Copyright (C) Mitsubishi Research Institute, Inc. MITSUBISHI RESEARCH INSTITUTE, INC. A Methodology of Forest Monitoring from Hyperspectral Images with Sparse Regularization A Methodology of Forest Monitoring from Hyperspectral Images with Sparse Regularization Jul. 26, 2011 Keigo YOSHIDA, Takashi OHKI, Masahiro TERABE, Hozuma SEKINE (MRI) Tomomi TAKEDA (ERSDAC) IGARSS 2011, Vancouver TU4.T08.1: Hyperspectral Monitoring of the Environment I
27
Embed
A Methodology of Forest Monitoring from Hyperspectral ......A Methodology of Forest Monitoring from Hyperspectral Images with Sparse Regularization A Methodology of Forest Monitoring
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Copyright (C) Mitsubishi Research Institute, Inc.
MITSUBISHI RESEARCH INSTITUTE, INC.
A Methodology of Forest Monitoring fromHyperspectral Images with Sparse RegularizationA Methodology of Forest Monitoring fromHyperspectral Images with Sparse Regularization
Modeling is not easy for several sensor dataof different physical property
Statistical or Data-driven approach is needed
Sensor fusion
Reflect diverse propertyof targets
Hard to bring out potential of big sensor data
[e.g.] NDVI use just 2 bands, or Red and IR& have to select optimal band combinations
Complexity of prediction model increases,resulting in poor prediction performance
Dimension is high but sample size is smalldue to limitation of field survey
This causes model overfitting
Hyperspectral sensor
provide detailed optical info.on forest physiognomy
growth situation character of tree species
etc.
ChallengesSensing Tech.
4Copyright (C) Mitsubishi Research Institute, Inc.
Research Outline
Utilize rich data by a machine learning technique (sparse regularization)and achieve accurate, informative, & less costly forest monitoring
Utilize rich data by a machine learning technique (sparse regularization)and achieve accurate, informative, & less costly forest monitoring
Output DataOutput Data
MethodologyMethodology
Remote Sensing Data Fusion(CASI-3 hyperspectral images + SAR signals)
Field Survey ResultsInput DataInput Data
Sparse Regularization
(Sparse Discriminant Analysis、LASSO regression)
Predicted Stand Factors of each subcompartmentsfor Forest Management
(Species, Canopy cover, Timber volume, Tree height)
Prediction Models
Subcompartment: a general spatial unit for forest monitoring
5Copyright (C) Mitsubishi Research Institute, Inc.
Target Site
Town-owned forest in Shimokawa,Hokkaido, Japan
Approx. 90 % of town is covered by forest Utilize local conifer resources for business Environmental model city for low-carbon society
Shimokawa
6Copyright (C) Mitsubishi Research Institute, Inc.
Remote Sensing Data
Species and Canopy-cover prediction: CASI-3 hyperspectrumVolume and Height prediction: Data fusion (CASI-3 + PALSAR)
Species and Canopy-cover prediction: CASI-3 hyperspectrumVolume and Height prediction: Data fusion (CASI-3 + PALSAR)
Remote Sensing Hyperspectral sensor (optical property) Airborne hyperspectral imager CASI-3 84 bands from 400 to 1060 nm (wavelenght res. : 8 nm) Original spatial res.: 2.0 m
→ Resolution is decreased to 30m to simulate satellite-based operation
PALSAR (shape or volume property) Microwave backscattering resizedorg. image
7Copyright (C) Mitsubishi Research Institute, Inc.
Field Survey
During aircraft obs., conduct field survey to collect data for modeling & validationDuring aircraft obs., conduct field survey to collect data for modeling & validation
Field Survey: Place 25-sq-m quadrats Inventory study for trees whose DBH > 5cm & and record tree species Canopy cover measurement with whole-sky camera Height measurement for sampled 10 trees
8Copyright (C) Mitsubishi Research Institute, Inc.
What is Sparse Regularization ? Why Do I Use it ?
“Sparse” means the model has a low # of nonzero parameters
■ Optimal Band Selection
Ineffective parameters will be removed from prediction modelautomatically by solving convex optimization problem
■ Higher Generalization Capability
simple model with smaller # of bands achieves lessoverfitting; better prediction performance
■ More Interpretable Model
9Copyright (C) Mitsubishi Research Institute, Inc.
norm(penalty)
Sparse Regularization:Theoretical Overview
Add penalty to loss function to obtain model with small num. of variablesAdd penalty to loss function to obtain model with small num. of variables
LASSO (R. Tibshirani et al., 96)
Optimal Scoring (T. Hastie et al., 94)
Perform Fisher’s linear discriminant analysis as regression by score
convert categorical variables for class membership into quantitative
Optimize and weight vector simultaneously
Loss function (LS)
10Copyright (C) Mitsubishi Research Institute, Inc.
