Ensemble Classification of Multispectral Remote Sensed Images for Change Detection

Download Full-Text PDF Cite this Publication

Text Only Version

Ensemble Classification of Multispectral Remote Sensed Images for Change Detection

Mr. S. Deepan

PG student, ECE Department

Sri Venkateswara College of Engineering Pennalur, Sriperumbudur Tk,

Mrs. D. Menaka

Assistant Professor,ECE Department, Sri Venkateswara College of Engineering

Pennalur, Sriperumbudur Tk,

Abstract The objective of the proposed work is to detect and process multispectral remote sensed images and classify them effectively using an ensemble classifier over a region. Thus, analyzing the changes present in that region over a period of time using machine learning methods. (As the traditional methods like surveying is inefficient in analyzing the difficult terrains such as forest cover). The ensemble classifier which is proposed in this work is that of a decision tree or E-Tree (Ensemble Tree classifier), effectively handles large volumes of data. It constantly updates the new classifier, thereby discarding the old ones. First, classification of forest areas has been proposed. A region of interest is selected and processed resulting in data sets to form a dictionary. The dictionary of data sets is interactively made to learn at classifier (Decision E-Tree) for further processing. The accuracy for the proposed classifier needs to be analyzed and compared with the other commonly used classifiers such as SVM, Subspace Discriminant etc. The performance of classifier is analyzed by metrics such as Accuracy, Scatter Plot, Receiver Operating Characteristics curve. Training of different data sets at different locations of forest areas across the globe has to be carried out, for further increase in the accuracy of the classifier. Test images over different periods of time is given as input to the classifier, a Region of Interest is selected and it is predicted based on the trained classifier and the variations in the landscapes is efficiently detected.

Keywords Change Detection, DecisionTree,Classifier, Remote Sensed Images,Scatter Plot,Receiver Operating Characteistics.


    Climate change across the globe resulting in severe changes over the land terrains across the world. a serious threat to mankind. Hence, analyzing the land covers over a period of time has become essential in order to maintain an ecological balance. Traditional methods (e.g, field surveys, map interpretations) are not effective enough to process land covers and terrains over a wide area such as forest covers and the yet- to be unraveled areas by mankind across the globe.

    Change Detection (CD) involves the use of multi temporal data sets to discriminate areas of land cover change between dates of imaging. In general, these methods can be categorized into two types: The first type covers those methods detecting binary change/no-change information. This performs CD without additional thematic information and produces a CD map in which changed areas are separated from unchanged ones. Here, the comparison is performed directly on the spectral data, resulting in a difference image which is analyzed

    to separate insignificant from meaningful changes. This category includes image differencing, image rationing, vegetation index differencing, and principal component analysis (PCA), among others .

    The second type of CD covers those approaches detecting detailed fromto change, such as post classification comparison, CVA, and hybrid CD methods based on the intensity of pixel values. The basis of CD is quantifying and labeling changes in the spectral space.

    The first known efficient work on change detection is done by differencing two remote sensed images at different time instants and statistical analysis was carried out[1].In order to analyze difficult land terrains like forest regions, an effective Artificial Neural Network framework was proposed to estimate the changes[2]. A regression model estimating radar backscattering amplitude over the forest region of Southern Sweden prior to the storm was developed and compared after the storm had affected the same region to estimate the changes[3]. Principle Component Analysis (PCA) and K-means Clustering techniques are employed (unsupervised method) in order to analyze the change detection. The difference or changed image over a period of time, is represented in the vector space and clustering technique is used to study the variations[4]. Hyperspectral remote sensing imagery at different time instants and resolutions are analyzed for variations by the proposed classification approach, a Support Cluster Machine (SCM) [5]. An entropy based approach is designed for analyzing the multispectral images and the classification approach is compared with maximum likelihood, parallelepiped, minimum distance classification etc.,[6]. The recent change detection approach proposed is that of a ensemble classification of multilayer perceptrons (Artificial Neural Network approach) to analyze the change detection[7]. A survey on the existing change detection approach, their characteristics, efficiencies, limitations etc., has been reported[8].


