A semi-automatic method for analysis of landscape elements using Shuttle Radar Topography Mission and Landsat ETM+ data
Introduction
Information about landforms and landscape is one of the fundamental requirements for a large variety of modeling problems in environmental science. Most researchers define landscape as an essentially visual phenomenon or as a particular configuration of topography, land use, vegetation cover and settlement pattern (Blankson and Green, 1991; Otero Pastor et al., 2007). Landscapes are dynamic systems that involve interrelation between physical characteristics (such as landform, soil) and anthropogenic processes (such as land use). Relationship between these physical properties and human impact on the land has led to the development of different analysis models. These models vary from visual analysis and quantitative techniques to rule-based geoecosystem techniques (Benefield and Bunce, 1982; Bernert et al., 1997; Blankson and Green, 1991).
However, all these models are based on the common task to find the basic elements of a heterogeneous landscape. The basic concept of natural landscape units, called geochores, was developed in the 1950s and 1960s. The term ‘geochore’ means a geographically defined or limited unit and can be regarded as mosaics of basic topic elements (Bastian, 2000). Landform as physical constituent of landscape may be extracted from digital elevation data using various approaches including classification of morphometric parameters (Dikau, 1989; Dehn et al., 2001), fuzzy set methods and unsupervised (ISODATA) classification (Adediran et al., 2004; Burrough et al., 2000; Irvin et al., 1997), probabilistic clustering algorithm (Stepinski and Collier, 2004; Stepinski and Vilalta, 2005), multivariate descriptive statistics (Dikau, 1989; Evans, 1972; Dehn et al., 2001) and double ternary diagram classification (Crevenna et al., 2005). Landforms possess at least two important properties. They are the result of past geomorphic and geologic processes and provide a controlling boundary condition for actual geomorphic processes (Dehn et al., 2001). For disciplines dealing with landforms, the properties of consideration are different. For geomorphologists both properties of landforms are important. But a common perspective of all landform studies regardless of discipline is to delimit homogeneous areas from digital elevation data. Digital elevation models (DEM) can be compiled from contour lines or other sources like the Shuttle Radar Topography Mission (SRTM). On 11 February 2000, the space shuttle Endeavour with the SRTM payload on board was launched. A single technique, synthetic aperture radar (SAR) interferometry, was used for producing a consistent DEM covering all landmasses on earth between 60°N and 57°S (Blumberg, 2006; Rabus et al., 2003; Wright et al., 2006). The 3 arc sec. (∼90 m) SRTM data are publicly available at http://seamless.usgs.gov.
Describing the shape of surface features on earth using a set of numerical measures (derivatives) such as profile curvature, plan convexity, slope, cross-sectional curvature, minimum and maximum curvature from DEMs is known as morphometry. Morphometric characterization identifies morphometric features such as saddle, channel, ridge and plane based on these measures (Fisher et al., 2004; Pike, 2000; Wood, 1996). Integrating land surface forms (morphometric features) with spectral information from remotely sensed data contributes to the explanation of relationships between landscape component processes (physical, biotic and human activities) on one hand and delimiting boundaries of homogenous landscape elements on the other hand.
In recent years, there has been considerable interest in using neural networks with remotely sensed data. Self-organizing map (SOM) is an unsupervised neural network algorithm, which clusters or visualizes high-dimensional input vectors into low-dimensional (usually two-dimensional) output based on regularities and correlations between them (Jianwen and Bagan, 2005; Kohonen, 2001; Li and Eastman, 2006). SOM has been used in a wide variety of areas such as classification of remote-sensing data (Duda and Canty, 2002; Jianwen and Bagan, 2005), information visualization and knowledge discovery (Koua et al., 2006), class modeling (Marini et al., 2005) and semi-automatic terrain analyses (Ehsani and Quiel, in press).
