Peter M. Atkinson

Introduction Neural networks in remote sensing

Abstract Over the past decade there have been considerable increases in both the quantity of remotely sensed data available and the use of neural networks. These increases have largely taken place in parallel, and it is only recently that several researchers have begun to apply neural networks to remotely sensed data. This paper introduces this special issue which is concerned specifically with the use of neural networks in remote sensing. The feed-forward back-propagation multi-layer perceptron (MLP) is the type of neural network most commonly encountered in remote sensing and is used in many of the papers in this special issue. The basic structure of the MLP algorithm is described in some detail while some other types of neural network are mentioned. The most common applications of neural networks in remote sensing are considered, particularly those concerned with the classification of land and clouds, and recent developments in these areas are described. Finally, the application of neural networks to m...

Generalised linear modelling of susceptibility to landsliding in the Central Apennines, Italy

Generalised linear modelling was used to model the relation between landsliding and several independent variables (geology, dip, strike, strata-slope interaction, aspect, density of lineaments and slope angle) for a small area of the central Apennines, Italy. Raster maps of landsliding and the independent variables were produced from air photographs, topographic and geological maps, and field checking. A logistic regression was then obtained between all slope movements and the independent variables (chosen to reflect conditions prior to landsliding). Not surprisingly, geology and slope angle were found to be the most significant factors in the model. The landslides in the region were then classified into dormant and active types and further linear models were obtained for each. While geology and slope angle were again the most significant factors in each model, slope aspect and strike were less significant for active landslides. Finally, further independent variables applicable to active landslides only (vegetation cover, soil thickness, horizontal curvature, vertical curvature, concavity of slope, local relief and roughness) were added to the model for active landslides. Interestingly, with these new variables added, vegetation cover and concavity of slope were found to be more significant than geology and slope angle.

Urbanization, malaria transmission and disease burden in Africa

Many attempts have been made to quantify Africas malaria burden but none has addressed how urbanization will affect disease transmission and outcome, and therefore mortality and morbidity estimates. In 2003, 39% of Africas 850 million people lived in urban settings; by 2030, 54% of Africans are expected to do so. We present the results of a series of entomological, parasitological and behavioural meta-analyses of studies that have investigated the effect of urbanization on malaria in Africa. We describe the effect of urbanization on both the impact of malaria transmission and the concomitant improvements in access to preventative and curative measures. Using these data, we have recalculated estimates of populations at risk of malaria and the resulting mortality. We find there were 1,068,505 malaria deaths in Africa in 2000 — a modest 6.7% reduction over previous iterations. The public-health implications of these findings and revised estimates are discussed.

Geostatistical classification for remote sensing: an introduction

Traditional spectral classification of remotely sensed images applied on a pixel-by-pixel basis ignores the potentially useful spatial information between the values of proximate pixels. For some 30 years the spatial information inherent in remotely sensed images has been employed, albeit by a limited number of researchers, to enhance spectral classification. This has been achieved primarily by filtering the original imagery to (i) derive texture ‘wavebands’ for subsequent use in classification or (ii) smooth the imagery prior to (or after) classification. Recently, the variogram has been used to represent formally the spatial dependence in remotely sensed images and used in texture classification in place of simple variance filters. However, the variogram has also been employed in soil survey as a smoothing function for unsupervised classification. In this review paper, various methods of incorporating spatial information into the classification of remotely sensed images are considered. The focus of the paper is on the variogram in classification both as a measure of texture and as a guide to choice of smoothing function. In the latter case, the paper focuses on the technique developed for soil survey and considers the modification that would be necessary for the remote sensing case.

Geostatistics and remote sensing

In geostatistics, spatial autocorrelation is utilized to estimate optimally local values from data sampled elsewhere. The powerful synergy between geostatistics and remote sensing went unrealized until the 1980s. Today geostatistics are used to explore and describe spatial variation in remotely sensed and ground data; to design optimum sampling schemes for image data and ground data; and to increase the accuracy with which remotely sensed data can be used to classify land cover or estimate continuous variables. This article introduces these applications and uses two examples to highlight characteristics that are common to them all. The article concludes with a discussion of conditional simulation as a novel geostatistical technique for use in remote sensing.

