Assessment of Current Field Plots and LiDAR

1
Assessment of Current Field Plots and LiDAR
Virtual Plots as Guides to Classification
Procedures for Multitemporal Analysis of Historic
and Current Landsat Data for Determining Forest
Age Classes

• Dr. William H. Cooke
• Department of Geosciences, GeoResources
  Institute,
• Mississippi State University

2
Objective

• The ultimate goal of this study is to map above
  ground biomass and relate the biomass estimates
  to the amount of carbon sequestered by even-aged
  pine stands.

3
Special Thanks toRobert Wallis (recently
employed)Chitra Prabhu (can you just do one
more thing)
4
Cooperative Study

• Curt Collins presented earlier this morning a
  multi-temporal analysis of Landsat data to
  determine forest age classes.
• Here, we present a alternative approach for
  determination of forest age classes using an NDVI
  threshold to discriminate forest from non-forest
  pixels followed by an unsupervised classification
  to further differentiate pines from other
  species.

5
Introduction

• This study is a preliminary data analysis study
  designed to provide a benchmark for comparisons
  of information derived from field data and recent
  Landsat ETM data alone with value added models
  developed using data from all sources.
• The study was undertaken to test the usefulness
  of field plots and Landsat TM data alone to
  estimate volume, which is closely related to
  above-ground biomass.

6
Introduction

• Results of these tests provide valuable
  information regarding the feasibility for using
  field plots and satellite data alone for volume
  estimation, and for improving classifications of
  historic Landsat imagery.

7
Questions of Interest

1. How well do NDVI thresholds discriminate forest
   from non-forest conditions
2. What proportion of field plots are necessary to
   optimize classification accuracies while
   maintaining a sufficient number of accuracy
   assessment plots
3. Are there optimum Euclidian distances that can be
   specified for signature generation based on field
   plots in a supervised classification approach
   and
4. What is the optimum combination of bands and band
   transformations for supervised classification of
   volume based on field plots?

8
Methods
9
Methods (Question 1)

• Preliminary studies helped determine that an NDVI
  threshold of 0.3 did a good job discriminating
  forest from non-forest pixels.
• Visual analyses for the results of the NDVI
  threshold are promising for the ETM scene that
  has been partitioned using this approach.
• An unsupervised classification was used to
  further differentiate pixels that were
  predominantly pine from other forest pixels.

10
(No Transcript)
11
Methods (Question 2)

• Once the scene was partitioned into pine pixels,
  the field plots were randomly drawn in equal
  proportion by age classes and used as seeds for
  generating volume signatures for a supervised
  classification process.
• Plots were grouped by forest age class and 10 of
  the plots in each age class were used to generate
  signatures.
• Since the distribution of plots by age class was
  relatively normally distributed around the 20-25
  year age class, the decision to sample
  proportionally by age class was predicated on the
  desire not to under-sample the low and high age
  classes.

12
Methods (Question 3)

• Euclidean distances of 10, 12, 14, 16, 18, and 20
  were chosen for analysis.
• Two replications are complete for tests of 34
  volume classes.
• Three replications are complete for 7 aggregated
  volume classes and also for 2 aggregated volume
  classes.

13
Methods (Question 4)

• Three replications are complete for the following
  bands and combinations
• - Raw bands only
• - NDVI only
• - 1st and 2nd Principle Components bands only
  
• - Raw bands NDVI
• - Raw bands 1st and 2nd Principle
  Components bands
• - Raw bands NDVI 1st and 2nd Principle
  Components bands

14
ResultsEuclidean Distances

• For 34 volume classes, for both draws,
  classification accuracy increases with Euclidean
  distance and reaches an optimum value at
  Euclidean distance of 18.

15
ResultsEuclidean Distances

• Classification accuracies for 7 volume classes
  reach an optimum value at Euclidean distance 18
  for draw 1. For draws 2 and 3, Euclidean distance
  16 gives the highest accuracy.

16
ResultsEuclidean Distances

• Results for 3 volume classes indicate that
  classification accuracies reach an optimum value
  for Euclidean distance 18 for draws 2 and 3, and
  16 for draw one.

17
ResultsEuclidean Distances

• For 2 volume classes, classification accuracies
  reach an optimum value at Euclidean distance 14.
  Additional replications are needed to determine
  whether the high classification accuracies for
  distance 14 are an anamoly.

18
ResultsMean by Draw

• Results for mean classification accuracies by
  draw indicate that Euclidean distance 14 gives an
  optimum value for 2 classes, Euclidean distance
  18 for 3 classes, and Euclidean distance 16 for 7
  classes. For 7 and 2 aggregated volume classes,
  Euclidean distance 18 gives the second highest
  accuracy.

19
ResultsBand and Transforms

• For 7 volume classes there are no improvements in
  the accuracies that accrue to the addition of
  transformed bands.

20
ResultsMean Band and Transforms

• Mean accuracy results for 7 classes indicate that
  addition of PC and NDVI transforms to the raw
  bands give a slightly higher accuracy compared to
  the raw bands alone. High classification
  accuracies for NDVI alone are significant since
  NDVI provides radiometric normalization
  advantages.

21
ResultsBand and Transforms

• For 3 volume classes, NDVI yields the highest
  classification accuracy for draw 2. For draw 1
  and draw 3, raw bands yield the highest
  classification accuracies.

22
ResultsRaw vs. NDVI
Tests for additional Euclidean distances were
performed to address scaling concerns of
transformed bands.
23
Conclusions

• A priori designation of Euclidean distances can
  help remove analyst subjectivity and enable
  front-end interface development where a text file
  of x,y coordinates is passed to the signature
  growing process.
• The ability to detect real change in biomass
  over time and over large geographic areas may
  depend on radiometric normalization.
• The results of these studies are too limited in
  geographic extent and number of replications to
  make definitive recommendation for automated
  processes.
• The real benefit of these studies is
  determination of the potential for automating
  processes in decision support systems that are
  designed for long-term monitoring of biomass.
• NDVI alone yields classification accuracies
  similar to raw bands and provides image
  radiometric normalization needed to track biomass
  changes over time.
• Euclidean distances can be chosen that optimize
  classification accuracies.