Extreme‐learning‐machine‐based land cover classification

The extreme learning machine (ELM), a single hidden layer neural network based supervised classifier is used for remote sensing classifications. In comparison to the backpropagation neural network, which requires the setting of several user‐defined parameters and may produce local minima, the ELM requires setting of one parameter, and produces a unique solution for a set of randomly assigned weights. Two datasets, one multispectral and another hyperspectral, were used for classification. Accuracies of 89.0% and 91.1% are achieved with this classifier using multispectral and hyperspectral data, respectively. Results suggest that the ELM provides a classification accuracy comparable to a backpropagation neural network with both datasets. The computational cost using the ELM classifier (1.25 s with Enhanced Thematic Mapper (ETM+) and 0.675 s with Digital Airborne Imaging Spectrometer (DAIS) data) is very small in comparison to the backpropagation neural network.     

SAR‐based land cover classification of Kuwait

Orbital synthetic aperture radar (SAR) C‐band data acquired by ERS‐1/2 in vv‐polarization and Radarsat in hh‐polarization during the period from 1996 to 1999 were used to evaluate their combined information potential for classification of land cover in the arid environment of Kuwait. Individual SAR scenes were orthorectified using a digital elevation model (DEM) of Kuwait, radiometrically adjusted for incidence angle effects, and mosaics were generated for the whole country. The data were coregistered as multichannel composites and integrated with geographical information system (GIS) layers of roads, hydrology, soils and vegetation. An adaptive spatial filter was used to increase the number of effective independent looks prior to generation of feature vectors based on SAR backscatter power values. A total of 13 classes of the joint ERS‐1/2 and Radarsat images were identified based on Bhattacharya distance and geospatial pattern. The C‐band radar backscatter observed by ERS and Radarsat was found to be related to vegetation cover, surface roughness, percentage of coarse material in the surface layer and moisture conditions. These factors are not independent, but are known to be correlated. The complexity of these dependencies made unambiguous classification of surface material difficult when using C‐band data alone. Nevertheless, class labels were assigned using a maximum likelihood supervised classification incorporating field measurements and ancillary data such as soil, and surface sediment maps. When used in a simple two‐class classification (e.g. low vs. high vegetation cover fraction, or smooth vs. rough soils), the overall accuracy of the combined ERS and Radarsat data was between 70 and 80%. The generated dataset is amenable to several label definitions based on the requirements of the intended use.     

A novel self‐organizing neuro‐fuzzy multilayered classifier for land cover classification of a VHR image

A novel self‐organizing neuro‐fuzzy multilayered classifier (SONeFMUC) is introduced in this paper, with feature selection capabilities, for the classification of an IKONOS image. The structure of the proposed network is developed in a sequential fashion using the group method of data handling (GMDH) algorithm. The node models, regarded as generic classifiers, are represented by fuzzy rule‐based systems, combined with a fusion scheme. A data splitting mechanism is incorporated to discriminate between correctly classified and ambiguous pixels. The classifier was tested on the wetland of international importance of Lake Koronia, Greece, and the surrounding agricultural area. To achieve higher classification accuracy, the image was decomposed into two zones: the wetland and the agricultural zones. Apart from the initial bands, additional input features were considered: textural features, intensity–hue–saturation (IHS) and tasseled cap transformation. To assess the quality of the suggested model, the SONeFMUC was compared with a maximum likelihood classifier (MLC). The experimental results show that the SONeFMUC exhibited superior performance to the MLC, providing less confusion of the dominant classes in both zones. In the wetland zone, an overall accuracy of 89.5% was attained.     

