Mapping surface water bodies from fine spatial resolution optical remote sensing imagery is essential for the understanding of the global hydrologic cycle. Although satellite data are useful for mapping, the limited spectral information captured by some satellite systems can be suboptimal for the task. For example, the very high-resolution images of Google Earth (GE) only contain RGB bands, which often means many water bodies and land objects are confused. Sentinel-2 (S2) imagery has a spectral resolution more suitable for mapping water bodies, but its medium spatial resolution limits the ability for detailed mapping of water-land boundaries. This letter proposes a deep feature and domain knowledge fusion network (DFDKFNet) for mapping surface water bodies by fusing GE and S2 images while incorporating domain knowledge. DFDKFNet uses the remote sensing indices of normalized difference water index (NDWI) and normalized difference vegetation index (NDVI) derived from the S2 image as the representative domain knowledge to better extract water bodies from terrestrial features. A similar pixel-based approach is used to downscale the NDWI and NDVI maps to match the spatial resolution between the GE and S2 images. The DFDKFNet uses the GE and downscaled NDWI and NDVI images to extract the deep semantic features of water bodies, which are fused with the domain knowledge extracted from the NDWI and NDVI images. DFDKFNet was compared with several state-of-the-art algorithms, and the results show that DFDKFNet can enhance water body mapping accuracy.

Spatiotemporal satellite image fusion (STIF) has been widely applied in land surface monitoring to generate high spatial and high temporal reflectance images from satellite sensors. This paper proposed a new unmixing-based spatiotemporal fusion method that is composed of a self-trained random forest machine learning regression (R), low resolution (LR) endmember estimation (E), high resolution (HR) surface reflectance image reconstruction (R), and residual compensation (C), that is, RERC. RERC uses a self-trained random forest to train and predict the relationship between spectra and the corresponding class fractions. This process is flexible without any ancillary training dataset, and does not possess the limitations of linear spectral unmixing, which requires the number of endmembers to be no more than the number of spectral bands. The running time of the random forest regression is about ~1% of the running time of the linear mixture model. In addition, RERC adopts a spectral reflectance residual compensation approach to refine the fused image to make full use of the information from the LR image. RERC was assessed in the fusion of a prediction time MODIS with a Landsat image using two benchmark datasets, and was assessed in fusing images with different numbers of spectral bands by fusing a known time Landsat image (seven bands used) with a known time very-high-resolution PlanetScope image (four spectral bands). RERC was assessed in the fusion of MODIS-Landsat imagery in large areas at the national scale for the Republic of Ireland and France. The code is available at https://www.researchgate.net/proiile/Xiao_Li52.

Superresolution mapping (SRM) is a commonly used method to cope with the problem of mixed pixels when predicting the spatial distribution within low-resolution pixels. Central to the popular SRM method is the spatial pattern model, which is utilized to represent the land cover spatial distribution within mixed pixels. The use of an inappropriate spatial pattern model limits such SRM analyses. Alternative approaches, such as deep-learning-based algorithms, which learn the spatial pattern from training data through a convolutional neural network, have been shown to have considerable potential. Deep learning methods, however, are limited by issues such as the way the fraction images are utilized. Here, a novel SRM model based on a generative adversarial network (GAN), GAN-SRM, is proposed that uses an end-to-end network to address the main limitations of existing SRM methods. The potential of the proposed GAN-SRM model was assessed using four land cover subsets and compared to hard classification and several popular SRM methods. The experimental results show that of the set of methods explored, the GAN-SRM model was able to generate the most accurate high-resolution land cover maps.

Over the last decade, the beaching of massive quantities of sargassum in the Caribbean has become a major problem. Relatively few studies have focused on beached sargassum which has numerous negative environmental, social and economic impacts. To aid understanding of the problem and develop approaches to effectively manage it, a satellite-based mapping and monitoring service is being developed on the basis of information generated in a set of user needs workshops. Here, we report on the foundations of this mapping and monitoring service. In particular, user needs obtained from a large and diverse set of stakeholders are used to identify appropriate satellite remote sensing systems to use as data sources. The system will use mainly PlanetScope and Sentinel-2 data to meet the spatio-temporal demands of users. It is planned to develop an operational system that may if desired be extended to the Caribbean as a whole and beyond.

