Presentation is loading. Please wait.

Presentation is loading. Please wait.

Classification and Regression Trees for Land Cover Mapping Lecture materials taken from a presentation by: B. Wylie, C. Homer, C. Huang, L. Yang, M. Coan.

Similar presentations


Presentation on theme: "Classification and Regression Trees for Land Cover Mapping Lecture materials taken from a presentation by: B. Wylie, C. Homer, C. Huang, L. Yang, M. Coan."— Presentation transcript:

1

2 Classification and Regression Trees for Land Cover Mapping Lecture materials taken from a presentation by: B. Wylie, C. Homer, C. Huang, L. Yang, M. Coan EROS Data Center, Sioux Falls Mnf1<=28 Mnf3<=19 Mnf13>56 Mnf1>19Mnf16<=54 decid. shrub Mnf3<=38 Mnf1<=28Mnf1<=25 Mnf3<=24 Mnf8<=28 shrub Mnf11<51 Mnf17>56 shrubcedar P. pine cedar Mnf2<=43cedar cedar

3 How do you eat an elephant? One bite at time! Divide&ConquerStratify&Predict

4 Specific Gravity Weight 6/6 = 100% 6/7 = 86% Separating Apples and Oranges (Lemons?) Classification Trees

5 Mnf1<=28 Mnf3<=19 Mnf13>56 Mnf1>19Mnf16<=54 decid. shrub Mnf3<=38 Mnf1<=28Mnf1<=25 Mnf3<=24 Mnf8<=28 decid. shrub Mnf11<51 Mnf17>56 shrubcedar decid. P. pinecedar Mnf2<=43cedar P. pine Example Decision Tree TrueFalse

6 Advantages of Decision Trees Rapid Rapid Repeatable Repeatable Nonparametric Nonparametric Utilize categorical data Utilize categorical data Non-linear relationships Non-linear relationships Less sensitive to errors in training data Less sensitive to errors in training data Disadvantages of Decision Trees Lots of training Lots of training Over-fitting Over-fitting Weights toward the relative % of training data Weights toward the relative % of training data Short-sighted (stepwise methods fail to identify optimal subset regressors –Fox,1991) Short-sighted (stepwise methods fail to identify optimal subset regressors –Fox,1991)

7 Other Methods: Unsupervised Clustering + Cluster busting can be time consuming + Cluster interpretation is subjective + Cannot include categorical inputs + Difficult to interpret if multiple date or if non spectral data include (DEM) + Parametric (assumes normal distribution) Supervised Classification + Parametric (assumes normal distribution) + Cannot include categorical inputs + Problematic multiple date or if non-spectral data include (DEM) + Difficult for large area applications

8 Other Methods: Unsupervised Clustering + Cluster busting can be time consuming + Cluster interpretation is subjective + Cannot include categorical inputs + Difficult to interpret if multiple date or if non spectral data include (DEM) + Parametric (assumes normal distribution) Supervised Classification + Parametric (assumes normal distribution) + Cannot include categorical inputs + Problematic multiple date or if non-spectral data include (DEM) + Difficult for large area applications Neural Nets + Long convergence times (training) + High CPU demands (training) + Grey box + Tricked by local minimums + Non-repeatability of results (random search functions) + Sensitive to errors in training data

9 Spectral variability: A monoculture of wheat Training data: capture variability of a class (sample size)

10 Predictor 1 Predictor 2 Would 2 examples of each produce a reliable separation?

11 Training samples Classification tree is a “Data Mining” method so it performs well with large training data sets. Sampling of classes should reflect their relative frequency in the study area. rare classes = few training points common classes = many training points Adequate but not over sampling of rare classes Samples should be widely distributed over the study area to minimize autocorrelation effects on the surface and on the image

12 Descriptive or Prediction decision trees? (De’ath and Fabricus 2000) DESCRIPTIVE TREES: 1) Usually a single tree 2) The objective is to understand important factors or functional relationships between predictor and dependent variables. 3) The decisions generated by the tree are as important as the predictions

13 PREDICTION TREES: 1)The objective is the Best Possible Prediction of the dependent variable 2)Usually consists of a combination of multiple trees aimed at producing higher accuracies and more stable and robust models (DeFries and Chan 2000)

14 Multiple Tree Approaches: Prediction 1)Bagging (bootstrap sampling of training data)--Splus & C5 2)Subset data layers—Splus & C5 3)Boosting * – C5

15 2) Subset of data layers—Splus & C5 Multiple Tree Approaches: Prediction soils spectral LUDA Tree 1 Tree 2 Tree 3 VOTEVOTE

16 3) Boosting (iterative trees try to trees try to account for account for previous trees previous trees errors)—C5 errors)—C5 Different over-fitting issues associated with each tree tend to be averaged out. Multiple Tree Approaches VOTE

17 Single tree Boosted

18 Boosting versus Single Tree (USGS NLCD Zone 16, Independent Test Data)

19 Mnf1<=28 Mnf3<=19 Mnf13>56 Mnf1>19Mnf16<=54 decid.shrub Mnf3<=38 Mnf1<=28Mnf1<=25 Mnf3<=24 Mnf8<=28 decid.shrub Mnf11<51 Mnf17>56 shrubcedar decid. P. pine cedar Mnf2<=43cedar cedar At each “terminal node” or “leaf” we: + know the number of training points correct, incorrect, and % right and % right Trees provide more than just land cover predictions

20 “leaf” map land cover % right or confidence Identifies terminal node number for each decision Model output Since each terminal node is coded with a probability, a confidence map can be generated

21 1)Reduce inputs to decision tree 2)Reduced tree may have improved accuracies 3)Increases speed that the tree can be applied to the study area 4)Interpretation of underlying functional relationships (drivers) 5)Produce multiple trees for class “voting” Use relative frequency of use of data layers in the training data as a crude index of data layer “utility”. Top “utility” data layers from 40 possible input layers

22 Accuracy Assessment: Cross Validation versus Independent Test, Zone 16, Utah

23 Uses of Cross Validation Accuracy Assessment Accuracy Assessment Optimal tree data sets Optimal tree data sets Pruning Pruning All training data used for prediction All training data used for prediction Cautions spatial autocorrelationspatial autocorrelation look for “significant” error changes when pruning or selecting tree parameterslook for “significant” error changes when pruning or selecting tree parameters

24 Past Experiences: hierarchical implementation of trees Landsat 7 ETM+ Mosaic (band 5,4,3) Mapping Zone 60, Spring, 2000 and 2001

25 Forest and Non-Forest Classification Mapping Zone 60, 2001 Established a classification tree model for mapping forest and non-forest class using 1700+ FIA plot data (669 forest and 1100+ non-forest plots). The classification was run using a 5-fold cross- validation procedure. The agreement between mapped and reference/validation data is 95% with standard error (SE) less than 1.0%.

26 Forest Classification Based on NLCD 2000 Classification System Mapping Zone 60, 2001 Established a classification tree model for mapping three MRLC forest classes using 669 FIA plots and 5-fold cross-validation procedure (134 plots for validation for each of the 5 runs). The agreement between mapped and reference/validation data is 80% with SE 1.0%

27 Forest Type Group Classification Based on USFS classification System Mapping Zone 60, 2001 Established a classification tree model for mapping six forest type groups using 669 FIA plots and a 5-fold cross-validation procedure (134 plots for validation for each of the 5 runs). The agreement between mapped and reference/validation data is 65% with SE 2.3%

28 Leaf-on Landsat 7 ETM+ scene mosaic (bands 5,4,3) for mapping zone 16 – Utah/Southern Idaho Leaf-on Landsat 7 ETM+ scene mosaic (bands 5,4,3) for mapping zone 16 – Utah/Southern Idaho

29 Forest/non-forest classification for mapping zone 16 – Utah/Southern Idaho Forest/non-forest classification for mapping zone 16 – Utah/Southern Idaho Missclassification 0 * 16.8% 0 * 16.8% 1 * 18.8% 1 * 18.8% 2 * 19.1% 2 * 19.1% 3 * 19.4% 3 * 19.4% 4 * 16.3% 4 * 16.3% Mean 18.1% Mean 18.1% SE 0.6% SE 0.6%

30 Deciduous/evergreen/ mixed classification for mapping zone 16 – Utah/Southern Idaho Deciduous/evergreen/ mixed classification for mapping zone 16 – Utah/Southern Idaho Fold Decision Tree ---- ---------------- Size Errors Size Errors 0 * 19.7% 0 * 19.7% 1 * 19.9% 1 * 19.9% 2 * 18.5% 2 * 18.5% 3 * 19.4% 3 * 19.4% 4 * 20.8% 4 * 20.8% Mean 19.7% Mean 19.7% SE 0.4% SE 0.4%

31 Forest type group classification for mapping zone 16 – Utah/Southern Idaho Fold Decision Tree ---- ---------------- Size Errors Size Errors 0 * 31.9% 0 * 31.9% 1 * 37.3% 1 * 37.3% 2 * 36.8% 2 * 36.8% 3 * 40.0% 3 * 40.0% 4 * 31.9% 4 * 31.9% 5 * 37.8% 5 * 37.8% 6 * 33.0% 6 * 33.0% 7 * 35.7% 7 * 35.7% 8 * 35.5% 8 * 35.5% 9 * 32.3% 9 * 32.3% Mean 35.2% Mean 35.2% SE 0.9% SE 0.9%

32 Mean and standard error (in parenthesis) of the overall accuracy (%) of classifications developed using a descriptive tree and an hierarchical tree approach in 5 repeated experiments. (Zone 16)

33 Recipe for Success Adequate and representative training (adequately represent rare classes, preserve relative proportions of training and population)Adequate and representative training (adequately represent rare classes, preserve relative proportions of training and population) Model construction assessed with Cross Validation Model construction assessed with Cross Validation (boosting, pruning, and data layer exclusions) Multiple trees for mapping (boosting) Multiple trees for mapping (boosting) Visually inspect land cover and add training in miss-classified areas and reconstruct modelVisually inspect land cover and add training in miss-classified areas and reconstruct model Consider “hierarchical” trees to allow trees to focus on problematic separationsConsider “hierarchical” trees to allow trees to focus on problematic separations

34

35 Edgematching

36

37 ACCURACY ASSESSMENT Accuracy assessment is based on a 20% sample of training sites for each mapped type.Accuracy assessment is based on a 20% sample of training sites for each mapped type. Since the 20% sample is not independent, this is not considered a true accuracy assessment.Since the 20% sample is not independent, this is not considered a true accuracy assessment. The 20% sample or validation set is not used in the classification tree model development.The 20% sample or validation set is not used in the classification tree model development. In addition to a validation with a 20% sample, we have supported our findings with a manual 5-fold cross validation with sample replacement to test model stability and help understand the effects of validation sample size on the model stability results.In addition to a validation with a 20% sample, we have supported our findings with a manual 5-fold cross validation with sample replacement to test model stability and help understand the effects of validation sample size on the model stability results.

38 Each site is considered correctly classified if the majority of pixels agree with sample polygon

39 Validation: Lower Wasatch Range MZ

40 Summary Statistics, 5-fold Validation This is a 5-fold cross validation with sample replacement to test model stability and help understand the effects of validation sample size on the model stability results.

41 Sample Size vs Accuracy and Coefficient of Variation

42 % of area mapped vs. available sample size (Great Salt Lake Desert)

43 % of area mapped vs. accuracy level

44 < 20 samples > 20 samples > 40 samples

45 Conclusions Current validation shows between 60-65% “accuracy” for completed mapping zones.Current validation shows between 60-65% “accuracy” for completed mapping zones. Variance of accuracies is directly linked to validation sample size.Variance of accuracies is directly linked to validation sample size. A validation set of 40 points seems to provide some stability to model results.A validation set of 40 points seems to provide some stability to model results. A minimum of 40 samples for validation per cover type should be a target size. This is difficult due to the preponderance of “rare” types and the lack of training sites.A minimum of 40 samples for validation per cover type should be a target size. This is difficult due to the preponderance of “rare” types and the lack of training sites. Using a 20% validation set, we need a minimum of 200 field sites per mapped type. This is difficult given the availability of unique sample sites for a particular cover type.Using a 20% validation set, we need a minimum of 200 field sites per mapped type. This is difficult given the availability of unique sample sites for a particular cover type. Overall map “accuracy” seems to coincide with expected results given the map detail.Overall map “accuracy” seems to coincide with expected results given the map detail. Rare ecological systems and systems that consist of a wide range of cover types (I.e. shrub steppe types) tend to lower accuracies.Rare ecological systems and systems that consist of a wide range of cover types (I.e. shrub steppe types) tend to lower accuracies.

46 A STRATEGY FOR ESTIMATING TREE CANOPY DENSITY USING LANDSAT 7 ETM+ AND HIGH RESOLUTION IMAGES OVER LARGE AREAS Chengquan Huang*, Limin Yang, Bruce Wylie, Collin Homer Raytheon ITSS EROS Data Center, Sioux Falls, SD 57198, USA


Download ppt "Classification and Regression Trees for Land Cover Mapping Lecture materials taken from a presentation by: B. Wylie, C. Homer, C. Huang, L. Yang, M. Coan."

Similar presentations


Ads by Google