Elsevier

Computers & Geosciences

Volume 33, Issue 10, October 2007, Pages 1261-1272
Computers & Geosciences

A linearised pixel-swapping method for mapping rural linear land cover features from fine spatial resolution remotely sensed imagery

https://doi.org/10.1016/j.cageo.2007.05.010Get rights and content

Abstract

Accurate maps of rural linear land cover features, such as paths and hedgerows, would be useful to ecologists, conservation managers and land planning agencies. Such information might be used in a variety of applications (e.g., ecological, conservation and land management applications). Based on the phenomenon of spatial dependence, sub-pixel mapping techniques can be used to increase the spatial resolution of land cover maps produced from satellite sensor imagery and map such features with increased accuracy. Aerial photography with a spatial resolution of 0.25 m was acquired of the Christchurch area of Dorset, UK. The imagery was hard classified using a simple Mahalanobis distance classifier and the classification degraded to simulate land cover proportion images with spatial resolutions of 2.5 and 5 m. A simple pixel-swapping algorithm was then applied to each of the proportion images. Sub-pixels within pixels were swapped iteratively until the spatial correlation between neighbouring sub-pixels for the entire image was maximised. Visual inspection of the super-resolved output showed that prediction of the position and dimensions of hedgerows was comparable with the original imagery. The maps displayed an accuracy of 87%. To enhance the prediction of linear features within the super-resolved output, an anisotropic modelling component was added. The direction of the largest sums of proportions was calculated within a moving window at the pixel level. The orthogonal sum of proportions was used in estimating the anisotropy ratio. The direction and anisotropy ratio were then used to modify the pixel-swapping algorithm so as to increase the likelihood of creating linear features in the output map. The new linear pixel-swapping method led to an increase in the accuracy of mapping fine linear features of approximately 5% compared with the conventional pixel-swapping method.

Introduction

Accurate mapping of linear land cover features is a well-researched task within the remote sensing literature (Quackenbush, 2004). A range of techniques for linear feature extraction exist, the majority of which focus on the extraction of road and rail networks (Priestnall et al., 2004; Song and Civco, 2004).

Rural linear features, such as hedgerows, are an important component of rural biodiversity. Hedgerows, in particular, support a rich diversity of plants, insects, birds and mammals and form an integral part of the UK Biodiversity Action Plan (UKBAP1), cited as a “priority habitat”. Over the last 40 years, up to half of the UK's hedgerows have been removed (McCollin, 2002), which is known to have had a profound effect on many different components of the UK's rural biodiversity (Robinson and Sutherland, 2002). Accordingly, accurate maps of these rural linear features would be of great utility to the organisations involved in the protection of UK hedgerows.

Remotely sensed imagery is a common source of data for creating land cover maps, and, in recent years, the accuracy with which land cover can be predicted from remotely sensed imagery has increased (Foody, 2002). This is largely due to focus on a common source of error in classified imagery—mixed pixels. Hard classification algorithms, such as the maximum likelihood classifier or the Mahalanobis distance classifier, assign individual pixels to single classes (Richards and Jia, 1999). In many cases, for example, where the frequency of spatial variation in land cover is greater than the frequency of sampling afforded by the sensor's spatial resolution (Woodcock and Strahler, 1987), pixels contain more than one land cover class. In these cases, allocating a pixel to a single land cover class will often not provide an accurate representation of land cover (Foody, 1996). This is important when considering features that are fine relative to the spatial resolution of the available imagery, such as hedgerows and pathways, which are often 1–5 m in width (Baudry et al., 2000; Baudry and Bunce, 2001), compared with, for example, Système Pour L’Observation de la Terre (SPOT) panchromatic imagery with a spatial resolution of 10 m.

Soft classification techniques, such as the fuzzy c-means (Bezdek et al., 1984) or the linear mixture model (Settle and Drake, 1993), offer a solution to the problem of mixed pixels. Pixels are ‘unmixed’ into individual component classes, using, for example, knowledge of spectral endmembers (pure pixels) or class means within a predefined set of available classes, and consequently pixels can be allocated to more than one class. However, the information provided is only the set of proportions of individual classes within a pixel: no information is provided on the spatial location of these classes within pixels.

Sub-pixel mapping techniques are one possible method of predicting the location of class proportions within pixels. By utilising the phenomenon of spatial dependence (the tendency for proximate observations to be more alike than those further apart; (Curran and Atkinson, 1998), and by decomposing pixels into smaller units (sub-pixels), sub-pixel mapping techniques can predict the locations of individual classes within pixels. This is achieved by adjusting the spatial location of the sub-pixels on the basis of maximising spatial dependence between sub-pixels within pixels. Atkinson, 2001, Atkinson, 2005, Verhoeye and De Wulf (2002) and Zhan et al. (2002) used these assumptions on simulated imagery and SPOT high resolution visible imagery to produce more accurate results than traditional hard classification. Zhan et al. (2002) reported that the technique was undeveloped for use in the “linear sub-pixel” case (Fisher, 1997), in which linear objects are less than two pixels wide. This highlighted the need for a suitable sub-pixel mapping technique where the features of interest are linear and small relative to the spatial resolution of the remotely sensed imagery.

Rural land cover features, such as pathways and hedgerows, exhibit many of the geometric characteristics normally associated with larger linear features such as roads. However, the techniques developed for extracting large linear features may not be suitable for fine linear features, for several reasons, including:

  • (1)

    Width (and size)

    Rural linear land cover features, such as hedgerows, tend to be between 1 and 5 m in width (Baudry and Bunce, 2001), as opposed to roads that tend to be greater than 10 m in width (Quackenbush, 2004). The width of a linear feature is an important aspect of the feature's character. For example, in the absence of class information, knowledge of the width of the feature can be used to increase the accuracy with which feature type can be predicted. For instance, ecologists might use the width of the feature (in addition to the length and height) to predict, for example, the species of vegetation within the hedgerow or the suitability of the hedgerow as a habitat for different species of fauna (Gillings and Fuller, 1998).