Updating land cover classification using a rule‐based decision system

This paper proposes a land cover classification methodology in agricultural contexts that provides satisfactory results with a single satellite image per year acquired during the growing period. Our approach incorporates ancillary data such as cropping history (inter‐annual crop rotations), context (altitude, soil type) and structure (parcels size and shape) to compensate for the lack of radiometric data resulting from the use of a single image. The originality of the proposed method resides in the three successive steps used: S1: per‐pixel classification of a single SPOT XS image with a restricted number of land cover classes (RL) chosen to ensure good accuracy; S2: conversion of RLs into a per‐parcel classification system using ancillary parcel boundaries; and S3: parcel allocation using exhaustive land cover classes (EL) and its refinement through the application of decision rules. The method was tested on a 120?km² area (Sousson river basin, Gers, France) where exhaustive knowledge of land cover for two successive years allowed complete validation of our method. It allocated 87% of the parcels with a 75% accuracy rate according to the exhaustive list (EL). This is a satisfactory result obtained with one SPOT XS image in a small agricultural parcel context.     

Object‐based land‐cover classification for the Phoenix metropolitan area: optimization vs. transportability

An object‐based approach was utilized in the development of two urban land‐cover classification schemes on high‐resolution (0.6 m), true‐colour aerial photography of the Phoenix metropolitan area, USA. An initial classification scheme was heavily weighted by standard nearest‐neighbour (SNN) functions generated by samples from each of the classes, which produced an enhanced accuracy (84%). A second classification was developed from the initial classification scheme in which SNN functions were transformed into a fuzzy‐rule set, creating a product transportable to different areas of the same imagery, or for land‐cover change detection with similar imagery. A comprehensive accuracy assessment revealed a slightly lower overall accuracy (79%) for the rule‐based classification. We conclude that the transportable classification scheme is satisfactory for general land‐cover analyses; yet classification accuracy can be enhanced at site‐specific venues with the incorporation of nearest‐neighbour functions using class samples.     

Object‐oriented classification for urban land cover mapping with ASTER imagery

This document demonstrates the potential of using an object‐oriented approach to map urban land cover. One objective of this work was to test the ability of the object‐oriented classification in the generation of urban land cover maps. Anotehr was to produce an updated land cover map for the city of Beijing from Advanced Spaceborne Thermal Emission and Reflecton Radiometer (ASTER) data, with an evaluation of its accuracy.     

A land‐cover classification for landslide susceptibility mapping by using feature components

Classifying original bands and/or image components may cause unsatisfactory results in fields that have heterogeneous reflectance. In such cases, the demand for accurate land‐use, land‐cover, vegetation, and forestry information may require more specific components. The components should represent peculiar information collected from several inputs for target land covers. In this study, a new technique of land‐cover classification was explored to prepare an input which increases the success of landslide susceptibility mapping in a subtropical region, Asarsuyu Catchment Area (Duzce). Land‐cover mapping is a difficult issue in this area by only carrying out field studies and aerial‐photo interpretations. Moreover, applying different classifications of Landsat Thematic Mapper bands and/or their secondary products does not produce acceptable results. For this reason, vegetation indices, soil/surface moisture indices, topographic wetness index and drainage density were calculated to produce feature representative components for the land‐cover classification process. Results obtained from the proposed technique show that feature representative components significantly improve the conventional classification accuracy from 77% to 89% and the resultant land‐cover map is such a valuable input for landslide susceptibility mapping that it increases the success of the landslide susceptibility map from 63% to 88%.     

Urban land cover multi‐level region‐based classification of VHR data by selecting relevant features

The limited spatial resolution of satellite images used to be a problem for the adequate definition of the urban environment. This problem was expected to be solved with the availability of very high spatial resolution satellite images (IKONOS, QuickBird, OrbView‐3). However, these space‐borne sensors are limited to four multi‐spectral bands and may have specific limitations as far as detailed urban area mapping is concerned. It is therefore essential to combine spectral information with other information, such as the features used in visual interpretation (e.g. the degree and kind of texture and the shape) transposed to digital analysis. In this study, a feature selection method is used to show which features are useful for particular land‐cover classes. These features are used to improve the land‐cover classification of very high spatial resolution satellite images of urban areas. The useful features are compared with a visual feature selection. The features are calculated after segmentation into regions that become analysis units and ease the feature calculation.     

