Elsevier

Applied Geography

Volume 31, Issue 2, April 2011, Pages 533-544
Applied Geography

Comparison of methods for land-use classification incorporating remote sensing and GIS inputs

https://doi.org/10.1016/j.apgeog.2010.11.006Get rights and content

Abstract

Over the last few decades, dramatic land-use changes have occurred throughout Israel. Previously-grazed areas have been afforested, converted to irrigated or rain-fed agriculture, turned into natural reserves, often used as large military training sites, converted to rural and urban settlements, or left unused. Land-use maps provided by the Israeli governmental are more detailed for agricultural and urban land-use classes than for others. While rangelands still account for a substantial part of the northern Negev, their extent today is not well defined. In light of continuous land-use changes and lack of regard to rangelands in existing land-use maps, there is a need for creating a current land-use information database, to be utilized by planners, scientists, and decision makers. Remote-sensing (RS) data are a viable source of data from which land-use maps could be created and updated efficiently. The purpose of this work is to explore low-cost techniques for combining current satellite RS data together with data from the Israeli Geographic Information System (GIS) in order to create a relatively accurate and current land-use map for the northern Negev. Several established methods for land-use classification from RS data were compared. In addition, ancillary land-use data were used to update and improve the RS classification accuracy within a GIS framework. It was found that using a combination of supervised and unsupervised training classes produces a more accurate product than when using either of them separately. It was also found that updating this product using ancillary data and GIS techniques can improve the product accuracy by up to 10%. The final product’s overall accuracy was 81%. It is suggested that applying the presented technique for more RS images taken at different times can facilitate the creation of a database for land-use changes.

Research highlights

► Land-use classification of the northern Negev was performed with 81% accuracy. ► Combined supervised and unsupervised training were used to achieve greater accuracy. ► Ancillary data was used post-classification to improve the accuracy by up to 10%. ► Rangelands were defined in a land-use map of this area for the first time.

Introduction

The observed biophysical cover of the earth’s surface, termed land-cover, is composed of patterns that occur due to a variety of natural and human-derived processes. Land-use, on the other hand, is human activity on the land, influenced by economic, cultural, political, historical, and land-tenure factors. Remotely-sensed data (i.e., satellite or aerial imagery) can often be used to define land-use through observations of the land-cover (Brown et al., 2000, Karl and Maurer, 2010). Up-to-date land-use information is of critical importance to planners, scientists, resource managers, and decision makers.

One way to extract land-use information from remote-sensing data is through visual interpretation. However, visual interpretation is limited to a single band or a three-band (RGB) color composite. Manual digitization of land-use patches is extremely tedious as well as subjective (Bolstad, Gessler, & Lillesand, 1990). Therefore, automatic classification of remote sensing is more suitable for mapping land-use in a large area. While land-use and land-cover patterns may be obvious to an image interpreter, automatically mapping them could be difficult because automated classification techniques do not possess the superior pattern recognition capabilities of the human brain (Hudak & Brockett, 2004). When automatically classifying a complex landscape from remote-sensing imagery, it is challenging to achieve an accurate classification (Manandhar, Odeh, & Ancev, 2009). It has been claimed before that the eastern Mediterranean landscapes are considered the most heterogeneous of all (Alrababah & Alhamad, 2006). Therefore, classifying the landscape in this region is not a trivial task.

Nevertheless, previous studies show that Landsat Thematic Mapper (TM) images with the spatial resolution of 30 m are sufficient to accurately classify a large variety of landscapes from the homogeneous tropical landscapes to the heterogeneous Mediterranean landscapes (Alrababah and Alhamad, 2006, Koutsias and Karteris, 2003, Manandhar et al., 2009, Sader et al., 1995, Schulz et al., 2010). Landsat has been providing a nearly continuous record of global land surface change since 1972 (Cohen & Goward, 2004). Currently, two Landsat sensors in orbit are operational: TM on board Landsat-5 and Enhanced Thematic Mapper Plus (ETM+) on board Landsat-7. Both sensors acquire measurements in all major portions of the solar electromagnetic spectrum (visible, near-infrared, and shortwave-infrared), providing significant advantage over less capable sensor systems. However, Landsat-7, launched in 1999, developed a problem with the scan-line corrector in 2003, leading to reduced data quality for land-use mapping applications (Powell, Pflugmacher, Kirschbaum, Kim, & Cohen, 2007). Today, Landsat-5, launched in 1984, has far exceeded its 3-year life expectancy but continues to provide quality data products, although it was expected to run out of fuel by late 2010 (Wulder et al., 2008). Landsat data are widely applied for land-use classification on a regional scale due to their relatively lower cost, longer history, and higher frequency of archives in comparison to other remote-sensing data sources.

It has been previously determined that satellite image classification results did not improve over a period of 15 years in spite of vigorous and creative efforts to establish new classification algorithms during this period (Wilkinson, 2005). Therefore, it was concluded there is little value in continued research efforts to improve classification algorithms in remote sensing (Manandhar et al., 2009). Recently, the trend amongst researchers has been to let geographical data “have a stronger voice” rather than let statistically-derived parameters dictate the analysis. Integration of remotely-sensed data with other sources of georeferenced information, such as previous land-use data, spatial texture, and digital elevation models (along with their derivatives: slope, aspect, etc.), geology, soils, hydrology, transportation network, vegetation, and climate enable greater classification accuracy to be achieved (Lillesand and Kiefer, 2000, Manandhar et al., 2009, Stefanov et al., 2001, Tateishi and Shalaby, 2007). The particular sources of data used and how and when they are employed in a given application are normally determined through a set of decision rules formulated by the image analyst. The integration of several data sources in a Geographic Information System (GIS) allows the analyst to develop a series of post-classification decision rules utilizing all the data sources in combination (Lillesand & Kiefer, 2000). The integration of remote-sensing data, GIS and “expert system” techniques to form Decision Support Systems (DSS) can provide better classification accuracies than any of the individual data sources used alone.

