,176,000 km². Regional expert knowledge was used to train passive optical remotely sensed imagery of the Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) sensors, JERS-1 active radar L-band imagery, and topographical indices derived from 3 arc sec elevation data of the Shuttle Radar Topography Mission (SRTM). All data inputs were resampled to a common 57 m resolution grid. A classification tree bagging procedure was employed to produce a final map of per-grid cell wetland probability. Thirty bagged trees were ranked and the median result was selected to produce the final wetland probability map. Thresholding the probability map at < 0.5 yielded a proportion of wetland cover for the study area of 32%, equivalent to 360,000 km². Wetlands predominate in the CARPE Lake Tele-Lake Tumba landscape located in the western part of the Democratic Republic of the Congo and the south-eastern Republic of Congo, where they constitute 56% of the landscape. Local topography depicting relative elevation for sub-catchments proved to be the most valuable discriminator of wetland cover. However, all sources of information (i.e. optical, radar and topography) featured prominently in contributing to the classification tree procedure, reinforcing the idea that multi-source data are useful in the characterization of wetland land cover. The method employed freely available data and a fully automated process, except for training data collection. Comparisons to existing maps and in situ field observations indicate improvements compared to previous efforts.     

A contextual classification method for recognizing land use patterns in high resolution remotely sensed data

This paper describes a CONtextual ANalysis procedure (CONAN) which is designed to recognize land use patterns in high resolution remotely sensed data by analysis of the local frequency distribution of scene components (i.e. ground cover type classes). The procedure was tested with randomly generated synthetic data developed to simulate the frequency distribution of cover type components for four land use classes. It was found that the accuracy in discriminating between the four test classes depends upon the size of the pixel neighborhood used to compute the component frequency distribution.     

Remote sensing of small and linear features: Quantifying the effects of patch size and length, grid position and detectability on land cover mapping

The accurate mapping of small, often fragmented and linear vegetation patches is of key importance for natural resource management because of their ecological significance. However, due to their small size and the quantised nature of remote sensing imagery they may be under-represented in the landscape when mapped using earth observation. This paper investigates the effect of patch area and patch elongation on the accurate mapping of these vegetation patches. Using synthetic images to simulate sub-pixel patch location, we investigated classification accuracy and extraction probability resulting from differences in the geometric properties of the raster grid and the feature alone. We simulated the effect of grid position, detectability, feature size and shape on classification. This represents the highest achievable accuracy using the remote sensing raster grid, where other factors influencing classification such as classification algorithm, radiometric calibration and sensor characteristics are excluded. We found that mapping error was highest when the scale of the feature and the raster grid coincided. We showed that the spatial resolution of the grid should be many times finer in order to extract these features accurately. For square patches with a mean classification accuracy of 75%, the grid pixel area was 11 times smaller than patch size. When patches were small and/or elongated, the probability of extraction was reduced, mapping accuracies decreased and variability in accuracy due to the effects of grid position increased. For example, a square shaped patch needed an area of at least 11 pixels to achieve a mean accuracy of 75%, whilst a linear patch with a width to length ratio of 4 needed an area of 12.3 pixels. This paper quantifies the limitations of remote sensing for the accurate detection of small and linear features and provides guidelines on the appropriate spatial resolution required to map these features. Using our results, map users can estimate the probability of a map classifying small and linear features independently of the error matrix. Furthermore, we provide a more precise estimate of the size of the smallest discernable feature taking into account the random position of the remote sensing grid with respect to the feature as well as its shape. An understanding of this phenomenon is critical for making good land management decisions based on a thorough understanding of the limitations of remote sensing data.     

A method for detecting large-scale forest cover change using coarse spatial resolution imagery

Many large countries, including Canada, rely on earth observation as a practical and cost-effective means of monitoring their vast inland ecosystems. A potentially efficient approach is one that detects vegetation changes over a hierarchy of spatial scales ranging from coarse to fine. This paper presents a Change Screening Analysis Technique (Change-SAT) designed as a coarse filter to identify the location and timing of large (>5-10 km²) forest cover changes caused by anthropogenic and natural disturbances at an annual, continental scale. The method uses change metrics derived from 1-km multi-temporal SPOT VEGETATION and NOAA AVHRR imagery (reflectance, temperature, and texture information) and ancillary spatial variables (proximity to active fires, roads, and forest tenures) in combination with logistic regression and decision tree classifiers. Major forest changes of interest include wildfires, insect defoliation, forest harvesting, and flooding. Change-SAT was tested for 1998-2000 using an independent sample of change and no-change sites over Canada. Overall accuracy was 94% and commission error, especially critical for large-area change applications, was less than 1%. Regions identified as having major or widespread changes could be targeted for more detailed investigation and mapping using field visits, aerial survey, or fine resolution EO methods, such as those being applied under Canadian monitoring programs. This multi-resolution approach could be used as part of a forest monitoring system to report on carbon stocks and forest stewardship.     

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.     

A linear mapping technique for optimizing binary templates in noise-free pattern matching

Dimensionality reduction of noise-free binary patterns is achieved here by linear mapping from higher dimensional vector space to a lower one. A systematic methodology is thus obtained for finding better templates for a collection of noise free patterns, a priori knowledge of the probability distribution of which is not required. Here the memory requirement of the classifier is reduced and the template matching made faster under certain conditions.     

Approaches to fractional land cover and continuous field mapping: A comparative assessment over the BOREAS study region

Subpixel land cover mapping involves the estimation of surface properties using sensors whose spatial sampling is coarse enough to produce mixtures of the properties within each pixel. This study evaluates five algorithms for mapping subpixel land cover fractions and continuous fields of vegetation properties within the BOREAS study area. The algorithms include a conventional “hard”, per-pixel classifier, a neural network, a clustering/look-up-table approach, multivariate regression, and linear least squares inversion. A land cover map prepared using a Landsat TM mosaic was adopted as the source of fine scale calibration and validation data. Coarse scale mixtures of five basic land cover classes and continuous vegetation fields, both corresponding to the field of view of SPOT-VEGETATION imagery (1.15-km pixel size), were synthesised from the TM mosaic using a modelled point spread function. Two measures of land cover distribution were used, fractions of fine scale land cover categories and continuous fields of vegetation structural characteristics. The subpixel algorithms were applied using both proximate (<100 km) and distant (>400 km) separation between training and validation regions. “Hard” classification performed poorly in estimating proportions or continuous fields. The neural network, look-up-table and multivariate regression algorithms produced good matches of spatial patterns and regional land cover composition for the proximate treatment. However, all three methods exhibited substantial biases with the distant treatment due to the characteristics of the training data. Linear least squares inversion offers a relatively unbiased but less precise alternative for subpixel proportion and fraction mapping as it avoids calibration to the a priori distribution of land cover in the training data. In general, a combination of multivariate regression for proximate training data and linear least squares inversion for distant training data resulted in woody fraction estimates within 20% of the Landsat TM classification-based estimates.     

Feature selection by genetic algorithms in object-based classification of IKONOS imagery for forest mapping in Flanders, Belgium

Obtaining detailed information about the amount of forest cover is an important issue for governmental policy and forest management. This paper presents a new approach to update the Flemish Forest Map using IKONOS imagery. The proposed method is a three-step object-oriented classification routine that involves the integration of 1) image segmentation, 2) feature selection by Genetic Algorithms (GAs) and 3) joint Neural Network (NN) based object-classification. The added value of feature selection and neural network combination is investigated. Results show that, with GA-feature selection, the mean classification accuracy (in terms of Kappa Index of Agreement) is significantly higher (p < 0.01) than without feature selection. On average, the summed output of 50 networks provided a significantly higher (p < 0.01) classification accuracy than the mean output of 50 individual networks. Finally, the proposed classification routine yields a significantly higher (p < 0.01) classification accuracy as compared with a strategy without feature selection and joint network output. In addition, the proposed method showed its potential when few training data were available.     

A new GLM-based method for mapping tree cover continuous fields using regional MODIS reflectance data

Knowledge about land cover and its change is an important input for the monitoring and modeling of ecological and environmental processes from the regional to the global scale. Considerable efforts have been made to develop global continuous fields for different land cover types at large spatial scales based on NOAA-AVHRR and TERRA-MODIS data and a range of techniques have been applied to depict the sub-pixel fraction of land cover types from these data. In this study, a new methodology is described for deriving and optimizing continuous fields of tree cover for complex topography at the regional scale of the European Alps using generalized linear models (GLM). MODIS data (MOD09) at a spatial resolution of 500 m were used to calibrate the models against regional training data of fractional tree cover. For evaluating the method we test the GLM model output to a regression tree model (using the same data structure). Further we test the resulting GLM-based tree cover continuous fields against two different, independent test data sets; one of which is spatially separated and the other is from within the calibration area. Finally, we compare the GLM model output with two available global data sets at spatial resolutions of 1 km and 3 km: (1) TERRA-MODIS Vegetation Continuous Fields product (MOD44), and (2) the NOAA-AVHRR vegetation continuous fields. Our GLM-based method results in high accuracy (MAE=9.1%) and low bias (−1.2%) across the combined evaluation and calibration area, and with small differences only between the calibration and the spatially separated evaluation area (1.3%). Compared to the regression tree model the results from the GLM model for all analyses are significantly better. Thus we conclude that generalized linear models are appropriate for deriving continuous fields of fractional tree cover for complex topography at the regional scale. GLMs can handle nonlinear relationships present in the training data set well, and the method is robust with respect to sample size and the number of months used for calibration. Regional calibrations of vegetation continuous fields may offer significantly improved predictions compared to globally calibrated models. Such regionally calibrated and optimized models may serve as valuable tools for regional monitoring of land cover pattern and its temporal change.