Integration of multitemporal/polarization C‐band SAR data sets for land‐cover classification

This paper investigates the potential of multitemporal/polarization C‐band SAR data for land‐cover classification. Multitemporal Radarsat‐1 data with HH polarization and ENVISAT ASAR data with VV polarization acquired in the Yedang plain, Korea are used for the classification of typical five land‐cover classes in an agricultural area. The presented methodologies consist of two analytical stages: one for feature extraction and the other for classification based on the combination of features. Both a traditional SAR signal property analysis‐based approach and principal‐component analysis (PCA) are applied in the feature extraction stage. Special concerns are in the interpretation of each principal component by using principal‐component loading. The tau model applied as a decision‐level fusion methodology can provide a formal framework in which the posteriori probabilities derived from different sensor data can be combined. From the case study results, the combination of PCA‐based features showed improved classification accuracy for both Radarsat‐1 and ENVISAT ASAR data, as compared with the traditional SAR signal property analysis‐based approach. The integration of PCA‐based features based on multiple polarization (i.e. HH from Radarsat‐1, and both VV and VH from ENVISAT ASAR) and different incidence angles contributed to a significant improvement of discrimination capability for dry fields which could not be properly classified by using only Radarsat‐1 or ENVISAT ASAR data, and thus showed the best classification accuracy. The results of this case study indicate that the use of multiple polarization SAR data with a proper feature extraction stage would improve classification accuracy in multitemporal SAR data classification, although further consideration should be given to the polarization and incidence angle dependency of complex land‐cover classes through more experiments.     

The use of the Minnaert correction for land‐cover classification in mountainous terrain

Land‐cover classifications in mountainous terrain are often hampered by the topographic effect. Several strategies can be pursued to correct for this. A traditional approach is to use training areas for the same land‐cover class for different topographic positions and later merge those into one class. Other solutions involve topographic corrections, such as a Minnaert correction. In this study the classification result of the traditional training‐area approach was compared with the classification result of a Minnaert‐corrected image. In order to derive the Minnaert constants, a SPOT XS scene of the Santa Monica Mountains, USA, was divided into three visually relatively homogeneous regions. Eighty per cent of the pixels were assigned the same land cover in both classifications. Differences in classification were mainly in the section of the image that had more diverse land cover than in the more homogeneous chaparral‐covered eastern section. This supports previous findings that the Minnaert constant needs to be derived for individual land‐cover classes. The findings also suggest that after the Minnaert correction the resulting classification is comparable to the classification obtained using a more traditional approach.     

Evaluation of segment‐based gap‐filled Landsat ETM+ SLC‐off satellite data for land cover classification in southern Saskatchewan,Canada

This paper describes single‐date and multi‐date land‐cover classification accuracy results using segment‐based, gap‐filled Landsat 7 Enhanced Thematic Mapper data compared with Landsat 5 Thematic Mapper data captured one day apart. Maximum likelihood and Decision tree classification algorithms were evaluated. The same training and verification sets of ground data were used for each classification evaluation. For the comparison with the single‐date classification, an average decrease of 2.8% in the classification accuracy was obtained with the use of the gap‐filled Landsat data. Area estimates for the mid‐summer images differed, on average, from 0.6% to 1.9% for a four‐class and eight‐class classification, respectively. A multi‐date land‐cover classification was also completed with the addition of a late spring Landsat 5 image, resulting in an average decrease in classification accuracy of 1.8%.     

