Landsat 8 vs. Landsat 5: A comparison based on urban and peri-urban land cover mapping

https://doi.org/10.1016/j.jag.2014.09.010Get rights and content

Highlights

  • Landsat TM and Landsat OLI are compared in the performance of landcover classification.

  • Commercial off-the-shelf (COTS) object based and pixel based methods have been applied.

  • The study area is a fragmented urban area in East Attica, Greece.

  • Pixel based classification by using SVM algorithm has the best results.

  • Landsat OLI has better classification accuracy comparing to Landsat TM.

Abstract

An image dataset from the Landsat OLI spaceborne sensor is compared with the Landsat TM in order to evaluate the excellence of the new imagery in urban landcover classification. Widely known pixel-based and object-based image analysis methods have been implemented in this work like Maximum Likelihood, Support Vector Machine, k-Nearest Neighbor, Feature Analyst and Sub-pixel. Classification results from Landsat OLI provide more accurate results comparing to the Landsat TM. Object-based classifications produced a more uniform result, but suffer from the absorption of small rare classes into large homogenous areas, as a consequence of the segmentation, merging and the spatial parameters in the spatial resolution (30 m) of Landsat images. Based exclusively on the overall accuracy reports, the SVM pixel-based classification from Landsat 8 proved to be the most accurate for the purpose of mapping urban land cover, using medium spatial resolution imagery.

Introduction

Thematic mapping is a prerequisite for several environmental and socioeconomic applications (Blaschke, 2010) and it is typically based on remotely sensed data and image classification (Chrysoulakis et al., 2010). One of the main issues when generating Land Cover (LC) maps from digital images is the confusion of spectral responses from different features. The accuracy of the classified map depends on the spatial and spectral resolution, the seasonal variability in vegetation cover types and soil moisture conditions. Landsat series of satellites are the most common Earth Observation (EO) data sources for LC mapping, even for urban, peri-urban and rural areas. Landsat Thematic Mapper (TM) started providing multispectral observations in 1984. In 2004, NASA sponsored the creation of the Global Orthorectified Landsat Dataset (Tucker et al., 2004). Recently, with the launch of the Landsat 8, carrying the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS), a new orthorectified dataset (L1T) became available (Roy et al., 2014). Landsat OLI, TIRS and TM imagery, having a nominal spatial resolution of 30 m are considered as low resolution (LR) (Strahler et al., 1986). Nevertheless, it can be used for mapping urban and peri-urban areas, especially in areas with moderate population density and complex landscape. On the contrary, high resolution (HR) imagery provided by sensors such as Worldview-2, is normally used in urban LC mapping. However, such datasets are costly and their analysis needs much more computing resources than the respective Landsat ones. The above LR and HR cases will be bridged by the expected Copernicus Sentinel-2, with a spatial resolution of 10 meters in visible and near infrared bands and a revisit time of 5 days (Drusch et al., 2012). Sentinel-2 is therefore expected to be used complementary with Landsat 8 for urban areas mapping and monitoring. Urban and peri-urban LC classification with the use of LR imagery is challenging due to the spectral mixing of different surface elements and the landscape complexity. Conventional pixel-based classifiers, such as Maximum Likelihood (MLC) (Jensen, 2000), cannot effectively handle the mixed-pixel problem in complex urban/peri-urban areas. Alternative approaches, such as the Support Vector Machine (SVM) (Foody and Mathur, 2004) and Geographic Object-Based Image Analysis (GEOBIA) (Blaschke, 2010) provide better results although they do not address the mixed-pixel problem.