The purpose of this work was to explore low-cost techniques for land-use mapping. Landsat TM imagery was classified by two widely used and established classification approaches, and these two methods were combined and compared. Next, the hypothesis that integrating current satellite remote-sensing data together with data from the Israeli GIS will improve the land-use mapping significantly was tested. The land-use classification technique presented in this work can be used to produce information pertaining changes in land-uses, such as monitoring of land-use conversion and land degradation. The information could be further used to study the relations between land-use changes and other phenomena such as carbon fixation, biodiversity, climate change, and sustainable management of natural resources.

Specific objectives of the current study are:

  • 1.

    To compare between supervised and unsupervised land-use classification techniques;

  • 2.

    To examine whether combining signatures from both supervised and unsupervised training data (hybrid classification) provides significantly more accurate results then each approach separately;

  • 3.

    To examine whether using a decision support system for updating the map based on expert knowledge and ancillary GIS data improves the classification accuracy significantly.

Section snippets

Study area

Located in the northern Negev, on the desert fringe, the study area (Fig. 1) is about 4000 km2 in size. The study area’s borders are delimited by Ramat-Hovav in the south, Yatir forest in the east, Kiryat-Gat and Ashkelon in the North, and the Mediterranean Sea, Gaza and Sinai in the west. This area is particularly diverse since it lies on the transition zone between arid, semi-arid, and Mediterranean climate zones. Average annual precipitation decreases along two climate gradients from north

Methodology

Initially, a Landat-5 TM image of the northern Negev was pre-processed and then classified in several ways using ERDAS IMAGINE 2010. Post-classification, a decision support system based on expert knowledge was used to update the classification products according to existing land-use databases using ArcGIS 9.3. The accuracy of each of the derived classification products was assessed in several ways, after which different product accuracies were compared using statistical means with STATISTICA

Reflective band classification

The accuracies of both the supervised and unsupervised classification of the reflective bands were assessed. Confusion matrixes and accuracy measures can be found in Table 3, Table 4. Judging by the overall accuracy and overall Kappa statistics, it is apparent that the unsupervised classification is superior to the supervised classification (overall accuracy of 70.67% vs. 60.83%, respectively, Kappa statistic of 0.65 vs. 0.53, respectively). A McNemer’s test confirmed that the unsupervised

Discussion

A Landsat TM image was pre-processed and classified using three methods: supervised, unsupervised, and hybrid classification methods. Following this, the classification products’ accuracy was assessed. A comparison of the products’ accuracy was conducted to find out if the accuracy differences are statistically significant. It was found that unsupervised training produces more accurate results than supervised training. A hybrid supervised–unsupervised classification also produced more accurate

Conclusions

  • When lacking intimate familiarization with a large, complex, and heterogeneous area, unsupervised classification has a potential to produce more accurate results than supervised classification.

  • Hybrid supervised–unsupervised classification produced more accurate classifications than the supervised classification; however, it did not improve the accuracy significantly in comparison to the unsupervised classification.

  • Using a decision support system for updating the map based on expert knowledge

Acknowledgements

This Research program was sponsored by the Mid-East Program, International Programs, United States Forest Service.

The authors wish to thank the helpful comments and advise from both anonymous reviewers.

References (50)

  • D.E. Orenstein et al.

    To populate or preserve? Evolving political-demographic and environmental paradigms in Israeli land-use policy

    Land Use Policy

    (2009)
  • S.A. Sader et al.

    Accuracy of Landsat-TM and GIS rule-based methods for forest wetland classification in Maine

    Remote Sensing of Environment

    (1995)
  • J.J. Schulz et al.

    Monitoring land cover change of the dryland forest landscape of central Chile (1975–2008)

    Applied Geography

    (2010)
  • C. Song et al.

    Classification and change detection using Landsat TM Data - when and how to correct atmospheric effects?

    Remote Sensing of Environment

    (2001)
  • W.L. Stefanov et al.

    Monitoring urban land cover change: an expert system approach to land cover classification of semiarid to arid urban centers

    Remote Sensing of Environment

    (2001)
  • M.A. Wulder et al.

    Landsat continuity: issues and opportunities for land cover monitoring

    Remote Sensing of Environment

    (2008)
  • G. Xian et al.

    Updating the 2001 National land cover database land cover classification to 2006 by using Landsat imagery change detection methods

    Remote Sensing of Environment

    (2009)
  • M. Zomeni et al.

    Historical analysis of landscape change using remote sensing techniques: an explanatory tool for agricultural transformation in Greek rural areas

    Landscape and Urban Planning

    (2008)
  • M.A. Alrababah et al.

    Land use/cover classification of arid and semi-arid Mediterranean landscapes using Landsat ETM

    International Journal of Remote Sensing

    (2006)
  • J.R. Anderson et al.

    A land use and land cover classification system for use with remote sensor data

    (1976)
  • P.V. Bolstad et al.

    Positional uncertainty in manually digitized map data

    International Journal of Geographical Information Systems

    (1990)
  • J.B. Campbell

    Introduction to remote sensing

    (1996)
  • P.S. Chavez

    Image-based atmospheric corrections - revisited and improved

    Photogrammetric Engineering and Remote Sensing

    (1996)
  • W.B. Cohen et al.

    Landsat’s role in ecological applications of remote sensing

    Bioscience

    (2004)
  • R.G. Congalton et al.

    Assessing Landsat classification accuracy using discrete multivariate analysis statistical techniques

    Photogrammetric Engineering and Remote Sensing

    (1983)
  • Cited by (229)

    View all citing articles on Scopus
    View full text