Imbalanced training sets are known to produce suboptimal maps for supervised classification. Therefore, one challenge in mapping land cover is acquiring training data that will allow classification with high overall accuracy (OA) in which each class is also mapped onto similar user's accuracy. To solve this problem, we integrated local adaptive region and box-and-whisker plot (BP) techniques into an iterative algorithm to expand the size of the training sample for selected classes in this article. The major steps of the proposed algorithm are as follows. First, a very small initial training sample (ITS) for each class set is labeled manually. Second, potential new training samples are found within an adaptive region by conducting local spectral variation analysis. Lastly, three new training samples are acquired to capture information regarding intraclass variation; these samples lie in the lower, median, and upper quartiles of BP. After adding these new training samples to the ITS, classification is retrained and the process is continued iteratively until termination. The proposed approach was applied to three very high-resolution (VHR) remote-sensing images and compared with a set of cognate methods. The comparison demonstrated that the proposed approach produced the best result in terms of OA and exhibited superiority in balancing user's accuracy. For example, the proposed approach was typically 2%-10% more accurate than the compared methods in terms of OA and it generally yielded the most balanced classification.

Due to the tradeoff between spatial and temporal resolutions commonly encountered in remote sensing, no single satellite sensor can provide fine spatial resolution land surface temperature (LST) products with frequent coverage. This situation greatly limits applications that require LST data with fine spatiotemporal resolution. Here, a deep learning-based spatiotemporal temperature fusion network (STTFN) method for the generation of fine spatiotemporal resolution LST products is proposed. In STTFN, a multiscale fusion convolutional neural network is employed to build the complex nonlinear relationship between input and output LSTs. Thus, unlike other LST spatiotemporal fusion approaches, STTFN is able to form the potentially complicated relationships through the use of training data without manually designed mathematical rules making it is more flexible and intelligent than other methods. In addition, two target fine spatial resolution LST images are predicted and then integrated by a spatiotemporal-consistency (STC)-weighting function to take advantage of STC of LST data. A set of analyses using two real LST data sets obtained from Landsat and moderate resolution imaging spectroradiometer (MODIS) were undertaken to evaluate the ability of STTFN to generate fine spatiotemporal resolution LST products. The results show that, compared with three classic fusion methods [the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), the spatiotemporal integrated temperature fusion model (STITFM), and the two-stream convolutional neural network for spatiotemporal image fusion (StfNet)], the proposed network produced the most accurate outputs [average root mean square error (RMSE)0.971].

Optical earth observation satellite sensors often provide a coarse spatial resolution (CR) multispectral (MS) image together with a fine spatial resolution (FR) panchromatic (PAN) image. Pansharpening is a technique applied to such satellite sensor images to generate an FR MS image by injecting spatial detail taken from the FR PAN image while simultaneously preserving the spectral information of MS image. Pansharpening methods are mostly applied on a per-pixel basis and use the PAN image to extract spatial detail. However, many land cover objects in FR satellite sensor images are not illustrated as independent pixels, but as many spatially aggregated pixels that contain important semantic information. In this article, an object-based pansharpening approach, termed object-based area-to-point regression kriging (OATPRK), is proposed. OATPRK aims to fuse the MS and PAN images at the object-based scale and, thus, takes advantage of both the unified spectral information within the CR MS images and the spatial detail of the FR PAN image. OATPRK is composed of three stages: image segmentation, object-based regression, and residual downscaling. Three data sets acquired from IKONOS and Worldview-2 and 11 benchmark pansharpening algorithms were used to provide a comprehensive assessment of the proposed OATPRK approach. In both the synthetic and real experiments, OATPRK produced the most superior pan-sharpened results in terms of visual and quantitative assessment. OATPRK is a new conceptual method that advances the pixel-level geostatistical pansharpening approach to the object level and provides more accurate pan-sharpened MS images.

