1. School of Geography, University of Nottingham, Nottingham, NG7 2RD, UK 2. Department of Computer Science, Maynooth University, Maynooth, Ireland 3. Ecosystems Services and Management Program, International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria 4. Faculty of Geo-Information Science and Earth Observation, University of Twente, Enschede, The Netherlands 5. IGN France, COGIT Laboratory, 73 Avenue de Paris, 94160 Saint-Mande, France 6. Department of Mathematics, University of Coimbra / INESC Coimbra, Coimbra, Portugal
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.
Jetz et al. (2004) have provided an interesting and thought provoking commentary on my earlier paper (Foody, 2004). Although their standpoint and interpretation is quite reasonable their commentary does not fully reflect either my own viewpoint or the content and intention of my paper. I am therefore taking this opportunity to highlight a few key issues in the hope that this will help to clarify some important features of geographically weighted regression (GWR) and its use in ecology. My earlier paper explicitly set out to explore spatial nonstationarity in the relationship between species richness and a set of determinants (e.g. see the statements made in the 'aim' section of the abstract and at the end of the introductory section). The implications of spatial nonstationarity, with particular regard to scale dependency, were then discussed. In their commentary, Jetz et al. have focused mainly upon issues connected with the strength of the relationships indicated by the R2 and some broad topics connected with the use of local and global statistical techniques. There are three points I would wish to stress in order to both emphasize the key thrusts made in my earlier paper and in response to the commentary: Jetz et al. comment that they see GWR as a supplement, and not an alternative, to standard regression. I could not agree more. GWR and standard regression are fundamentally different and aim to do different things — one does not replace the other but, consequentially, one can be more appropriate than the other depending on the application. GWR was not presented as an alternative to global regression in my earlier paper. Rather, GWR was used to do what it is appropriate for, namely to explore spatial nonstationarity in a relationship and to suggest that it would help in identifying additional variables to include in order to ensure a standard global model was appropriately specified. As such, GWR may be a useful technique to apply in the steps toward the development of an ultimate global model. The fundamental differences between global and local techniques should be recognized. In particular, it is important to note that conventional global regression fits with the view that a single parameter estimate applies over the entire region of study. With local statistics, such as GWR, the focus, however, is not on regularities but rather on differences, making it highly attractive for data exploration. The variations in parameter values observed in my earlier paper do cast doubt over the use and interpretation of standard regression analyses. As discussed in the earlier paper, the single estimate for a parameter derived from a conventional global regression may not represent conditions locally or even at any site within the study area. The value of the derived global model for descriptive and predictive purposes is therefore questionable. Although GWR is not problem-free, it does highlight that the model parameters may vary markedly, in sign and direction, while those from a global regression are constant. This is an issue noted by Jetz et al. who observe that local variation may arise in a global analysis because of missing variables or interaction terms. The key point to stress here, however, is that if the global model is not fully specified, its realism may vary over space, limiting the model's descriptive and predictive value. Again, the GWR results may be used to inform attempts to fully specify a global model. In the commentary, relatively little attention was paid to the issues of scale dependence raised in my earlier article. This issue was fundamental to the paper (see its title!) and is an issue on which standard regression yields little useful information. Jetz et al. comment that GWR does provide a framework for evaluating the effects of changing the scale of an analysis. This is an important attribute as scale dependency is often observed in ecological studies and Fig. 3 in my earlier paper highlights how spatial nonstationarity in a relationship can give rise to varying trends in scale dependence effects. Tools to explore spatial nonstationarity, such as GWR, should therefore be a welcome addition to the ecologist's armoury of techniques. Jetz et al. do raise some interesting issues, notably in relation to the spatial autocorrelation of residuals and magnitude of R2. Comparison of models using the R2 is difficult and may be more appropriately undertaken using the Akaike Information Criterion (AIC) or similar variable. Spatial autocorrelation of residuals is often observed in standard regression modelling as a function of forcing a global model when the relationship is actually spatially nonstationary. GWR offers the ability to model spatial nonstationarity and its effects directly rather than through some postanalysis interpretation of the manifestation of nonstationarity effects in terms of residuals (Fotheringham et al., 2002). Although some clustering of residuals may remain in a GWR analysis, the problem is usually reduced. However, this is an issue worthy of further thought and attention. I am grateful to Jetz et al. for their observations. Their perspective, which seems based on a standard global view, is common and appropriate when wishing to make global statements and ultimately laws. This is a perfectly valid perspective but it is not necessarily always superior to others and a local perspective has much to offer. I hope that the diversity of views, tools and motivations of researchers can be viewed positively and used together to advance understanding. I am very grateful to Walter Jetz for kindly forwarding a draft of the commentary and Stewart Fotheringham for some helpful comments on GWR.
(1988). Review of: “Satellite Remote Sensing: An Introduction”. By R. HARRIS. (London and New York: Routledge and Kegan Paul, 1987.) [Pp. 220.] Price £22·50 (Cloth), £10·95 (Paperback). International Journal of Remote Sensing: Vol. 9, No. 4, pp. 823-823.
Remote sensing has considerable potential for vegetation mapping. The model of vegetation distribution represented in an image classification, however, may not always be appropriate as the algorithms typically used give a ‘hard’ class allocation. Here the output of three classification techniques, a maximum likelihood, artificial neural network and fuzzy sets classification, are softened and shown to be able to reflect the class composition of image pixels and so be able to provide a better representation of some vegetation from remotely sensed imagery.
Mis-registration of data sets is one of the largest sources of error in many remote sensing studies. An initial contribution to this error arises through the mis-location of ground control points (GCPs) used to derive geometrical transformation equations. Here, it is proposed that a soft classification of land cover may be used to direct the estimation of GCP location. The soft classification provides an estimate of the class composition of each image pixel. The spatial distribution of a pixel's component land covers may then be modeled over the area it represents and used to reduce the error in estimating the location of a GCP that lies within this area. An example is provided in which the error in locating a set of GCPs was reduced by up to 35.7 percent when information from a soft classification was available to aid the estimation of their location at a sub-pixel scale.
No abstract is provided for this article.
Synthetic aperture radar data possess many different characteristics to conventional remotely sensed data. They therefore should not be analysed in the same fashion. However, since there is no standard technique for the classification of radar data techniques formulated for use on other data sets are often used. Since these make no allowance for the unusual properties of radar data they will not exploit fully the data's information content. A method which compensates for some of the main characteristics of the radar data and significantly increases classification accuracy is proposed in this article.
Although soft classification analyses can reduce problems such as those associated with mixed pixels that impact negatively on conventional hard classifications their accuracy is often low. One approach to increasing the accuracy of soft classifications is the use of an ensemble of classifiers, an approach which has been successful for hard classifications but rarely applied for soft classifications. Four methods for combining soft classifications to increase soft classification accuracy were assessed. These methods were based on (i) the selection of the most accurate predictions on a class‐specific basis, (ii) the average of the outputs of the individual classifications for each case, (iii) the direct combination of classifications using evidential reasoning and (iv) the adaptation of the outputs to enable the use of a conventional (hard classification) ensemble approach. These four approaches were assessed with classifications of National Oceanic and Atmospheric Administration (NOAA) Advanced Very High‐Resolution Radiometer (AVHRR) imagery of Australia. The data were classified using two neural networks and a probabilistic classifier. All four ensemble approaches applied to the outputs of these three classifiers were found to increase classification accuracy. Relative to the most accurate individual classification, the increases in overall accuracy derived ranged from 2.20% to 4.45%, increases that were statistically significant at 95% level of confidence. The results highlight that ensemble approaches may be used to significantly increase soft classification accuracy.
Fine spatial resolution remotely sensed imagery has considerable potential for mapping a shoreline. Although fine spatial resolution imagery typically allows the instantaneous shoreline to be mapped with high accuracy, interest is normally focused on a reference shoreline, defined on a stable vertical datum, which is generally not apparent in the imagery unless acquired at a time carefully coordinated with the tidal characteristics of the region. To map a tide‐coordinated shoreline, such as the mean sea level (MSL), information on terrain topography, bathymetry and tidal characteristics is required. In this study, IKONOS imagery was used to derive topographic and bathymetric information for an extract of the Malaysian coast and combined with a tide chart for the region to map the MSL. The digital elevation model (DEM) derived had a root mean square error (RMSE), calculated on independent control points, of ∼2.2 m while the bathymetric model had an RMSE of 0.87 m. The shoreline derived from the combination of the DEM, bathymetry and tidal information was mapped with an RMSE of 1.8 m. Acknowledgements We are grateful for the support of the Malaysian Centre of Remote Sensing (MACRES), especially for the provision of the data used and assistance with fieldwork, and the Malaysian Government for providing a scholarship to A.M. The research reported was undertaken while both authors were based at the University of Southampton. Finally, we thank the referees for their constructive comments on the original manuscript.
The Earth is undergoing an accelerated rate of native ecosystem conversion and degradation and there is increased interest in measuring and modelling biodiversity from space. Biogeographers have a long-standing interest in measuring patterns of species occurrence and distributional movements and an interest in modelling species distributions and patterns of diversity. Much progress has been made in identifying plant species from space using high-resolution satellites (QuickBird, IKONOS), while the measurement of species movements has become commonplace with the ARGOS satellite tracking system which has been used to track the movements of thousands of individual animals. There have been significant advances in land-cover classifications by combining data from multi-passive and active sensors, and new classification techniques. Species distribution modelling has been growing at a striking rate and the incorporation of spaceborne data on climate, topography, land cover, and vegetation structure has great potential to improve models. There have been significant advances in modelling species richness, alpha diversity, and beta diversity using multisensors to quantify land-cover classifications and landscape metrics, measures of productivity, and measures of heterogeneity. Remote sensing of nature reserves can provide natural resources managers with near real-time data within and around reserves that can be used to support conservation efforts anywhere in the world. Future research should focus on incorporating recent spaceborne sensors, more extensive integration of available spaceborne imagery, and the collection and dissemination of high-quality field data. This will improve our understanding of the distribution of life on earth.
