Land cover mapping of large areas using chain classification of neighboring Landsat satellite images

https://doi.org/10.1016/j.rse.2009.01.010Get rights and content

Abstract

Satellite imagery is the major data source for regional to global land cover maps. However, land cover mapping of large areas with medium-resolution imagery is costly and often constrained by the lack of good training and validation data. Our goal was to overcome these limitations, and to test chain classifications, i.e., the classification of Landsat images based on the information in the overlapping areas of neighboring scenes. The basic idea was to classify one Landsat scene first where good ground truth data is available, and then to classify the neighboring Landsat scene using the land cover classification of the first scene in the overlap area as training data. We tested chain classification for a forest/non-forest classification in the Carpathian Mountains on one horizontal chain of six Landsat scenes, and two vertical chains of two Landsat scenes each. We collected extensive training data from Quickbird imagery for classifying radiometrically uncorrected data with Support Vector Machines (SVMs). The SVMs classified 8 scenes with overall accuracies between 92.1% and 98.9% (average of 96.3%). Accuracy loss when automatically classifying neighboring scenes with chain classification was 1.9% on average. Even a chain of six images resulted only in an accuracy loss of 5.1% for the last image compared to a reference classification from independent training data for the last image. Chain classification thus performed well, but we note that chain classification can only be applied when land cover classes are well represented in the overlap area of neighboring Landsat scenes. As long as this constraint is met though, chain classification is a powerful approach for large area land cover classifications, especially in areas of varying training data availability.

Introduction

Large area land cover maps derived from satellite images play a key role in global, regional and national land cover and land use assessments, carried out for example by the United Nations (UN), the Food and Agricultural Organization (FAO), or the United States Geological Survey (USGS) (Cihlar, 2000, Franklin and Wulder, 2002, Homer et al., 2004, Vogelmann et al., 2004). Such classifications allow assessments of broad-scale forest fragmentation (Riitters et al., 2002), carbon sequestration potential (Cruickshank et al., 2000, Niu and Duiker, 2006), or the Wildland Urban Interface (Radeloff et al., 2005). Therefore, large area land cover classifications present a basic prerequisite for many scientific applications (Wulder et al., 2008).

Landsat satellite data is the most widely used data type for land cover mapping because of its 35-year data record and its relatively high spatial resolution (Cohen and Goward, 2004, Wulder et al., 2008). Landsat data will become even more valuable as the Landsat Data Continuity Mission (NASA [National Aeronautic Space Administration], 2008, Wulder et al., 2008) ensures future data availability. Decreasing costs, the availability of free Landsat data in the Geocover dataset (Tucker et al., 2004), the “Mid-decadal Global Land Survey” (Khatib et al., 2007) and the USGS' decision to provide free access to all Landsat data holdings offer opportunities for large area land cover classifications using Landsat imagery.

Unfortunately, Landsat image classifications are commonly conducted on one scene at a time, which limits the rapid analysis of large areas (Cihlar et al., 1998, Cihlar, 2000) and requires that adequate ground truth data are available for each scene. For large area classifications, three approaches have been proposed and tested before: single scene classification and subsequent mosaicking, mosaicking of images and subsequent classification of the image mosaic as a whole (Cihlar, 2000), and signature extension. In signature extension, a classifier is trained on one scene and the resulting signatures are applied to different scenes in space or time (Pax-Lenney et al., 2001). Signature extension is promising, but has to account for differences in topography, phenology, illumination, landscape variability, and atmosphere that result in spectral differences among images. Tests in northwest Oregon showed that accuracy declined by 8–13% (depending on the atmospheric correction method applied) when extending the classifier from an initial training image across space to nearby scenes (Pax-Lenney et al., 2001). Across northern Canada, classification accuracy dropped approximately by 50% when using signature extension for images that were about 1500 km apart (Olthof et al., 2005).

A promising approach for mosaicking images prior to classification is ‘applied radiometric normalization’ (Cohen et al., 2001). Here, the overlap area between neighboring Landsat images is used to extend information gained from a source image to neighboring images, thereby creating a seamless mosaic for the classification. The first step is to develop a relationship between the spectral measurements in the source image, and continuous forest variables, such as percent vegetation cover or stand age that are available from ground truth data (Cohen et al., 2001). The second step is to apply the regression equations that were developed, and predict the forest structure attributes across the entire source image. In the third step, the map with the predictions in the overlap area is used as ground truth to develop new regression equations for the neighboring image, which has most likely different phenology and atmospheric conditions. In the fourth step, these regression equations are then applied to the entire neighboring image. The resulting map of continuous forest structure attributes for the entire study area can then be classified into different forest types. When testing this approach in a 73,000 km2 study area in western Oregon based on two Landsat TM source images, estimates for four forest cover attributes resulted in an overall accuracy of 66% (Cohen et al., 2001).

Signature extension and the mosaicking of images prior to classification have great potential for classifying large areas using Landsat imagery. However, they require considerable effort to match multiple images radiometrically. Here, we propose a new approach to large area land cover classification that fills a gap between single scene classification on one hand and signature extension or mosaicking on the other hand. We suggest the term ‘chain classification’ for this method.