The brick-making industry of the `Brick Belt' (spanning mostly parts of Bangladesh, Nepal, India and Pakistan) is a major employer, providing work for tens of millions people. Unfortunately, it is known via locally-based human rights groups that many of those working in the brick kilns are modern-day slaves. However, reliable and timely, spatially explicit and scalable data on the full extent of this slavery is currently unavailable. Since brick kilns leave a visible impact on the Earth's land cover that can be detected from space, and given that satellite sensor images are readily available globally for the past >40 years, using these data could afford a better understanding of the spatio-temporal dynamics of this slavery activity, affording optimal intervention. This research focuses on aging of brick kilns in the `Brick Belt' based on Landsat time series data, in order to reflect the evolution of brick kilns. Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper (ETM+) and Landsat 8 Operational Land Imager (OLI) time series data from 1984 to 2018 were downloaded from Google Earth Engine according to the coordinates of a collection of kilns identified in the `Brick Belt'. A break detection method applied to the time series of data and random forest classifications were used to age the kilns. Results showed that the accuracy of kiln aging, determined using Google Earth as ground data, was approximately 83%, and the root-mean-square-error (RMSE) between the Landsat prediction date and Google Earth date was about 3.60 years, both highlighting the potential to age kilns across the broader region. The results also show that many kilns were younger than 10 years, indicating that the brick making industry has been highly active in recent years, being driven by the demand for bricks as the region's economy grows.

A decade ago, Volunteered Geographical Information (VGI) was identified as a new source of information that would blur the traditional boundary between producers and the consumers of data (Goodchil...

Super-resolution mapping (SRM) aims to determine the spatial distribution of the land cover classes contained in the area represented by mixed pixels to obtain a more appropriate and accurate map at a finer spatial resolution than the input remotely sensed image. The image-based SRM models directly use the observed images as input and can mitigate the uncertainty caused by class fraction errors. However, existing image-based SRM models always adopt a fixed set of endmembers used in the entire image, ignoring the spatial variability and spectral uncertainty of endmembers. To address this problem, this letter proposed an optimal endmember-based SRM (OESRM) model, which considers the spatial variations in endmembers, and determines the best-fit one for each coarse resolution pixel using the spectral angle and the spectral distance as the spectral similarity indexes. A Sentinel-2A and a Landsat-8 multispectral images were used to analyze the performance of OESRM, by comparing with three other SRM methods which adopt a fixed endmember set or multiple endmember sets. The results showed that OESRM generated resultant land cover maps with more spatial detail, and reduced the confusion between land cover classes with similar spectral features. The proposed OESRM model produced the results with the highest overall accuracy in both experiments, showing its effectiveness in reducing the effect of endmember uncertainty on SRM.

The mixed pixel problem is common in remote sensing. A soft classification can generate land cover class fraction images that illustrate the areal proportions of the various land cover classes within pixels. The spatial distribution of land cover classes within each mixed pixel is, however, not represented. Super-resolution land cover mapping (SRM) is a technique to predict the spatial distribution of land cover classes within the mixed pixel using fraction images as input. Spatial–temporal SRM (STSRM) extends the basic SRM to include a temporal dimension by using a finer-spatial resolution land cover map that pre- or postdates the image acquisition time as ancillary data. Traditional STSRM methods often use one land cover map as the constraint, but neglect the majority of available land cover maps acquired at different dates and of the same scene in reconstructing a full state trajectory of land cover changes when applying STSRM to time-series data. In addition, the STSRM methods define the temporal dependence globally, and neglect the spatial variation of land cover temporal dependence intensity within images. A novel local STSRM (LSTSRM) is proposed in this paper. LSTSRM incorporates more than one available land cover map to constrain the solution, and develops a local temporal dependence model, in which the temporal dependence intensity may vary spatially. The results show that LSTSRM can eliminate speckle-like artifacts and reconstruct the spatial patterns of land cover patches in the resulting maps, and increase the overall accuracy compared with other STSRM methods.

