The use of Support Vector Machines (SVM) for Land Cover Mapping from Remote Sensor Imagery

Land Cover Mapping:

Land cover mapping has become one of the essential studies in environmental and remote sensing science.  No later than the introduction of remote sensing, the lad cover mapping has gained popularity among scientists and researchers. Land cover mapping is a technique of mapping the features on the earth’s surface using remote sensing imageries and deploying different tools and algorithms. Basically, it is a science of generating various thematic products essential for numerous environmental monitoring and resource management applications at local, regional, and global levels. In other words, the pattern of ecological resources and human activities dominating different areas of Earth’s surface. These patterns are observable and can be mapped by ground surveys or remote sensing.

Over the years, various pattern recognition techniques have been developed to automate this process from remote sensor imagery. Support vector machines (SVM) is one of them.

Support Vector Machine (SVM):

It is a group of relatively novel statistical learning algorithms that have demonstrated their robustness in classifying homogeneous and heterogeneous land cover types. Because of its robustness, SVM is implemented to map various land cover types from a remote sensor image covering an urban area, demonstrating the robustness of this type of pattern recognition technique for mapping heterogeneous landscapes. In terms of classification accuracy, SVM can perform better than maximum likelihood (MLC) or decision tree (DC) in terms of classification accuracy, and multilayer perceptron neural networks (MLP). 

However, there are some parametric and non-parametric factors that can affect the performance of SVM. Since the SVM is a supervised classifier by nature, both the size and quality of the training sample and the noise in training samples can affect the classification accuracy.

SVM in Land Cover Mapping

The basic idea behind the SVM is to construct separating hyperplanes between classes in feature space through the use of support vectors that are lying at the edges of class domains; SVM seeks the optimal hyperplane that can separate classes from each other with the maximum margin. The two hyper-planes are selected so as not only to maximize the distance between the two given classes but also not to include any points between them. The overall goal is to find out in which class the new data points fall.

The success of the SVM depends on how well the process is trained. The first step in the classification process was the development of the classification scheme. You can define your land cover classification scheme depending upon the projects you are doing. Some popular land cover classification consists of these classes: forest, grassland, agricultural land, water, urban, and so on.

The easiest way to train the SVM is by using linearly separable classes. Overall, the SVMs are reported to produce results of higher accuracies compared with the traditional approaches but the outcome depends on the kernel used, choice of parameters for the chosen kernel, and the method used to generated SVM.

For example, if the training data with k number of samples is represented as {Xi, yi}, i = 1, …, k where X € RN is an N-dimensional space and y € {-1, +1} is a class label then these classes are considered linearly separable if there exists a vector W perpendicular to the linear hyper-plane (which determines the direction of the discriminating plane) and a scalar b showing the offset of the discriminating hyperplane from the origin. For the two classes, i.e. class 1 represented as -1 and class 2 represented as +1, two hyper-planes can be used to discriminate the data points in the respective classes. These are expressed as:

WXi + b ≥ +1 for all y = 1; i.e. a member of class 1

WXi + b ≤ -1 for all y = -1; i.e. a member of class 2

Some useful references:

  • Alberti M (2008) Advances in urban ecology: integrating humans and ecological processes in urban ecosystems. Springer, New York
  • Bazi Y, Melgani F (2006) Toward an optimal SVM classification system for hyperspectral remote sensing images. Geosci Remote Sens IEEE Trans 44(11):3374–3385
  • Camps-Valls G, Gomez-Chova L, Mu~noz-Mari J, Vila-Frances J, Calpe-Maravilla J (2006) Composite kernels for hyperspectral image classification. IEEE Geosci Remote Sens Lett 3(1):93–97
  • Congalton RG (1991) A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens Environ 37(1):35–46
  • Del Frate F, Pacifici F, Schiavon G, Solimini C (2007) Use of neural networks for automatic classification from high-resolution images. IEEE Trans Geosci Remote Sens 45(4):800–809
  • Demir B, Ertu¨rk S (2009) Clustering based extraction of border training patterns for accurate SVM classification of hyperspectral images. IEEE Geosci Remote Sens Lett 6(4):840–844
  • Dixon B, Candade N (2008) Multispectral landuse classification using neural networks and support vector machines: one or the other, or both? Int J Remote Sens 29(4):1185–1206 Duda RO, Hart PE, Stork DG (2001) Pattern classification. Wiley, New York
  • Foody GM (2008) RVM-based multi-class classification of remotely sensed data. Int J Remote Sens 29(6):1817–1823
  • Blaschke, T., Lang, S., 2006. Object based analysis for automated information extraction-a synthesis. In: MAPPS/ASPRS Fall Conference, San Antonio, TX.
  • Brodley, C.E., Utgoff, P.E., 1992. Multivariate versus univariate decision trees. Technical Report 92-8. University of Massachusetts, Amherst, MA, USA. Erdas Inc., 1999.
  • Erdas Field Guide. Erdas Inc., Atlanta, Georgia. FAO, 2005. Land Cover Classification System (LCCS), Classification Concepts and Users Manual. FAO, Rome, Italy.
  • Foody, G.M., 1986. Approaches for the production and evaluation of fuzzy land cover classification from remotely sensed data. International Journal of Remote Sensing 17, 1317–1340.
  • Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote Sensing of the Environment 80, 185–201. Geneletti, D., Gorte, B.G.H., 2003. A method for object oriented land cover classification combining landsat TM and aerial photographs. International Journal of Remote Sensing 24, 1273–1286.
  • Hay, G., Blaschke, T., et al., 2003. A comparison of three image-object methods for the multiscale analysis of the landscape structure. ISPRS Journal of Photogrammetry and Remote Sensing 57, 327–345.

Leave a Reply

%d bloggers like this: