On October 23rd, 2014, we updated our
By continuing to use LinkedIn’s SlideShare service, you agree to the revised terms, so please take a few minutes to review them.
Molinier - Feature Selection for Tree Species Identification in Very High resolution Satellite Images.pptPresentation Transcript
Feature Selection for Tree Species Identification in Very High Resolution Satellite Images Matthieu Molinier and Heikki Astola VTT Technical Research Centre of Finland [email_address] , [email_address] IGARSS 2011 Vancouver, 28.7.2011
A 1.5-year study (2009-2010) funded by The Finnish Funding Agency for Technology and Innovation (TEKES), with Finnish Companies (forest) and Research Organizations (VTT and University of Eastern Finland UEF)
Improve methods for operative forest inventory from remote sensing data
Species-wise estimates (e.g. stem volume) not accurate enough (a ccuracy vs. cost)
NewForest approach in forest variable estimation Modelling based on satellite image pixel reflectances and contextual features Individual tree crown (ITC) detection and crown width estimation Combining data to predict total amount and size variation by species segmentation estimates Refined, more accurate species-wise estimates
Tree species classification : training data from 20 pure species field plots
Testing data – from 178 field plots (mixed species)
178 field plots acquired in 2009, limited spatial distribution (several plots per forest stand)
In total : 1164 ground objects mapped (276 pines, 277 spruces, 347 deciduous, 264 non-trees)
GeoEye image : 10.5 km x 11.5 km
Input for feature selection – 35 + 4 features R G B NIR PAN mean intensity within 1.5 m radius around tree candidates ( TC ) SPECTRAL (5) – set A CONTEXTUAL (9) – set B From PAN , 7.5 m radius around TC mean mean / median skewness kurtosis contrast pm1 : mean of brightest pixels ps1 : std of brightest pixels pm2 : mean of darkest pixels ps2 : std of darkest pixels SEGMENT-WISE (21) – set C From PAN , 3 segment sizes : 50 m 2 , 85 m 2 , 125 m 2 mean mean / median skewness kurtosis std : standard deviation pmean : partial mean pstd : partial standard deviation Probe variables random vectors or random permutations of a feature vector probe_gauss1 , probe_gauss2 probe_shuffle1 , probe_shuffle2
Class definitions and training scheme WHOLE DATASET (1164 samples) 900 trees, 264 non-trees TESTING (391) MODEL DESIGN (773) 2 / 3 1 / 3 TRAINING (512) VAL (261) 2 / 3 1 / 3 stratified sampling to preserve classes proportions model building ranking Class # Class name 1 pine 2 spruce 3 deciduous 4 shadow 5 open area / sunlit 6 bare ground 7 green vegetation Tree classes Non-tree classes
Feature selection preparation (Guyon et al., 2003)
Feature normalization to the range [0, 1]
Visual screening of scatter plots on the 35 real features : no obvious correlations, very few outlier samples
Variable ranking – assessing features one by one with the most simple classifier (single threshold), one(+) vs all(-) . 4 scores :
Fisher criteria F , scaled to [0 1]
R 2 – Pearson correlation coefficient for a single feature vs +/- labels
AUC : Area under ROC curve (Receiver-Operative Curve)
sum of previous scores ( FR2AUC )
All scores computed for every class, then averaged to rank the variables for all 7 classes and for tree classes only (1,2,3).
No single feature outperformed significantly and consistently the others
Feature selection and image classification
Classification accuracy on validation set VAL (261) as a score
Sequential Forward Selection ( SFS ) with three classification methods :
Linear Discriminant Analysis ( LDA )
k-nearest neighbor ( kNN ) classifier, k [2 9]. Feature selection and choice of k at the same time.
Find the best minimal feature subset by a brute-force approach
10 best features from the SFS
retrain the best model using all modeling dataset (TRAIN + VAL) and test with the independent TEST set
brute force approach tractable in this case with simple classifiers
overcome the sub-optimality of SFS
6-10 features is enough Spectral features performed best segment-wise features not suited to mixed species study Overall classification accuracy on tree classes over 80% Probe variables selected more often in the first places with LDA than with kNN : linear classifier too simple. Quadratic LDA was overfitting. kNN, k=5 best overall performance, and lowest difference from training to validation error => lower risk of overfitting
Example of tree species classification map pine : 76 % spruce : 76 % deciduous : 88 % non-forest
Pan-sharpened GeoEye image extract of 1 km x 1 km
Individual tree crown classification with 5-NN classifier trained with pure species training data
Non-forest mask generated with
k-means clustering + cluster labeling
Predicted species-wise stem numbers vs. field plot data Nspruce [stems/ha] Npine [stems/ha] Predicted [stems/ha] Ndecid [stems/ha]
Predicted stem number per species plot against test data (178 test plots)
Systematic under-estimation of predicted stem number with spruce and deciduous classes
Noise partly due the small collecting radius (r = 8 m) of test data, and to geolocation differences between satellite and ground data
0 500 1000 1500 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2200 True number of spruces/field plot Predicted number of spruces/field plot y=0.98*x + 137.1 y=0.98*x + 137.1 y=0.98*x + 137.1 R 2 = 0.24 y=0.98*x + 137.1 y=0.33*x + 239.8 y=0.56*x + 21.0 R 2 = 0.54 True number of broadleaved/field plot Predicted number of broadleaved/field plot y=0.85*x + 45.0 R 2 = 0.34 True number of pines/field plot Predicted number of pines/field plot 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
The methodology could detect individual treetops, identify their species and determine species proportions in mixed forest .
Feature ranking and feature selection was performed on a set of 35 features for tree species classification.
Several classifiers (model including a feature subset and a classification method ) were built. The best turned out to be 5-NN with a subset of 6 features, mostly spectral . Segment-wise features could be discarded.
The tree species proportion accuracy was good (1.4% to 3.5%), but the correlation of stem numbers / species not as good as expected.
Model selection with more elaborate classifiers (e.g. SVMs)
Embedding feature selection into a cross-validation scheme
Improve stem number estimation with adaptive filtering
Tree crown width estimation validation with ground data