The development of remote sensing has enabled the acquisition of information on land-cover change at different spatial scales. However, a tradeoff between spatial and temporal resolutions normally exists. Fine-spatial-resolution images have low temporal resolutions, whereas coarse-spatial-resolution images have high temporal repetition rates. A novel superresolution change detection method (SRCD) is proposed to detect land-cover changes at both fine spatial and temporal resolutions with the use of a coarse-resolution image and a fine-resolution land-cover map acquired at different times. SRCD is an iterative method that involves endmember estimation, spectral unmixing, land-cover fraction change detection, and superresolution land-cover mapping. Both the land-cover change/no-change map and from-to change map at fine spatial resolution can be generated by SRCD. In this paper, SRCD was applied to a synthetic multispectral image, a Moderate-Resolution Imaging Spectroradiometer multispectral image, and a Landsat-8 Operational Land Imager multispectral image. The land-cover from-to change maps are found to have the highest overall accuracy (higher than 85%) in all of the three experiments. Most of the changed land-cover patches, which were larger than the coarse-resolution pixel, were correctly detected.

Superresolution mapping (SRM) is a method to produce a fine-spatial-resolution land cover map from coarse-spatial-resolution remotely sensed imagery. A popular approach for SRM is a two-step algorithm, which first increases the spatial resolution of coarse fraction images by interpolation and then determines class labels of fine-resolution pixels using the maximum a posteriori (MAP) principle. By constructing a new image formation process that establishes the relationship between the observed coarse-resolution fraction images and the latent fine-resolution land cover map, it is found that the MAP principle only matches with area-to-point interpolation algorithms and should be replaced by deconvolution if an area-to-area interpolation algorithm is to be applied. A novel iterative interpolation deconvolution (IID) SRM algorithm is proposed. The IID algorithm first interpolates coarse-resolution fraction images with an area-to-area interpolation algorithm and produces an initial fine-resolution land cover map by deconvolution. The fine-spatial-resolution land cover map is then updated by reconvolution, back-projection, and deconvolution iteratively until the final result is produced. The IID algorithm was evaluated with simulated shapes, simulated multispectral images, and degraded Landsat images, including comparison against three widely used SRM algorithms: pixel swapping, bilinear interpolation, and Hopfield neural network. Results show that the IID algorithm can reduce the impact of fraction errors and can preserve the patch continuity and the patch boundary smoothness simultaneously. Moreover, the IID algorithm produced fine-resolution land cover maps with higher accuracies than those produced by other SRM algorithms.

Superresolution mapping (SRM) is a technique for generating a fine-spatial-resolution land cover map from coarse-spatial-resolution fraction images estimated by soft classification. The prior model used to describe the fine-spatial-resolution land cover pattern is a key issue in SRM. Here, a novel learning-based SRM algorithm, whose prior model is learned from other available fine-spatial-resolution land cover maps, is proposed. The approach is based on the assumption that the spatial arrangement of the land cover components for mixed pixel patches with similar fractions is often similar. The proposed SRM algorithm produces a learning database that includes a large number of patch pairs for which there is a fine- and coarse-spatial-resolution representation for the same area. From the learning database, patch pairs that have similar coarse-spatial-resolution patches as those in the input fraction images are selected. Fine-spatial-resolution patches in these selected patch pairs are then used to estimate the latent fine-spatial-resolution land cover map by solving an optimization problem. The approach is illustrated by comparison against state-of-the-art SRM methods using land cover map subsets generated from the USA's National Land Cover Database. Results show that the proposed SRM algorithm better maintains the spatial pattern of land covers for a range of different landscapes. The proposed SRM algorithm has the highest overall accuracy and kappa values in all of these SRM algorithms, by using the entire maps in the accuracy assessment.

The quality of the training data used in a supervised image classification can impact on the accuracy of the resulting thematic map obtained. Here the effects of mis-labeled training cases on the accuracy of classifications by discriminant analysis and a support vector machine were explored. The accuracy of both classifiers varied with the amount and nature of mis-labeled training cases. In particular, the SVM, which has been claimed to be relatively insensitive to training data error, showed the greatest sensitivity with overall accuracy declining by 8% with the use of a training set containing 20% mis-labeled cases; the difference in accuracy from that obtained without mis-labeled cases was statistically significant at the 95% level of confidence. Training data quality needs consideration when undertaking a supervised classification and should be considered in the selection of a classifier as the effects will be classifier-specific.