The main objective of this paper is characterization of landscape elements through the combination of morphometric parameters and remotely sensed spectral data. The emphasis is on morphologically homogeneous landscape elements characterized mainly by similar slope conditions. SOM as unsupervised paradigm of neural networks is used to reduce large multidimensional data to one output layer consisting of 20 map units. This output indicates land cover of landscape elements on one hand and impact of geologic and geomorphological processes (morphometric features) on the other hand.
Section snippets
Study area
The study area is centered on the common border point of Poland, Slovakia and Ukraine and located between 48°52′N and 49°25′N latitude, 21°59′E and 23°1′E longitude with a total area of about 4500 Km2 (Fig. 1). It covers the biosphere reserve “Eastern Carpathians” with the Bieszczady national park in Poland, Uzanski national park in Ukraine and Poloniny national park in Slovakia.
Historically, the region had similar land management policies. After World War II, fundamental changes in political
Data
The data set in this study consists of:
- •
Landsat ETM+ data path 186, row 26 dated 30 September 2000 (Fig. 1) were acquired from the Global Land Cover Facility (GCLF) server at the University of Maryland, Institute for Advanced Computer Studies (UMIACS). GLCF provides free access to an integrated collection of critical land cover and earth science data (http://glcf.umiacs.umd.edu).
- •
The 3 arcsec digital elevation model derived from SRTM data (∼90 m) was acquired from the USGS server in geographic
Optimal self-organizing map
Learning of SOM was performed with randomly initialized weights of the map units. Table 3 shows the 16 SOM with different configuration of learning control parameters.
The initial radius for learning was set to 3, 2 and 1, respectively. The map with an initial radius of three and a final radius of one produced the highest average quantization error. Results indicate that the final radius should be much smaller than 1. In several trials with different configurations of learning parameters, SOM 16
Conclusion
This study demonstrates the effectiveness of self-organizing map (SOM) algorithms of artificial neural networks as semi-automatic methods to extract both spectral and morphometric information from SRTM and Landsat 7 data. Analyzing and interpreting this information yielded 20 classes of homogeneous landscape elements. Morphometric analysis of first- and second-order derivatives of DEM data, such as slope, cross-sectional curvature, maximum curvature and minimum curvature, led to the description
Acknowledgements
We are grateful to the Swedish Institute for funding all travel expenses in the framework of the Visby program. We thank all our colleagues, especially, Docent Ivan Kruhlov, Department of Physical Geography; Ivan Franko, University in Lvov, Ukraine and Dr. Mieczyslaw Sobik, Institute of Geography and Regional Development, University Wroclaw, Poland for interesting discussions and for providing facilities and support. We thank Dr. Kuemmerle for providing his land use map of the project area.
References (37)
- et al.
Computer-assisted discrimination of morphological units on north-central Crete (Greece) by applying multivariate statistics to local relief gradients
Geomorphology
(2004) - et al.
The self organizing map, the Geo-SOM, and relevant variants for geosciences
Computers and Geosciences
(2005) Landscape classification in Saxony (Germany)—a tool for holistic regional planning
Landscape and Urban Planning
(2000)- et al.
Use of landscape classification as an essential prerequisite to landscape evaluation
Landscape Urban Planning
(1991) Analysis of large aeolian (wind-blown) bedforms using the Shuttle Radar Topography Mission (SRTM) digital elevation data
Remote Sensing of Environment
(2006)- et al.
High-resolution landform classification using fuzzy k-means
Fuzzy Sets and Systems
(2000) - et al.
Principles of semantic modeling of landform structures
Computers and Geosciences
(2001) - et al.
SRTM-based morphotectonic analysis of the Pocos de Caldas Alkaline Massif, southeastern Brazil
Computers and Geosciences
(2007) - et al.
Fuzzy and isodata classification of landform elements from digital terrain data in Pleasant Valley, Wisconsin
Geoderma
(1997) - et al.
