Académique Documents
Professionnel Documents
Culture Documents
DEPARTMENT OF GEOSPATIAL
ENGINEERING AND SPACE TECHNOLOGY
TITLE: A COMPARISON OF
CLASSIFICATION METHODS FOR
MAPPING LAND COVER
FGE 447
LECTURER: Dr Faith N. Karanja
ABSTRACT
Classification of land cover is one of the most important tasks and one of the primary objectives in the
analysis of remotely sensed data.
The classification process aims at assigning each pixel from the analysed scene to a particular class of
interest, such as urban area, forest, water, roads, etc. The image resulting from the labelling of all
pixels is henceforth referred to as a thematic map.
Such maps are very useful in many remote sensing applications especially those concerned with
agricultural production monitoring, land change cover and environmental protection. Conventional
classification methods classify each pixel independently by considering only its observed intensity
vector. The result of such methods has often a salt and pepper appearance which is a main
characteristic of misclassification. In particular of remotely sensed satellite imagery, adjacent pixels
are related or correlated, both because imaging sensors acquire significant portions of energy from
adjacent pixels and because ground cover types generally occur over a region that is large compared
with the size of a pixel. It seems clear that information from neighbouring pixels should increase the
discrimination capabilities of the pixel-based measured data, and thus, improve the classification
accuracy and the interpretation efficiency. This information is referred to as the spatial contextual
information. In this report I present several classification methods
Decision tree analysis (Per-pixel algorithm)
Decision Trees are a non-parametric supervised learning method used
for classification and regression. The goal is to create a model that predicts the value of a target
variable by learning simple decision rules inferred from the data features.
Decision tree builds classification or regression models in the form of a tree structure. It breaks down
a dataset into smaller and smaller subsets while at the same time an associated decision tree is
incrementally developed. The final result is a tree with decision nodes and leaf nodes. A decision
node has two or more. Leaf node (e.g., Play) represents a classification or decision. The topmost
decision node in a tree which corresponds to the best predictor called root node. Decision trees can
handle both categorical and numerical data.
In remote sensing, the construction of decision tree requires supervised training; therefore it
is necessary to have a training dataset consisting of response and explanatory variables. In
classification problems involving remote sensing dataset, the response variables are generally
land use land cover classes and explanatory variables are spectral bands or information
derived from these. The classification structure defined by a decision tree is estimated from
training data using a statistical procedure. The "tree" is made of a root node, internal nodes
and leaves. Nodes are where trees branch or split the data set; terminal nodes are called
leaves which contain most homogeneous classes. If in a training set T, there is k number of
classes (C) and a total of |T| cases, the expected information from such a system is,
info(T)=(Tfreq(Cj,T)/T)log2 (freq(Cj,T)/T) where Tfreq(Cj,T)/T is the probability of
occurrence of class Cj in training set T. If we partition the training set T in accordance with
any response variable X (e.g. NDVI), there may be 'n' number of cases.
There are several types of decision tree classification algorithms
Univariate decision tree
Multivariate decision tree
Hybrid decision tree
Performance and Comparison
Several studies have compared decision tree classification methods with other classifiers. Otukei and
Blaschke (2010) compared decision tree, maximum likelihood and support vector machine based
techniques for land cover change assessment using Landsat TM and ETM+ data and found decision
tree based methods performed better than others. Punia et al. (2011) used C 5.0 based decision tree
classifiers to classify IRS-P6 AWiFS data and reported very high accuracy. Duro et al. (2012)
compared decision tree, support vector and random forest methods for the classification of
agricultural landscapes using SPOT-5 HRG imagery in both pixel and object oriented domain and
found for the specific case study, all the algorithms equally performed.
Fuzzy set classification logic takes into account the heterogeneous and imprecise nature (mix pixels)
of the real world. Proportion of the m classes within a pixel (e.g., 10% bare soil, 10% shrub, 80%
forest). Fuzzy classification schemes are not currently standardized.
To use the fuzzy approaches detailed ground data are required. While this inevitably increases the cost
and complexity of an investigation the benefits accrued, particularly in terms of improved
representation and accuracy, must be considered. It is also worth stressing that detailed ground data
may be required for conventional classification analyses, as ground data for a pixel are supposed to
describe the class membership properties of the area on the ground represented by the pixel;
Performance and Comparison
Fuzzy approach holds advantages over both conventional hard methods and partially fuzzy
approaches, in which fuzziness in only the remotely sensed imagery is accommodated. It is usually
found that Kappa coefficients more than double when applying the fuzzy evaluation technique as
opposed to the hard evaluation technique.
Conlusion
Mixed pixels are common in remotely sensed data sets and depending on the Land cover mosaic on
the ground and the sensors spatial resolution, may dominate an image. It is therefore inappropriate to
use conventional hard classifier techniques to map land cover from remotely sensed data, as these
techniques are only appropriate for pure pixels. If remotely sensed data are to be used as a source of
land cover data, then the presence of pixels with multiple and partial class membership (incorporation
of contextual information) should be accommodated.
Techniques appropriate for pure pixels are often used in training and testing a supervised
classification. A fuzzy classification strategy may enable a suitable and effective classification of
remotely sensed imagery depicting inherently fuzzy phenomena and their evaluations, and allows for
locational and quantitative examinations of the misclassification in classified data. To use the fuzzy
approaches detailed ground data are required.
References
1. Decision tree approach for classification of remotely sensed satellite data using open source
support by Richa Sharma, Aniruddha Ghosh, and P K Joshi
2. A comparison of object-oriented and pixel-based classification methods for mapping land
cover in northern Australia. By T. Whiteside and Ahmad, W.
3. A fuzzy classification of sub-urban land cover from remotely sensed Imagery.
Int. j. remote sensing, 1998, vol. 19, no. 14, 2721 2738
4. Fully Fuzzy Supervised Classification of Land Cover from
Remotely Sensed Imagery with an Artificial Neural Network
G. M. Foody
5. Department of Geography, University of Southampton, Southampton, UK
An Evaluation of the ICM Algorithm for image Reconstruction
R. H. GLENDINNING
6. Contextual classification of remotely sensed data using MAP approach and MRF
R. Khedama, A. Belhadj-Aissaa.