Super-Resolution Land Cover Pattern Prediction Using a Hopfield Neural Network

Landscape pattern represents a key variable in management and understanding of the environment, as well as driving many environmental models. Remote sensing can be used to provide information on the spatial pattern of land cover features, but analysis and classification of such imagery suffers from the problem of class mixing within pixels. Soft classification techniques can estimate the class composition of image pixels. However, their output provides no indication of how such classes are distributed spatially within the instantaneous field-of-view (IFOV) represented by the pixel. Techniques to provide an improved spatial representation of land cover targets larger than the size of a pixel have been developed. However, the mapping of subpixel scale land cover features has yet to be investigated. We recently described the application of a Hopfield neural network technique to super-resolution mapping of land cover features larger than a pixel, using information of pixel composition determined from soft classification, and now show how our approach can be extended in a new way to predict the spatial pattern of subpixel scale features. The network converges to a minimum of an energy function defined as a goal and several constraints. Prior information on the typical spatial arrangement of the particular land cover types is incorporated into the energy function as a semivariance constraint. This produces a prediction of the spatial pattern of the land cover in question, at the subpixel scale. The technique is applied to synthetic and simulated Landsat Thematic Mapper (TM) imagery, and compared to results of an existing super-resolution target identification technique. Results show that the new approach represents a simple, robust, and efficient tool for super-resolution land cover pattern prediction from remotely sensed imagery.

Spatial Scale Problems and Geostatistical Solutions: A Review

The concept of spatial scale is fundamental to geography, as are the problems of integrating data obtained at different scales. The availability of GIS has provided an appropriate environment to re-scale data prior to subsequent integration, but few tools with which to implement the re-scaling. This sparsity of appropriate tools arises primarily because the nature of the spatial variation of interest is often poorly understood and, specifically, the patterns of spatial dependence and error are unknown. Spatial dependence can be represented and modelled using geostatistical approaches providing a basis for the subsequent re-scaling of spatial data (e.g., via spatial interpolation). Geostatistical techniques can also be used to model the effects of re-scaling data through the geostatistical operation of regularization. Regularization provides a means by which to re-scale the statistics and functions that describe the data rather than the data themselves. These topics are reviewed in this paper and the importance of the spatial scale problems that remain is emphasized.

Sub-pixel Target Mapping from Soft-classified, Remotely Sensed Imagery

A simple, efficient algorithm is presented for sub-pixel target mapping from remotely-sensed images. Following an initial random allocation of “soft” pixel proportions to “hard” subpixel binary classes, the algorithm works in a series of iterations, each of which contains three stages. For each pixel, for all sub-pixel locations, a distance-weighted function of neighboring sub-pixels is computed. Then, for each pixel, the sub-pixel representing the target class with the minimum value of the function, and the sub-pixel representing the background with the maximum value of the function are found. Third, these two sub-pixels are swapped if the swap results in an increase in spatial correlation between sub-pixels. The new algorithm predicted accurately when applied to simple simulated and real images. It represents an accessible tool that can be coded and applied readily by remote sensing investigators.

The integration of spectral and textural information using neural networks for land cover mapping in the Mediterranean

The aim of this study was to develop an efficient and accurate procedure for classifying Mediterranean land cover with remotely sensed data. Combinations of artificial neural networks (ANN) and texture analysis on a per-field basis were used to classify a Landsat Thematic Mapper image of the Cukurova Deltas, Turkey, into eight land cover classes. This study integrated spectral information with measures of texture, in the form of the variance and the variogram. The accuracy of the ANN was greater than that of maximum likelihood (ML) when using spectral data alone and when using spectral and textural data. The use of texture measures through the per-pixel and per-field majority rule approaches were found to reduce classification accuracy because the field boundaries were enlarged and so overwhelmed the measures of texture. In contrast, the per-field approach (where the field was specified prior to analysis) combined with texture information increased significantly classification accuracy. However, the accuracy decreased as the variogram lag increased. The accuracy with which land cover could be classified in this region was maximised at 89% by using a per-field, ANN approach in which semivariance at a lag of 1 pixel was incorporated as textural information. This is 15% greater than the accuracy achieved using a standard per-pixel ML classification. The primary limitation of the use of the per-field approach was noted to be the need for prior knowledge of field boundaries which may be resolved using existing data or through some form of edge-detection routine.

Fine spatial resolution simulated satellite sensor imagery for land cover mapping in the United Kingdom

This article presents a set of techniques developed to classify land cover on a per-parcel (herein termed per-field) basis by integrating fine spatial resolution simulated satellite sensor imagery with digital vector data. Classification, based on the spectral and spatial properties of the imagery, was carried out on a per-pixel basis. The resulting classified images were then integrated with vector data to classify on a per-field basis. Four tools were adopted or developed to increase the accuracy and utility of the per-field classification and a fifth was proposed. The spectral variability within agricultural fields resulted in misclassification within the per-pixel classification, and this was overcome using a per-field classification. Mixed land cover in urban areas also resulted in misclassification. A low pass smoothing filter and a “texture” filter applied to the per-pixel classified image increased the classification accuracy of this land cover prior to per-field classification. The flexibility of the integration process enabled the exploitation of spectral and spatial variation between pixels within individual parcels to produce new classes during per-field classification and to identify fields with a high likelihood of misclassification.