A method for classification accuracy evaluation for a precipitation‐oriented classifier

A method of classification accuracy evaluation for a cloud and precipitation classifier applied to geostationary meteorological satellite data is presented. The method has been developed to evaluate the accuracy of a rather precise classification algorithm. The algorithm produces nine classes, four of which involve precipitation. The classes are: (1) clear or insignificant cloud, (2) low thin cloud with no rain, (3) low or middle thin cloud with no rain, (4) low or middle thick cloud with no rain, (5) middle or high cloud with no rain, (6) middle or high cloud with the possibility of rain, (7) middle or high cloud with light–moderate precipitation, (8) middle–high cloud with moderate–heavy precipitation, (9) heavy thunderstorm. The evaluation classifier has been tested for its accuracy (ground truth) using comparison between actual meteorological weather reports and classification results derived from the algorithm applied. For the estimation of classification accuracy, the omission/commission method is applied between the observed and the classification‐produced values. The classifier used has proved to be very reliable for classifying major cloud types and precipitation, tested during the synoptic situation of depression systems approaching the south Balkan Peninsula from the west. In that synoptic situation, different intensities of rainfall as well as heavy thunderstorm were present, and the results are very satisfactory. The method can be used to evaluate classification results produced by algorithms applied to meteorological satellite data, classifying precipitation areas as well as the heaviness of precipitation.     

Fuzzy segmentation for object‐based image classification

This Letter proposes an object‐based image classification procedure which is based on fuzzy image‐regions instead of crisp image‐objects. The approach has three stages: (a) fuzzification in which fuzzy image‐regions are developed, resulting in a set of images whose digital values express the degree of membership of each pixel to target land‐cover classes; (b) feature analysis in which contextual properties of fuzzy image‐regions are quantified; and (c) defuzzification in which fuzzy image‐regions are allocated to target land‐cover classes. The proposed procedure is implemented using automated statistical techniques that require very little user interaction. The results indicate that fuzzy segmentation‐based methods produce acceptable thematic accuracy and could represent a viable alternative to current crisp image segmentation approaches.     

MERIS‐FR potential for land use–land cover mapping in Spain

Abstract

Since 1984, national and international agencies have sought to improve their ability to forecast famine in sub-Saharan Africa. A number of early warning systems have been implemented for this purpose that monitor physical and social variables that may indicate the likelihood and magnitude of famine. Several famine early warning systems use satellite remote sensing data to supplement ground-based observations. These systems have demonstrated the advantages in timeliness and consistency of remote sensing data. Although user needs have not been clearly defined, experience gained in the operation of early warning systems and the results of related research suggest that: (a) at the continental scale AVHRR GAC data offer many advantages over traditional, ground data sources;

(b) quantitative crop yield estimates might be improved through consideration of both photosynthetic activity of the vegetation and length of growing season;

(c)qualitative comparisons of crop years have provided useful inputs to current early warning needs; and (d) stratification of the region into coherent geographical areas would improve all estimates:     

A resource limited artificial immune system algorithm for supervised classification of multi/hyper‐spectral remote sensing imagery

The resource limited artificial immune system (RLAIS), a new computational intelligence approach, is being increasingly recognized as one of the most competitive methods for data clustering and analysis. Nevertheless, owing to the inherent complexity of the conventional RLAIS algorithm, its application to multi/hyper‐class remote sensing image classification has been considerably limited. This paper explores a novel artificial immune algorithm based on the resource limited principles for supervised multi/hyper‐spectral image classification. Three experiments with different types of images were performed to evaluate the performance of the proposed algorithm in comparison with other traditional image classification algorithms: parallelepiped, minimum distance, maximum likelihood, K‐nearest neighbour and back‐propagation neural network. The results show that the proposed algorithm consistently outperforms the traditional algorithms in all the experiments and hence provides an effective new option for processing multi/hyper spectral remote sensing images.     

Per‐field urban land use classification based on tax parcel boundaries

The objective of this study is to test a per‐field approach for classifying detailed urban land use, such as single‐family, multi‐family, industrial and commercial. Tax parcel boundaries are used as the field boundaries for classification. Twelve attributes of parcels, such as parcel sizes, parcel shape, building counts and building heights, are used as the discriminant factors between different land use types. For our study area that consists of 33 025 parcels, we first derived parcel attributes from geographic information system (GIS) and remote sensing data. We then converted the parcel vector data to an image of 12 bands with pixel values from parcel attributes. After that, we performed a standard supervised classification to classify the image into nine land use types. The best classification result with a decision tree classifier had an overall accuracy of 93.53% and a Kappa Coefficient of 0.7023. This study shows the feasibility of applying a per‐field approach based on tax parcel boundaries to classify detailed urban land use.     