Land-use classification using ASTER data and self-organized neural networks
International Journal of Applied Earth Observation and Geoinformation
(2005)
Cross-border comparison of land cover and landscape pattern in Eastern Europe using a hybrid classification technique
Remote Sensing of Environment
Class-modeling using Kohonen artificial neural networks
Analytica Chimica Acta
Landscape evaluation: comparison of evaluation methods in a region of Spain
Journal of Environmental Management
The Shuttle Radar Topography Mission—a new class of digital elevation models acquired by spaceborne radar
ISPRS (International Society for Photogrammetry and Remote Sensing) Journal of Photogrammetry and Remote Sensing
Using self-organizing maps to identify patterns in satellite imagery
Progress in Oceanography
An assessment of Shuttle Radar Topography Mission digital elevation data for studies of volcano morphology
Remote Sensing of Environment
A quantitative method for delineating regions: an example for the western Corn Belt plains ecoregion of the USA
Environmental Management
Cited by (26)
A fine digital soil mapping by integrating remote sensing-based process model and deep learning method in Northeast China
2024, Soil and Tillage ResearchSelf-Organizing Maps for identification of zeolitic diagenesis patterns in closed hydrologic systems on the Earth and its implications for Mars
2021, International Journal of Sediment ResearchCitation Excerpt :SOMs are different from other artificial neural network (ANN) methods since they apply competitive learning instead of error-correction learning, such as backpropagation, and they organize the clusters in a two dimensional space according to the similarity of the cluster centroids themselves (Dowla & Rogers, 1995; Wehrens & Buydens, 2007). SOMs can be applied in every field of geosciences including geochemistry (e.g. Penn, 2005), petrology (e.g. Chang et al., 2002), seismology (e.g. Kohler et al., 2009), GIS (e.g. Bacao et al., 2005) and remote sensing (e.g. Ehsani & Quiel, 2009), geomorphology (e.g. Bue & Stepinski, 2006), volcanology (e.g. Ersoy et al., 2007), and marine geology (e.g. Ferentinou et al., 2012). SOMs can be used for both continuous and categorical data.
Hyper-temporal remote sensing data in bare soil period and terrain attributes for digital soil mapping in the Black soil regions of China
2020, CatenaCitation Excerpt :These processed data could represent some characteristics of the study area in a new form (Kunkel et al., 2011; Dobos et al., 2001). Previous studies show that the combination of terrain factors and RS image data could provide more accurate soil type information (Ehsani and Quiel, 2009; Dobos et al., 2000). Terrain is one of the five major soil-forming factors, and is also one of the most commonly used covariates for predicting soil types and properties, especially for DSM (McBratney et al., 2003).
Hyper-temporal remote sensing for digital soil mapping: Characterizing soil-vegetation response to climatic variability
2017, GeodermaCitation Excerpt :Terrain attributes and spectral indices often exhibit considerable cross correlation due to the strong interrelationships between factors in the soil environment. Nonetheless, previous studies have shown that the combination of both terrain and RS data produces the most accurate classifications of soils in complex geomorphic landscapes (Dobos et al., 2000; Ehsani and Quiel, 2009; Martin and Franklin, 2005; Taramelli and Melelli, 2009), indicating that terrain and spectral indices each possess some unique amount of explanatory power. Future research is needed to examine the added benefit of incorporating both hyper-temporal RS and terrain attributes into soil prediction models.
Uncoupling the complexity of forest soil variation: Influence of terrain indices, spectral indices, and spatial variability
2016, Forest Ecology and ManagementCitation Excerpt :While considerable shared variance can exist between terrain and spectral indices, large portions of modeled variance is also unique to each data type. Previous studies also confirm this where the combination of both terrain and remotely sensed spectral data has been shown to produce the most accurate classifications of soils in complex geomorphic landscapes (Dobos et al., 2000; Ehsani and Quiel, 2009; Martin and Franklin, 2005; Miller et al., 2015; Taramelli and Melelli, 2009). For example, Dobos et al. (2000) found that AVHRR data accurately delineated soils formed on different parent materials, i.e., sandy regions, clayey fluvial deposits, loess areas; however, within each parent material type, different soil classes were found depending upon differences in elevation.
- 1
Postal address: Brinellvägen 34, 10044 Stockholm, Sweden.