Remote sensing has much to gain from citizen sensing. This is particularly evident in relation to the provision of ground reference data for use in the training and testing stages of supervised image classification analyses used to generate thematic maps from remotely sensed data. Citizens are able to provide data over large geographical areas inexpensively, addressing potential problems connected with ground data samples and authoritative good practices. The great potential of citizen sensing is, however, constrained by concerns, notably with the quality of the data generated. This paper provides an overview of some of the key issues in citizen sensing to support thematic mapping from remote sensing. It highlights especially some of the ways that citizen sensing can aid remote sensing studies as a source of ground reference data.

Volunteered Geographic Information (VGI) represents a growing source of potentially valuable data for many applications, including land cover map validation. It is still an emerging field and many different approaches can be used to take value from VGI, but also many pros and cons are related to its use. Therefore, since it is timely to get an overview of the subject, the aim of this article is to review the use of VGI as reference data for land cover map validation. The main platforms and types of VGI that are used and that are potentially useful are analysed. Since quality is a fundamental issue in map validation, the quality procedures used by the platforms that collect VGI to increase and control data quality are reviewed and a framework for addressing VGI quality assessment is proposed. A review of cases where VGI was used as an additional data source to assist in map validation is made, as well as cases where only VGI was used, indicating the procedures used to assess VGI quality and fitness for use. A discussion and some conclusions are drawn on best practices, future potential and the challenges of the use of VGI for land cover map validation.

Post-disaster building damage maps are an important component of the disaster response chain and may be derived from remotely sensed data. The usefulness of the maps is a function of their accuracy and hence information on map accuracy is desirable. The assessment of building damage map accuracy is, however, a challenging task as high quality ground reference data are usually scarce or absent. Here, binary and ordinal level latent class analyses were used to evaluate the accuracy of five maps produced after the 2010 earthquake in Haiti. The quality of the estimates derived from the analyses could be evaluated in this case as a ground data set was available. The results showed that the latent class analyses were able to yield an accurate assessment of the relative accuracy of the maps, allowing maps to be ranked in order of quality. This feature may help disaster relief activities by ensuring the highest quality data are used.

Accuracy assessment should be a fundamental part of a land cover mapping programme but often constrained by the lack of ground data. Here, two sources of volunteered data are used to illustrate the potential of neogeographical activity in map validation. Ground based photographs acquired by an internet-based collaborative project and interpreted by a set of volunteers provided the reference data to support evaluation of the Globcover map's representation of tropical forest in West Africa. The results highlight concerns with volunteered data, notably the low levels of agreement but also show that imperfect data may be used to derive useful information on map properties and accuracy.

A contour-based pixel swapping method for super-resolution mapping, which combines contouring and pixel swapping super-resolution mapping approaches, that seeks to exploit the positive features of contouring and pixel swapping to produce a method that is more accurate than each alone is proposed. The accuracy of super-resolution mapping with the individual and combined techniques is explored. When combined, the error with which objects of varying shape were represented was typically greatly reduced relative to that observed from the application of the methods individually. For example, the root mean square error in mapping the boundary of an aeroplane represented in relatively fine spatial resolution imagery decreased from 14.43m with contouring and 2.95m with pixel swapping to 2.18m when the approaches were combined.

This letter evaluates three Envisat Medium Resolution Imaging Spectrometer Terrestrial Chlorophyll Index (MTCI)-based models for the estimation of terrestrial gross primary productivity (GPP) across a range of vegetation types. Correlations between flux tower measures of GPP and models for years between 2003 and 2007 were established for 30 sites across USA, Canada, and Brazil. Correlations were seen to range from very strong to weak, depending on seasonal variation in photosynthetic capacity (which is influenced by chlorophyll content) exhibited by the vegetation at each site. At least one of the three models obtained a statistically significant relationship with GPP at every site. Results indicate that chlorophyll content (as measured by the MTCI) is a most relevant community property for estimating primary productivity and chlorophyll-related vegetation indexes provide favorable approximations of the GPP of terrestrial vegetation. The inclusion of radiation information (photosynthetically active radiation (PAR) and fraction of photosynthetically active radiation (fPAR)) into the models extended the applicability of the models and the accuracy of the GPP estimate. Although further investigation is required to fully understand the applicability of these models and their parameters, these results point to the possibility of a total remote sensing approach to GPP estimation.

Accuracy assessment should be a fundamental component of an image classification analysis and is typically undertaken following either a non-site- or a site-specific methodology. The assessment of classification accuracy is, however, often difficult, with many challenges associated with the ground data typically required. Using a series of classifications of two test sites, this paper shows that accuracy assessment from both perspectives is possible through the use of a latent class modeling approach in the absence of ground data. This is possible because the parameters of a latent class model that explains the observed associations in class labeling made by a series of classifications provide estimates of class cover and conditional probabilities of class membership that equate to popular non-site- and site-specific (producer's accuracy) measures of accuracy, respectively. Additionally, the latent class model provides a new classification that could be evaluated by traditional means if ground data are available. The classification of each test site derived from the latent class model was accurate, being of equivalent accuracy to a conventional ensemble classification that was based on the same series of classifications for a site. The ability to derive a highly accurate classification and yield estimates of classification accuracy without ground data to form a testing set indicates the considerable promise of the method and a means to reduce demands for costly ground data that may also be a source of error due to imperfections.

This paper explores the potential of using the Medium Resolution Medium Spectrometer (MERIS) Terrestrial Chlorophyll Index (MTCI) for estimating gross primary productivity across a range of vegetation cover types. Correlations between flux tower measures of GPP and corresponding MTCI for years between 2003 and 2007 were established for 30 sites across the Americas. Correlations were seen to range from very strong to weak depending on variation in chlorophyll content exhibited. Moreover, comparison of the performance of the MTCI against the MOD17 GPP product at selected sites indicated that overall, the correlation between flux tower and MOD17 GPP estimates were not as strong as obtained when using the MTCI. These results demonstrate the potential of this approach and should be developed further.

This paper presents a super-resolution mapping technique as a means to gain accurate information of land cover, and especially its spatial pattern, at a sub-pixel scale. This technique extends the application of an established Hopfield neural network of super-resolution mapping technique by providing its input with a fusion of a time series coarse spatial but fine temporal resolution images. To illustrate this technique, a series of daily MODIS 250m images was acquired and fused. Using a Landsat ETM+ 30m image as ground data, results demonstrated that a Hopfield network that uses time series information produces significantly more accurate representation of land cover mapping in terms of thematic accuracy and spatial pattern prediction than by using a single image into Hopfield network or into hard classification techniques.

Support vector machines (SVM) are attractive for the classification of remotely sensed data with some claims that the method is insensitive to the dimensionality of the data and, therefore, does not require a dimensionality-reduction analysis in preprocessing. Here, a series of classification analyses with two hyperspectral sensor data sets reveals that the accuracy of a classification by an SVM does vary as a function of the number of features used. Critically, it is shown that the accuracy of a classification may decline significantly (at 0.05 level of statistical significance) with the addition of features, particularly if a small training sample is used. This highlights a dependence of the accuracy of classification by an SVM on the dimensionality of the data and, therefore, the potential value of undertaking a feature-selection analysis prior to classification. Additionally, it is demonstrated that, even when a large training sample is available, feature selection may still be useful. For example, the accuracy derived from the use of a small number of features may be noninferior (at 0.05 level of significance) to that derived from the use of a larger feature set providing potential advantages in relation to issues such as data storage and computational processing costs. Feature selection may, therefore, be a valuable analysis to include in preprocessing operations for classification by an SVM.