Chain classification is similar to applied radiometric normalization in that it uses the overlap area among neighboring Landsat scenes, but we propose classifying one initial scene and then using the classification in the overlap area to train a classifier for a neighboring image. Once the second image is classified, it can be used as a new initial scene to classify a third image and so forth. One potential advantage of chain classification is that it does not require atmospheric correction or regression matching of scenes to account for radiometric differences. It can be applied both in horizontal directions (across track), and in vertical direction (along track). Furthermore, large area land cover maps often cover several countries or different land ownership regimes. The availability of spatially well distributed training and validation data is often limited in such situations. Chain classification may offer a solution to this problem by using the image with the best available ground truth data as the starting image in the image chain, and by providing training data for neighboring images from the image chain itself.

In principal, any classification algorithm could be used for chain classification. However, Support Vector Machines (SVMs), a fairly recently developed non-statistical classifier based on machine learning theory (Vapnik, 1999) offer some method-inherent advantages. Comparisons with other classification algorithms show that SVMs outperform or are at least as accurate as other parametric or non-parametric classifiers (Huang et al., 2002, Pal and Mather, 2005, Dixon and Candade, 2008).

SVMs are able to separate complex classes (Melgani & Bruzzone, 2004) such as in forest change analysis (Huang et al., 2008). In the SVM, the location of decision boundaries for optimal class separation is determined using kernel functions representing non-linear decision surfaces (Pal and Mather, 2005, [45]). By constructing the optimum hyperplane in feature space between two classes, an SVM is a binary classifier focusing on the classes of interest only. To determine this hyperplane, only the edges between the class distributions are described based on a relatively small amount of training data (Foody and Mathur, 2004, Foody et al., 2007, Mathur and Foody, 2008).

In summary, the overarching goal of this study was to develop a simple, robust, and reproducible method for large area land cover classification with minimal requirements for image pre-processing and training data. To do so, we tested chain classification of forest and non-forest based on the overlapping areas between Landsat scenes in the Carpathian Mountains in Eastern Europe.

Section snippets

Study area

We selected the Carpathians as a study area to test chain classification. The Carpathians represent a fairly homogeneous ecoregion with mostly similar environmental conditions. However, the study area includes seven countries with significant differences in forest type, non-forest land cover classes, geology, and land use patterns, and exhibits elevation-dependent vegetation gradients. This variability generates an interesting test case to investigate the feasibility of chain classification.

The

Across track chain classifications

The results for chain-classifying the direct neighbors had overall accuracy losses ranging from 0.26% for scene 5 to 6 (kappa loss 0.0091) to 4.50% for scene 1 to 2 (kappa loss 0.0882) with an average of 1.91% (kappa loss 0.0408) (Table 3 — Test A; Fig. 5). Both tests including scene 2 as a target scene, resulted in the highest overall accuracy and kappa losses.

Accuracy and kappa loss tend to increase as more scenes were added to a classification chain (Table 3). Average overall accuracy loss

Discussion

We tested ‘chain classification’, a new approach to classify land cover for large areas that uses a classification in one image to train a classifier for a neighboring image. The average loss of accuracy when comparing across track chain classifications to reference classifications was only 1.91% with two images in a chain, 2.81% with three, 2.47% with four, 4.76% with five and 5.11% with six scenes, respectively. Average pixel-wise agreement between two individually derived chain

Acknowledgements

We gratefully acknowledge support by the Humboldt-University Berlin and by the NASA Land Cover and Land Use Change Program for this research. T. Kuemmerle is supported by the Alexander von Humboldt Foundation. S. Schmidt, M. Ozdogan, and two anonymous reviewers provided valuable comments that greatly improved this manuscript.

References (49)

  • Pax-LenneyM. et al.

    Forest mapping with a generalized classifier and Landsat TM data

    Remote Sensing of Environment

    (2001)
  • SteeleB.M.

    Maximum posterior probability estimators of map accuracy

    Remote Sensing of Environment

    (2005)
  • TurnockD.

    Ecoregion-based conservation in the Carpathians and the land-use implications

    Land Use Policy

    (2002)
  • WulderM.A. et al.

    Landsat continuity: Issues and opportunities for land cover monitoring

    Remote Sensing of Environment

    (2008)
  • BoserB.E. et al.

    A training algorithm for optimal margin classifiers

  • BurgesC.J.C.

    A tutorial on support vector machines for pattern recognition

    Data Mining and Knowledge Discovery

    (1998)
  • BurmanP.

    A comparative-study of ordinary cross-validation, v-fold cross-validation and the repeated learning–testing methods

    Biometrika

    (1989)
  • CERI

    The status of the Carpathians

  • ChangC.-C. et al.

    LIBSVM: A library for support vector machines

    (2001)
  • CihlarJ.

    Land-cover mapping of large areas from satellites: Status and research priorities

    International Journal of Remote Sensing

    (2000)
  • CihlarJ. et al.

    Classification by progressive generalization: A new automated methodology for remote sensing multichannel data

    International Journal of Remote Sensing

    (1998)
  • CohenW.B. et al.

    Landsat's role in ecological applications of remote sensing

    Bioscience

    (2004)
  • CohenW.B. et al.

    Modelling forest cover attributes as continuous variables in a regional context with Thematic Mapper data

    International Journal of Remote Sensing

    (2001)
  • CortesC. et al.

    Support vector networks

    Machine Learning

    (1995)
  • Cited by (210)

    • 10 m crop type mapping using Sentinel-2 reflectance and 30 m cropland data layer product

      2022, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text