    1. MultiSpectral Image

      A multispectral image is one that captures image data at specific frequencies across the electromagnetic spectrum. The wavelengths may be separated by filters or by the use of instruments that are sensitive to particular wavelengths,

      including light from frequencies beyond the visible light range, such as infrared. Spectral imaging can allow extraction of additional information the human eye fails to capture with its receptors for red, green and blue. It was originally developed for space-based imaging. Satellite images are multispectral in nature.

    2. .Decision Tree Classifier

    Decision tree learning uses a decision tree as a predictive model which maps observations about an item to conclusions about the item's target value. It is one of the predictive modeling approaches used in statistics, data mining and machine learning. Tree models where the target variable can take a finite set of values are called classification trees. In these tree structures, leaves represent class labels and branches represent conjunctions of features that lead to those class labels. Decision trees where the target variable can take continuous values(typically real numbers)are called regression trees.

    C Intensity Based Feature Extraction

    Intensity based feature extraction in multi spectral image specifies the retrieval of intensity levels across multiple bands (red, green, near infra red, visible light etc..) over the specified region. Each region in a multi spectral image correlates with a specified range of intensity levels over a particular band. Hence, intensity based feature extraction enables to determine the specified region, based on the trained data(spectral levels across green, near infrared and visible light).


  1. Satellite Imagery:

    Satellite imagery is defined as the collection or archiving of images of earth by satellites orbiting around it. These are multispectral spectral in nature, providing vital information of any region across the earth.

  2. Selecting Appropriate Bands For Training:

The training data set used describes the remote sensing study which mapped different forest types based on their spectral characteristics at visible-to-near infrared wavelengths, using ASTER satellite imagery. The output (forest type map) can be used to identify and/or quantify the ecosystem services (e.g. carbon storage, erosion protection) provided by the forest. (Courtesy: https://archive.ics.uci.eu/ml/datasets/Forest+type+mapping )

The proposed work is illustrated in the form of block diagram as shown below:

Satellite Imagery

Selecting Appropriate Bands for Training

Decision Tree Classification

Performance and Accuracy Measurements

Fig 1. Block Diagram For The Proposed Work

  1. Decision Tree Classification:

    Decision tree classifier is designed based on the following parameters:

    • Maximum Number of Splits: Maximal number of decision splits (or branch nodes) per tree.

      Max Splits size= (X,1)-1 (1)

    • Minimum Parent Size: Each splitting node in the tree has at least 'Min Parent Size' observations.

      Min Parent Size=max(Min Parent Size,2*Min Leaf Size)


    • Predictor Names: Matrix of predictors used to train the decision tree. It is represented as [X]

    • Response Names: Name of the response variable. It is represented as [Y].

    • Pruning: Pruning is a technique in machine learning that reduces the size of decision trees by removing sections of the tree that provide little power to classify instances. Pruning reduces the complexity of the final classifier, and hence improves predictive accuracy.

    • Validation Accuracy: Validation Accuracy is an accuracy estimation to determine the performance of classifier.

      Validation Accuracy = 1 k fold Loss (partitioned Model, 'Loss Fun', 'Classification Error') (3)

      1. (b) (c)

        (d) (e) (f)

        Fig 2.1 (a)-(f) Scatter plot for different band levels (b4-b7,b1-b4,b2-b5,b5-b8,b3-b6,b6-b9.

        • Score Transform Symmetric

    Score Transform function, specified as the comma- separated pair consisting of 'Score Transform' and a function handle for transforming scores. The function should accept a matrix (the original scores) and return a matrix of the same size (the transformed scores). The classified decision tree is depicted in the below figure 3.

    Fig 3 Decision tree classification layout for the trained data set

  2. Performance And Accuracy Measurements

  • Scatter Plot: Fig 2.1-1.6 illustrates the scatter plot

    for the different band levels (b4-b7,b1-b4,b2-b5,b5-b8,b3- b6,b6-b9) where, b1 represents green band,b2 represents red band and b3 represents near infrared during September 26,2010. b4 represents green band,b5 represents red band and b6 represents near infrared during March 19,2011 .b7 represents green, b8 represents red and b9 represents near infrared band levels during May 8,2011.