Ground reference data error is a major source of bias in the estimation of land cover change and of change detection accuracy. This paper explores the magnitude and direction of some of the main biases introduced into studies of land cover dynamics by remote sensing based upon the popular binary change detection error matrix. It is shown that substantial mis-estimation may arise through the use of imperfect reference data, even if of a very high accuracy. The magnitude and direction of the bias is a function of the size and nature of the errors. One key issue, however, is that it may sometimes be possible to reduce the effects of imperfect reference data and so derive more accurate estimates. These various issues are discussed with the use of simulated data to control for other potential error sources.

Support vector machines (SVMs) have considerable potential for supervised classification analyses, but their binary nature has been a constraint on their use in remote sensing. This typically requires a multiclass analysis be broken down into a series of binary classifications, following either the one-against-one or one-against-all strategies. However, the binary SVM can be extended for a one-shot multiclass classification needing a single optimization operation. Here, an approach for one-shot multi- class classification of multispectral data was evaluated against approaches based on binary SVM for a set of five-class classifications. The one-shot multiclass classification was more accurate (92.00%) than the approaches based on a series of binary classifications (89.22% and 91.33%). Additionally, the one-shot multi- class SVM had other advantages relative to the binary SVM-based approaches, notably the need to be optimized only once for the parameters C and 7 as opposed to five times for one-against-all and ten times for the one-against-one approach, respectively, and used fewer support vectors, 215 as compared to 243 and 246 for the binary based approaches. Similar trends were also apparent in results of analyses of a data set of larger dimensionality. It was also apparent that the conventional one-against-all strategy could not be guaranteed to yield a complete confusion matrix that can greatly limit the assessment and later use of a classification derived by that method.

The impacts of intra-class spectral variation on the use of soft classification outputs for super-resolution mapping was assessed. The accuracy of soft classification and super- resolution mapping was negatively related to the degree of intra-class spectral variation present in the data set. The provision of a distribution of possible sub-pixel fractional covers from a soft classification may reflect the impacts of intra-class variation and help to enhance super-resolution mapping. A possible approach to reduce the impacts of intra-class spectral variation was investigated. This was based on an approach that reduces the degree of intra-class spectral variation by defining spectral subclasses for use in the soft classification. The use of this approach increased the accuracy of soft classification predictions from r = 0.87 to r = 0.94 and decreased the RMSE in super-resolution mapping of an inter-class boundary from 44.7 m to 37.2 m. The results highlighted that reducing intra-class spectral variation may be used to increase the accuracy of soft classification and super-resolution mapping.

Remote sensing is a major source of land-cover information. Commonly, interest focuses on a single land-cover class. Although a conventional multiclass classifier may be used to provide a map depicting the class of interest, the analysis is not focused on that class and may be suboptimal in terms of the accuracy of its classification. With a conventional classifier, considerable effort is directed on the classes that are not of interest. Here, it is suggested that a one-class-classification approach could be appropriate when interest focuses on a specific class. This is illustrated with the classification of fenland, a habitat of considerable conservation value, from Landsat Enhanced Thematic Mapper Plus imagery. A range of one-class classifiers is evaluated, but attention focuses on the support-vector data description (SVDD). The SVDD was used to classify fenland with an accuracy of 97.5% and 93.6% from the user's and producer's perspectives, respectively. This classification was trained upon only the fenland class and was substantially more accurate in fen classification than a conventional multiclass maximum-likelihood classification provided with the same amount of training data, which classified fen with an accuracy of 90.0% and 72.0% from the user's and producer's perspectives, respectively. The results highlight the ability to classify a single class using only training data for that class. With a one-class classification, the analysis focuses tightly on the class of interest, with resources and effort not directed on other classes, and there are opportunities to derive highly accurate classifications from small training sets

Soft classification accuracy was negatively related to the degree of intra-class variation. Moreover, the utility of a single value prediction of the fractional cover of a class derived from a soft classification declined as level of intra-class variation increased and it may be preferable instead to derive a distribution of possible fractional covers. The distributional output may also help in the evaluation of change derived from the use of a post-classification comparison analysis.

