Wetland Mapping Using Remote Sensing Imagery and ModelMap by pptfiles


									                  Xuan Zhu
Monash University, Australia
Remote Sensing of Wetlands
 Some wetland mapping studies have successfully
  utilised supervised rule-based or classification tree-
  based methods
 Typical classification tree-based methods include
  classification tree analysis (CTA), Random Forests
  (RF) and Stochastic Gradient Boosting (SGB)
Classification Tree Analysis
 To determine a set of if-then logical (split) conditions that
  permit accurate prediction or classification of cases via
  tree-building algorithms
 In image classification, recursively parsing the training
  observations in a form of binary partitioning based on the
  values of the selected explanatory variables such as spectral
  responses and ancillary data
Classification Tree Analysis (cont’d)
 The classification trees used as classification rules
 No assumptions are made regarding the underlying
  distribution of values of the explanatory variables
   can handle numerical data as well as categorical variables
 However, not necessarily produce the optimal classification
  tree, and the results may be adversely affected by
  inaccuracies and outliers in the training data
 Several ensemble classification methods: Random Forests
  (RF) and Stochastic Gradient Boosting (SGB)
Random Forests (RF)
 Developing multiple classification trees by selecting
  random subsets of the explanatory variables and
  original training data
    Each classification tree is trained on a re-sampled set of the
     original training data, and searches only a random sample of
     the explanatory variables for a split at each tree node
 For classification, each tree casts a unit vote for the most
  frequent class to the input data
 The output is determined by a majority vote of the trees
 Not sensitive to noise or overtraining
Stochastic Gradient Boosting (SGB)
 Start with a standard classification tree built from the
  entire sample of the training data, and use iterative re-
    the incorrectly classified training data are given higher weighting,
     which results in a new classification tree that emphasises the most
     difficult classification problems in the training data
 At each iteration, a classification tree is built from a random
  subset of the training data producing an incremental
  improvement in the classification
 Finally, all the small classification trees are stacked together as a
  weighted sum of terms
 The overall accuracy of classification improves progressively with
  each additional term
Comparison of RF and SGB
 Comparison of the two ensemble
 classification methods for wetland mapping
   using Landsat ETM+ imagery
   using the Jiuzhaigou Nature Reserve in China as
   a case study
 Conducted in ModelMap
   A classification tree-based predictive modelling
   package in R
The Jiuzhaigou Nature Reserve
 Located in the Min Mountains in
    the north of Sichuan Province
   Cover an area of > 700 square km
   Karst landforms
   Designated as a World Heritage
    Site and a World Biosphere Reserve
   The current human activity is
    mainly tourism
   Have all palustrine, riverine, and
    lacustrine wetland systems
   Not to classify these different types
    of wetland, but to focus on the
    separation of wetland from upland
    at a coarse level
 Data: Landsat-TM images acquired in September
  2004, DEM generated from the 1:100,000
  topographic map
 Variables: Six bands of the TM images (excluding
  Band 6 - the thermal band) as spectral variables,
  slope as an ancillary variable
 876 training sites were developed for five broad
  types of land cover
   wetland, woodland, pasture, dry land and
    glacier/permanent snow
 Identifying major types of land cover from the TM images
  and generating probability maps of each type of land cover
    Using six spectral variables and slope as the explanatory
     variables, RF and SGB models were built with ModelMap
 Combining the probability maps to derive wetland maps
    Five RF-derived and five SGB-derived probability maps
     respectively for wetland, woodland, pasture, dry land and
     glacier/permanent snow
    Combined via overlay to create the RF and SGB-derived land
     cover maps, in which each location was assigned to the type
     of land cover that has the highest probability at that location
    Each of the two land cover maps was finally reclassified to
     generate a wetland map
 Assessing mapping accuracies
Methods for Accuracy Assessment
 Four accuracy measures for the performance of RF and
 SGB for mapping the probability of presence of each
 type of land cover
   Sensitivity - measures the proportion of actual positives
    which are correctly identified
   Specificity - measures the proportion of true negatives which
    are correctly identified
   Kappa - measures the agreement on classification by taking
    into account the agreement occurring by chance
   The area under the ROC (Receiver Operating Characteristic)
    curve (or AUC) - a measure of overall performance of a
       A good classifier should have an AUC near 1, while a poor classifier
        has an AUC near 0.5.
Methods for Accuracy Assessment
 Standard error matrix measures for the accuracies of
 the RF-derived and SGB-derived wetland maps
   overall accuracy
   producers accuracy
   users accuracy
 The accuracy assessment data
    69 random points sampled in wetlands
    731 random points sampled in non-wetland areas
RF-derived probability maps
RF-derived probability maps
SGB-derived probability maps
SGB-derived probability maps
Accuracy measures for each type of land cover with RF and SGB
       (the thresholds were chosen to maximise Kappa)

                 Kappa        Sensitivity    Specificity        AUC
Land Cover
               RF     SGB    RF      SGB    RF      SGB    RF     SGB

  Wetland      0.99   0.99   0.98    0.99   0.99    0.98   0.99   0.99

 Woodland      0.65   0.66   0.79    0.90   0.87    0.77   0.91   0.91

  Pasture      0.53   0.57   0.62    0.63   0.95    0.95   0.89   0.89

  Dryland      0.12   0.11   0.25    0.32   0.92    0.88   0.69   0.70

Glacier/snow   0.84   0.86   0.90    0.87   0.94    0.97   0.97   0.98
Wetland Maps

     RF-Derived   SGB-Derived
 In individual land cover mapping
   RF and SGB achieved a similar level of agreement on
    classification on all types of land cover
   Both performed well in estimating probabilities of the
    presence of wetland and glacier/permanent snow,
    moderately well for pasture and woodland, but poorly in
    predicting the presence of dryland
 In the final wetland mapping
   Both achieved high overall accuracies and class
   SGB produced more commission errors than RF

To top