Rogan J, Miller J, Stow D, et al. JUL 2003. Land-cover change monitoring with classification trees using Landsat TM and ancillary data. PE&RS 69 (7): 793-804.

Questions:

1.      Where did ancillary data contribute little to change map accuracy, what factors may have caused this, and why use ancillary data then if it contributed so little to two of the three levels of classification?
Ancillary data improved change map accuracy only two percent in Level 1 and 2.  Level 1 has only two classes (change and no change) and level 2 has only four classes so spectral data provided enough discriminatory information.  Level 3 is the most complicated measurement space and ancillary data shows their importance in discriminating land-cover change with a large number of classes.  Ancillary data increases dimensionality and helps separate classes.

2.      Why was Landsat TM Multitemporal Kauth Thomas (MKT) linear transformation used?   The MKT produced six features of interest:  three features representing change in brightness, greenness and wetness, and three features of mean or stable brightness, greenness and wetness between image dates.  MKT is like multitemporal principal component analysis (PCA) where major components are termed stable components and minor components are termed change components but recent research shows MKT to be superior to PCA,  MKT produces biophysical based features versus scene-dependent components from PCA.

  1. Why is it necessary for the number of training samples per class to be kept roughly equal when used in a classification tree?
    Classification trees are sensitive to large discrepancies in the number of training samples among individual classes. Classes with a larger number of training pixels might have greater weight in the analysis.

 

  1. Classification trees are a particular type of machine learning algorithm.  What are 2 factors (There are more than 2) contributing to the success of machine learning classifiers in resolving land-cover and land-cover change classes from often complex measurement spaces?
    1. Due to their non-parametric natures, they deal well with multi-modal, noisy and missing data.
    2. They can readily accommodate both categorical and continuous ancillary data
    3. They allow users to investigate the relative importance of input layers in contribution to classification accuracy.
    4. They are flexible and can be adapted to improve performance for particular problems.

 

 

Object-based Analysis of Ikonos-2 Imagery for Extraction of Forest Inventory Parameters

 

 

Q1:      If the spectral response is represented in digital imagery as a series of discrete pixels covering a wide range of spectral values, for forest inventory purposes, the stand is interpreted as a single homogenous polygon; a fundamental problem of incorporating digital imagery into the forest inventory process.

            What do the authors give as one solution in approaching this problem?

A:         One solution is to aggregate the individual pixels representing the forest stand into an ‘image object’ represented spectrally as the combined response of all underlying pixels. The image objects become the carriers of image information and form the basic units of subsequent analysis. [Pg 383]

 

Q2:      A pixel typically contains a vector of information representing each band or layer in a data set where as digital imagery’s  spectral response information is related as digital numbers (DN). In contrast , image objects are composed of _____________, enable the calculation of aggregative __________, such as ____________ and _________ ________, from an object’s underlying Dns. [Pg 384]

            What words complete the above question?

A:         Multi-pixel Groups, Statistics, Mean, Standard Deviation.

 

Q3:      What is ‘binary recursive partitioning’ in Tree Models? [Pg 384]

A:         When data is split repeatedly into increasingly homogeneous groups, or nodes, using combinations of explanatory variables that best distinguish the variation of the response variable.

 

Q4:      The authors generated image objects through an image segmentation procedure in eCognition termed ‘multiresolution segmentation.’ The authors influenced the output of the segmentation process through specification and weighting of input data and definition of parameters affecting the size, spectral homogeneity, spatial homogeneity, and shape of the resulting image objects. Bands 1,2, and 3 were equally weighted against band 4 and the panchromatic band. [Pg 385]

            What were the three objectives of the segmentation phase, how many images were produced and what was the scale for each?

A:         (a) to delineate homogeneous forest stand components, (b) to isolate field plot locations within these areas; and (c) to characterize the resulting image objects in terms of image texture. Three image object levels were created, the first at a scale of (166), second 50% finer scale than original, and third (1/8)th of the original scale.

 

Q5:      What is the over-all accuracy result from Table 2 Land Cover-Classification? [Pg 389]

A:         0.93, According to J. Duh this is to good to be true!