Huang et al. (2002) compared classification products from four different classification algorithms implemented on Landsat TM data and found that SVM outperformed the rest three approaches. Similar results were provided by Duro et al. (2012), who examined the performance of three classification algorithms SPOT-5 high resolution geometrical (HRG). They found that in the pixel-based classification results, even if the overall accuracy between the above three approaches (Decision Tree-DT, Random Forest-RF and SVM) was not statistical significant, the SVM classification achieved better discrimination between riparian vegetation and grasslands and thus obtained less speckles in such areas. Robertson and King (2011) analyzed Landsat TM images from different periods and found that the accuracies for the object-based classifications were lower than the respective accuracies of pixel-based methods. This could be related to the segmentation process and the scale factor that is selected for the segmentation. Duro et al. (2012) provide similar results, using SPOT imagery, while Gong et al. (2013) create a global landcover dataset from Landsat TM and ETM+ by applying SVM. Classification maps can be acceptable depending on the purpose of the classification product (input for hydrological or fire propagation models, cartography purposes, etc.). The objective of this work is to compare the performance of Landsat 8 OLI against Landsat TM for urban and peri-urban LC mapping, using common training and validation data.

Section snippets

Classification algorithms

To assess the differences of OLI and TM for LC mapping, the performances of the following image classification methods were compared:

  • Pixel based MLC and SVM.

  • ENVI Feature Extraction with SVM & k-NN and Feature Analyst (FA).

  • Sub-pixel Linear Spectral Mixture Analysis (LSMA).

MLC and SVM have been extensively discussed in previous studies and thus we will not discuss them (Lu and Weng, 2007).

Concerning Object-based Machine Learning methods, the FA and ENVI Feature Extraction were used in this study.

Study area and datasets

The study area covers the catchment of Rafina Municipality (Fig. 1), an area of 123 km2, located in Attika, Greece. It is a recently developed area, close to the “Eleftherios Venizelo” International Airport of Athens and to the Attiki Odos highway (A6 highway). This highway connects the study area with the city of Athens, favoring urban sprawl (Chrysoulakis et al., 2013). It should be however noted that, given the present economic status of Greece the urbanization rate in the study area has been

Classification scheme

Knowledge of the study area and visual inspection of the available images assisted in developing the classification scheme. False color composites, RGB: 4-3-2 for TM and RGB: 5-4-3 for OLI clearly depicted the forest vegetation in dark red, the water associated vegetation in bright red and the urban surface materials in light bluish tones, while it was difficult to distinguish dry cultivated areas and low vegetation. The latter one was possible using the true color combination RGB: 3-2-1 for TM

Pixel-based classification results

The main elements of the anthropogenic environment detected in the study area are compact urban areas; small settlements; isolated building infrastructures; and the main road network in the central and southwest area. The main elements of the natural environment are: forest islets; an extended area of mixed cultivations to the south; low vegetated areas in the north; and an extended network of streams with riparian vegetation. The per-pixel classification maps from Landsat TM and OLI images are

Conclusions

In this study, a set of different classification methods were applied to Landsat TM and OLI data in order to assess and compare their performance in classifying peri-urban areas. In highly fragmented landscapes, like the peri-urban area of Athens, by analyzing medium resolution imagery such as provided by Landsat TM and OLI, using advanced methods like the SVM, it is possible to result in highly accurate classification maps. SVM can maintain the spatial characteristics of such landscapes, like

Acknowledgement

This research project is funded by the FLIRE (LIFE11ENV/GR/975), a LIFE + European Union project.

References (31)

  • N. Chrysoulakis et al.

    A comparative analysis of the urban web of the greater Athens agglomeration for the last 20 years period on the basis of Landsat imagery

    Fresenius Environ. Bull.

    (2013)
  • M.J. Collins et al.

    Mapping subalpine forest types using networks of nearest neighbor classifiers

    Int. J. Remote Sens.

    (2004)
  • D. Gadkari

    Image Quality Analysis Using GLCM

    (2004)
  • R.M. Haralick et al.

    Textural features for image classification

    IEEE Trans. Syst. Man Cybern.

    (1973)
  • M. Herold et al.

    The spectral dimension in urban remote sensing

    Remote Sensing of Urban and Suburban Areas

    (2010)
  • Cited by (0)

    View full text