    Also, In the figure 2.1-2.6, s- represents evergreen forests, ,h- deciduous forests, d- mixed deciduous forests and o- other barren lands. (s,h,d,o- responses; b1-b9- predictors for the classifier).

  • Receiver Operating Characteristics:

    Fig 4 Receiver Operating Characteristics (ROC) curve for the trained data sets.

    Area under the curve for the selected data set is found to be 0.8923 (89.23%). Fig 4 depicts the ROC curve for the decision tree.

  • Validation Accuracy: Validation Accuracy with respect to other classifiers is depicted in the below table.


Classifier Used


Accuracy (%)


Coarse Gaussian SVM



Medium KNN



Cosine KNN



Decision Tree


Table 1 Validation Accuracy analysis of decision tree with respect to other classifiers.


The test data sites used for the proposed work is described below:

  1. North And East Regions Of Alabama:

    The test images show the observed changes in forest cover north and east of Tuscaloosa, Alabama. They also show the change after an EF5 tornado caused massive destruction on April 27, 2011. The two Landsat images show the forested areas in 2000 and 2011. The tornado path can be clearly seen in the second image.

    1. (b)

      Fig 5 (a) & (b)- Multispectral test sites across the regions of Alabama in the years 2000 and 2011.

  2. Forest Regions Across Palani Hills:

    Another test site across the forest regions of Palani Hills dated March 22,2015 and December 25,2011.

    1. (b)

Fig 6 (a) & (b)- Multispectral test site 2 across Palani Hills in the year 2015 and 2011.

The two test sites are analyzed by selecting a region of interest from the images and then tested with the trained classifier to detect the type of forest regions.


Thus, the decision tree is trained with the data sets and then analyzed for its performance and accuracy measures which founds to hold better (validation accuracy=94.5%) when compared with other classifiers. Then, the trained classifier is used to detect the type of forest regions from the images chosen at different time instants.


  1. John D.Villasenor, Dennis R.Fatland, Larry D.Hinzman, Change Detection on Alaskas North Slope Using Repeat-Pass ERS-1 SAR Images,IEEE Transactions on Geoscience and Remote Sensing, Vol. 31, No. 1, January 1993.

  2. Sucharita Gopal, Curtis Woodcock, Remote Sensing of Forest Change Using Artificial Neural Networks,IEEE Transactions on Geoscience and Remote Sensing, Vol. 34, No. 2, March 1996.

  3. Johan E. S. Fransson, Fredrik Walter, Kristina Blennow, Anders Gustavsson, and Lars M. H. Ulander, Detection of Storm-Damaged Forested Areas Using Airborne CARABAS-II VHF SAR Image Data IEEE Transactions on Geoscience and Remote Sensing,Vol.40,No.10,October 2002.

  4. Turgay Celik.,Unsupervised Change Detection in Satellite Images Using Principal Component Analysis and k-Means Clustering, IEEE Transactions on Geoscience and Remote Sensing, Vol. 6, No.4,October 2009.

  5. Mingimin Chi., Qun Qian and J´on Atli Benediktson, Cluster Based Ensemble Classification For Hyperspectral Remote Sensed Images. IEEE Conference Volume.,1,I-209 – I-212.

  6. D.I.Long and Vijay P.Singh, An Entropy Based Multispectral Image Classification Algorithm. IEEE Trans., Vol.51., No.12., Pages:5225-5238,2013.

  7. Moumita Roy,Dipen Routaray, Susmita Goush And Ashish Goush., Ensemble Of Multilayer Perceptrons For Change Detection In Remote Sensed Images IEEE Trans., Vol.11,No.1., Pages: 49-53, 2014.

  8. Richard J.Radke,Srinivas Andra,Omar-Al-Kofahi and Badrinath Roysam, Image Change Detection Algorithms: A Systematic Survey IEEE Transactions on Image Processing Vol 14, No.3, March 2005.

  9. Forest Mapping Data Set: https://archive.ics.uci.edu/ml/datasets/Forest+type+mapping

Leave a Reply

Your email address will not be published. Required fields are marked *