Object‐based classification of residential land use within Accra,Ghana based on QuickBird satellite data

A segmentation and hierarchical classification approach applied to QuickBird multispectral satellite data was implemented, with the goal of delineating residential land use polygons and identifying low and high socio‐economic status of neighbourhoods within Accra, Ghana. Two types of object‐based classification strategies were tested, one based on spatial frequency characteristics of multispectral data, and the other based on proportions of Vegetation–Impervious–Soil sub‐objects. Both approaches yielded residential land‐use maps with similar overall percentage accuracy (75%) and kappa index of agreement (0.62) values, based on test objects from visual interpretation of QuickBird panchromatic imagery.     

SPOT‐4 VEGETATION multi‐temporal compositing for land cover change studies over tropical regions

Multi‐temporal compositing of SPOT‐4 VEGETATION imagery over tropical regions was tested to produce spatially coherent monthly composite images with reduced cloud contamination, for the year 2000. Monthly composite images generated from daily images (S1 product, 1‐km) encompassing different land cover types of the state of Mato Grosso, Brazil, were evaluated in terms of cloud contamination and spatial consistency. A new multi‐temporal compositing algorithm was tested which uses different criteria for vegetated and non‐vegetated or sparsely vegetated land cover types. Furthermore, a principal components transformation that rescales the noise in the image—Maximum Noise Fraction (MNF)—was applied to a multi‐temporal dataset of monthly composite images and tested as a method of additional signal‐to‐noise ratio improvement. The back‐transformed dataset using the first 12 MNF eigenimages yielded an accurate reconstruction of monthly composite images from the dry season (May to September) and enhanced spatial coherence from wet season images (October to April), as evaluated by the Moran's I index of spatial autocorrelation. This approach is useful for land cover change studies in the tropics, where it is difficult to obtain cloud‐free optical remote sensing imagery. In Mato Grosso, wet season composite images are important for monitoring agricultural crop cycles.     

Study on optimization‐based layered classification for separation of wetlands

The classification and dynamics monitoring of wetlands using remotely sensed data is a complicated, time‐consuming process involving high costs, and the accuracy varies depending on the techniques used for image processing and analysis, and the time and costs required for training, etc. This paper presents an optimization‐based layered classification method for the classification and dynamics monitoring of wetlands based on a decision procedure of multiple objectives. Four driven factors including techniques to be used, classification accuracy, and the time and cost needed for the classification, were selected as the indicators of criteria in optimization of the layered classification. This method was applied to Thematic Mapper (TM) image classification of the wetlands in Minjiang River estuary and led to the overall correct percentage of 85.13% for seven categories of the wetlands.     

Incorporating domain knowledge and spatial relationships into land cover classifications: a rule‐based approach

For some tropical regions, remote sensing of land cover yields unacceptable results, particularly as the number of land cover classes increases. This research explores the utility of incorporating domain knowledge and multiple algorithms into land cover classifications via a rule‐based algorithm for a series of satellite images. The proposed technique integrates the fundamental, knowledge‐based interpretation elements of remote sensing without sacrificing the ease and consistency of automated, algorithm‐based processing. Compared with results from a traditional maximum likelihood algorithm, classification accuracy was improved substantially for each of the six land cover classes and all three years in the image series. Use of domain knowledge proved effective in accurately classifying problematic tropical land covers, such as tropical deciduous forest and seasonal wetlands. Results also suggest that ancillary data may be most useful in the classification of historic images, where the greatest improvement was observed relative to results from maximum likelihood. The cost of incorporating contextual knowledge and extensive spatial data sets may be justified, since results from the proposed technique suggest a considerable improvement in accuracy may be achieved.