  • (2)

    Curvature

    Rural linear features, such as hedgerows, often exhibit directional variation along their length, for example, where they follow natural boundaries, such as streams and rivers, where they follow human-made linear land cover objects, such as minor roads, or where they have been planted to serve a specific purpose, such as separators between agricultural fields. The frequency of variation in curvature is diverse, from minor changes in direction (for example, where a feature has been modified to suit modern farming practices (McCollin, 2002) to major changes in direction (for example, where a hedgerow curves around the corners of a field or deviates from a straight line, where it follows the natural course of a stream or river). Curvature is an intrinsic component of a feature's character, but is often not represented accurately in land cover maps created by traditional classification techniques.

  • (3)

    Embedded objects

    Trees and other land cover objects are often found within rural linear features such as hedgerows, which, in addition to breakpoints and localised changes in size (e.g., increased growth in small sections due to spatial variation in soil type, proximity to a water source or the use of fertilizers in adjacent fields), alter the width of the feature at points along its length. Some linear feature extraction techniques predict linear features as a set of parallel edges. In some cases, these techniques might not be suitable for the prediction of hedgerows where, for example, the location of a tree within the hedgerow alters the position of one of the edges of the feature, or where other changes in width remove the parallelism of the edges. Additionally, these techniques are unlikely to predict accurately the location of land cover objects found within linear features.

The three factors noted above represent key considerations in the development of techniques for the accurate extraction of rural linear features from remotely sensed imagery.

The objective of this paper was to develop and evaluate a new sub-pixel mapping technique to super-resolve more accurately linear rural land cover objects, such as hedgerows, from fine spatial resolution remotely sensed imagery. The new method involves ‘linearising’ an existing pixel-swapping algorithm.

Section snippets

Pixel swapping

The pixel-swapping algorithm used here for sub-pixel mapping is simple and efficient. It was first presented by Atkinson (2001), where it was demonstrated using simulated imagery, and by Thornton et al. (2006), where it was applied to Quickbird satellite sensor imagery.

Land cover class proportions for each pixel, obtained from a classification, are input to the pixel-swapping algorithm. In each pixel, a fixed number of sub-pixels are created, based on a selected zoom factor, the number of

Fieldsite and data

Aerial photography at a spatial resolution of 0.25 m of Christchurch, Dorset, UK, and its surroundings was acquired. A small fieldsite (400×400 pixels) was selected from the aerial photography, which contained a simple linear feature: a hedgerow combined with four trees, one distinct from the hedgerow and three embedded within it (Fig. 4a). The imagery was hard classified into four unique classes (hedgerow, trees, cereal and shadow) using the Mahalanobis distance classifier, and the shadow class

Analysis

The pixel-swapping techniques were applied to the input imagery. In each of the 2.5 and 5 m data sets, sub-pixels were allocated randomly (Figs. 5c and d) and each of the three versions of the pixel-swapping algorithm (original, linear 3×3, linear 5×5) were applied until the algorithm converged to a solution. A zoom factor of 10 was used on the 2.5 m spatial resolution data and a zoom factor of 20 was used on the 5 m spatial resolution data, resulting in super-resolved output images at the same

Discussion

The super-resolved output from each of the three techniques described above maps the predicted sub-pixel locations of land cover classes with reasonable accuracy. In the absence of fine spatial resolution remotely sensed imagery, the sub-pixel map, in most cases, will be of greater utility than a standard hard classification, particularly in the case of mapping features that are fine relative to their surroundings.

The methods of accuracy assessment used in this study were effective at providing

Conclusion

The pixel-swapping technique is an efficient method for mapping land cover objects at the sub-pixel scale. The super-resolved output maps the predicted locations of land cover classes, particularly fine rural features such as hedgerows and trees, with greater accuracy than a conventional hard classification.

Development of the technique has increased the ability to map linear features from coarse spatial resolution remotely sensed imagery than was previously possible, with up to 5% increases in

References (22)

  • S. Gillings et al.

    Changes in bird populations on sample lowland English farms in relation to loss of hedgerows and other non-crop habitats

    Oecologia

    (1998)
  • Cited by (78)

    • Principles and methods of scaling geospatial Earth science data

      2019, Earth-Science Reviews
      Citation Excerpt :

      Moreover, neither spatial autocorrelation-based nor spatial pattern-based SRM methods effectively preserve the structure of linear patterns. Anisotropic models (Thornton et al., 2007), linear subpixel mapping agents (Xu et al., 2014), central line control (Ai et al., 2014), and linear templates (Ge et al., 2016a) have been developed to enhance the mapping of linear features. However, as for spatial autocorrelation-based SRM, the spatial distribution pattern-based SRM methods are prior knowledge approaches that require specific models to characterize the spatial distribution pattern or the spatial information of LCLU features.

    • Subpixel Mapping of Hyperspectral Image Based on Multiscale and Multifeature

      2023, IEEE Transactions on Geoscience and Remote Sensing
    • Unsupervised Bayesian Subpixel Mapping Autoencoder Network for Hyperspectral Images

      2023, IEEE Transactions on Geoscience and Remote Sensing
    • Gradient Prior Dilated Convolution Network for Remote Sensing Image Super-Resolution

      2023, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
    View all citing articles on Scopus
    View full text