Three possible methods of combining soft classification outputs to increase soft classification accuracy were assessed. These methods were (i) an approach that selects the most accurate predictions on a class-specific basis, (ii) Dempster-Shafer theory of evidence and (iii) an approach which degrades the soft classification output into a set of ordered classes and then combines these through the use of a conventional ensemble approach. The potential of these approaches was assessed using coarse spatial resolution NOAA AVHRR imagery of Australia. The data were classified using two neural networks (a multi-layer perceptron and a radial basis function network) as well as a probabilistic classifier. All three approaches to combine the classifications were applied to combine the soft classification outputs and had been shown to increase classification accuracy. Relative to the most accurate individual classification, the increases in overall accuracy derived ranged from 2.73 to 4.45% and large increases in individual class accuracy were also observed. The results highlight that ensemble based approaches may be used to increase soft classification accuracy.

The accuracy of supervised classification is dependent to a large extent on the input training data. In general, the analyst aims to capture a large training set to fully describe the classes spectrally with the conventional statistical classifier in mind. However, it is not always necessary to provide a complete description of the classes if using a support vector machine (SVM) as the classifier. A key attraction of the SVM based approach to classification is that it seeks to fit an optimal hyperplane between the classes and since it uses only the training samples that lie at the edge of the class distributions in feature space (support vectors) it may require only a small training sample. The paper shows the potential of SVM of using only a fraction of the training data (support vectors) collected by the usual random scheme for a study carried in the south western part of Punjab state of India

Support vector machines (SVMs) have considerable potential as classifiers of remotely sensed data. A constraint on their application in remote sensing has been their binary nature, requiring multiclass classifications to be based upon a large number of binary analyses. Here, an approach for multiclass classification of airborne sensor data by a single SVM analysis is evaluated against a series of classifiers that are widely used in remote sensing, with particular regard to the effect of training set size on classification accuracy. In addition to the SVM, the same datasets were classified using a discriminant analysis, decision tree, and multilayer perceptron neural network. The accuracy statements of the classifications derived from the different classifiers were compared in a statistically rigorous fashion that accommodated for the related nature of the samples used in the analyses. For each classification technique, accuracy was positively related with the size of the training set. In general, the most accurate classifications were derived from the SVM approach, and with the largest training set the SVM classification was significantly (p90% correct, the classifiers differed in the ability to correctly label individual cases and so may be suitable candidates for an ensemble-based approach to classification.

The relationship between middle infrared reflectance and various scenarios of preceeding rainfall for a range of tropical forest types is investigated. Statistically significant correlations for the relationship were generally observed when the rainfall data were acquired over a period of about a month with a short time lag before image acquisition.

Methods for mapping the shoreline at a sub-pixel level are evaluated. The most accurate predictions of shoreline location were made from an approach based on simulated annealing applied to the output of a soft classification (RMSE=2.25 m).

Two types of neural network were used to derive measures of biodiversity from Landsat TM data of a tropical rainforest. A feedforward neural network was used to estimate species richness while a Kohonen neural network was used to provide information on species composition. The results indicate the potential of remote sensing as a source of maps of biodiversity.

Remote sensing is an important source of land cover data required by many GIS users. Land cover data are typically derived from remotely–sensed data through the application of a conventional statistical classification. Such classification techniques are not, however, always appropriate, particularly as they may make untenable assumptions about the data and their output is hard, comprising only the code of the most likely class of membership. Whilst some deviation from the assumptions may be tolerated and a fuzzy output may be derived, making more information on class membership properties available, alternative classification procedures are sometimes required. Artificial neural networks are an attractive alternative to the statistical classifiers and here one is used to derive a fuzzy classification output from a remotely–sensed data set that may be post–processed with ancillary data available in a GIS to increase the accuracy with which land cover may be mapped. With the aid ancillary information on soil type and prior knowledge of class occurrence the accuracy of an artificial neural network classification was increased by 29–93 to 77–37 per cent. An artificial neural network can therefore be used generate a fuzzy classification output that may be used with other data sets in a GIS, which may not have been available to the producer of the classification, to increase the accuracy with which land cover may be classified.