Remote sensing is an attractive source of data for land cover mapping applications. Mapping is generally achieved through the application of a conventional statistical classification, which allocates each image pixel to a land cover class. Such approaches are inappropriate for mixed pixels, which contain two or more land cover classes, and a fuzzy classification approach is required. When pixels may have multiple and partial class membership measures of the strength of class membership may be output and, if strongly related to the land cover composition, mapped to represent such fuzzy land cover. This type of representation can be derived by softening the output of a conventional 'hard' classification or using a fuzzy classification. The accuracy of the representation provided by a fuzzy classification is, however, difficult to evaluate. Conventional measures of classification accuracy cannot be used as they are appropriate only for 'hard' classifications. The accuracy of a classification may, however, be indicated by the way in which the strength of class membership is partitioned between the classes and how closely this represents the partitioning of class membership on the ground. In this paper two measures of the closeness of the land cover representation derived from a classification to that on the ground were used to evaluate a set of fuzzy classifications. The latter were based on measures of the strength of class membership output from classifications by a discriminant analysis, artificial neural network and fuzzy c-means classifiers. The results show the importance of recognising and accommodating for the fuzziness of the land cover on the ground. The accuracy assessment methods used were applicable to pure and mixed pixels and enabled the identification of the most accurate land cover representation derived. The results showed that the fuzzy representations were more accurate than the 'hard' classifications. Moreover, the outputs derived from the artificial neural network and the fuzzy c-means algorithm in particular were strongly related to the land cover on the ground and provided the most accurate land cover representations. The ability to appropriately represent fuzzy land cover and evaluate the accuracy of the representation should facilitate the use of remote sensing as a source of land cover data.
Aim To explore the impacts of imperfect reference data on the accuracy of species distribution model predictions. The main focus is on impacts of the quality of reference data (labelling accuracy) and, to a lesser degree, data quantity (sample size) on species presence–absence modelling. Innovation The paper challenges the common assumption that some popular measures of model accuracy and model predictions are prevalence independent. It highlights how imperfect reference data may impact on a study and the actions that may be taken to address problems. Main conclusions The theoretical independence of prevalence of popular accuracy measures, such as sensitivity, specificity, true skills statistics (TSS) and area under the receiver operating characteristic curve (AUC), is unlikely to occur in practice due to reference data error; all of these measures of accuracy, together with estimates of species occurrence, showed prevalence dependency arising through the use of a non-gold-standard reference. The number of cases used also had implications for the ability of a study to meet its objectives. Means to reduce the negative effects of imperfect reference data in study design and interpretation are suggested.
Studies of land-cover change using satellite remote sensing are often constrained to depict land-cover conversions only, with the equally important modifications undetected or misrepresented, resulting in significant error. Desert fluctuations within the Sahel were examined using an approach that indicated the magnitude of land-cover changes. This showed that the conventional post-classification comparison method of change detection appeared to underestimate the area of land-cover change and, where a change was detected, typically overestimate its magnitude. At the regional scale, the land-cover changes detected were strongly related to rainfall variability. This relationship did not, however, explain changes at a finer spatial scale and indicated that dryland degradation, and its causes, may remain far from understood.
RÉSUMÉLa correction radiométrique des données du radar à antenne de synthèse (RAS) permet d'améliorer leur précision pour la cartographie de l'occupation du sol. Pour vérifier cette assertion, des données RAS, recueillies dans le cadre du projet RAS-580 de l'Agence spatiale européenne, ont été corrigées radiométriquement. Des données corrigées et non corrigées ont été utilisées pour la cartographie de l'occupation du sol dans la vallée de la Tamise au Royaume-Uni. Les corrections radiométriques ont été nécessaires afin de supprimer le déséquilibre dans la tonalité, présent sur les données du RAS-580. Ceci a permis d'accroître la précision cartographique presque du double, soit 52–73%, pour un niveau de confiance de 95%. L'auteur conclut que pour la cartographie de l'occupation du sol, il faut également tenir compte d'autres facteurs tels que l'effet de l'angle d'incidence.SUMMARYThe accuracy with which synthetic aperture radar (SAR) data can be used to map land cover is increased if the data are radio-metrically corrected. To evaluate this assertion, SAR data, collected as part of the European Space Agency's SAR-580 project, were radio-metrically corrected and both the corrected and uncorrected data were used to map land cover in the Thames Valley, UK. This correction was required to remove the tonal imbalance present in the SAR-580 data. Radiometric correction of the SAR data almost doubled the accuracy with which land cover could be mapped, to 52–73 % at the 95% confidence level. It was concluded that for land cover mapping, other factors, such as the effect of incidence angle, must also be considered.