Intuitive Explanation of Sparse Regularization
To reduce empirical errors,W moves away from 0,then penalty increases
<penalty>
L1-norm: attraction force to 0 is const.-> Small values in W tend to be 0
L2-norm: attraction force is small around 0-> Small values in W remain
L1-regularization L2-regularization
Coefficients
<attraction force to 0>
11Copyright (C) Mitsubishi Research Institute, Inc.
Experimental Flow
2. Prediction for Subcompartments
1. Modeling
LASSO Regression
Forest Pixel Extraction
Subcompartment Prediction
Hyperspectral Reflectance(ave. w/in each quadrat)
Hyperspectral Reflectance(ave. w/in each quadrat)
Hyperspectral Reflectance(30m x 30m pixels)
Hyperspectral Reflectance(30m x 30m pixels) Semisupervised LDA
PredictionPerformancePrediction
Performance
RegressionModel
RegressionModel
Forest PixelsForest Pixels
PredictedForest Condition
PredictedForest Condition
Averaged Reflectancew/in each Subcomp.
Averaged Reflectancew/in each Subcomp. Obtained Model
Sparse LDA ClassificationModel
ClassificationModel
PALSAR SignalsPALSAR Signals
12Copyright (C) Mitsubishi Research Institute, Inc.
Variety in a Subcompartment
(Subcompartment)
There is a large variety inside a subcompartment
Non-forest area• Deforestation area• Canopy gaps
Invading woods other than planted species• they’re not recorded on forest register
There is a large variety inside a subcompartment
Non-forest area• Deforestation area• Canopy gaps
Invading woods other than planted species• they’re not recorded on forest register
13Copyright (C) Mitsubishi Research Institute, Inc.
Dataset: Target category: 4 species
Larix kaempferi, Abies sachalinensis, Picea glehnii, other Broadleaf
Source Hyperspectral reflectance by CASI-3
84 bands, 400 – 1060 nm
9 signals given by PALSAR data polarimetries (HH/HV/VV) Three scattering components proposed by Freeman
14Copyright (C) Mitsubishi Research Institute, Inc.
Experimental Setting (2/2)
Validation: 100 times iteration of 5-fold cross valiadtion
Comparison: Methodology Classification Spectral Angular Mapper ; SAM Regularized Discriminant Analysis; RDA (L2-norm regularization) ν-Support Vector Machines; SVM (w/ Linear and RBF kernel)
RegressionPartial Least Squares; PLS
Input data pseudo multi-spectral imageASTER image simulated from CASI-3 data 3 bands 760 - 860Band 3
630 – 690Band 2
520 – 600Band 1
wavelength range (nm)#
15Copyright (C) Mitsubishi Research Institute, Inc.
■ Prediction accuracy (Mean of 100 times 5-fold CV)
0.0 %6.2 %1.8 %92.0 %Larix
0.8 %4.3 %16.3 %78.8 %0.1 %15.1 %84.4 %0.4 %
Abies0.0 %0.3 %76.4 %23.3 %
PredictedUpper:HyperspecLower: Multi-spec
Actual
BroadleafPiceaAbiesLarix
67.1 %0.0 %0.2 %32.7 %86.4 %6.0 %0.0 %7.6 %
Braodleaf
0.0 %77.6 %0.0 %22.4 %2.8 %70.8 %6.2 %20.2 %
Picea
SDA achieves highest performance (87% prediction accuracy)Results by using hyperspectral images outperform pseudo-images
SDA achieves highest performance (87% prediction accuracy)Results by using hyperspectral images outperform pseudo-images
■ Confusion Matrix(SDA, Mean of 100 times 5-fold CV)
84.0
SVM(Lin.)
83.0 %82.2 %69.0 %87.0 %CV-Accuracy
SVM(RBF)RDASAMSDAMethod
Result: Quadrat Species Classification
16Copyright (C) Mitsubishi Research Institute, Inc.
Result: Quadrat Stand Factor Regression
LASSO with hyperspectral data provides best performance for all stand factorLASSO with hyperspectral data provides best performance for all stand factor
RMSE(10-fold CV)
Canopy Cover Timber Volume Tree Height
poor results
Prediction(10-fold CV)
17Copyright (C) Mitsubishi Research Institute, Inc.
Grass & Bamboo
Bare ground
Cloud
Clouds
deforestation
Sparse forest
Validation with 39 pixels selected manually Forest vs. non-forest: 100 % Overall Accuracy: 97.4 % (38/39)
※ Dots indicate top-leftpoint of each pixel
Result: Forest Extraction by Semi-supervised LDA
18Copyright (C) Mitsubishi Research Institute, Inc.
Invading broadleaf trees were found
Abies sachalinensis
Broadleaf
Confirm consistency between actual and estimated species by field surveyConfirm consistency between actual and estimated species by field survey
30 m resized pixels
Original CASI-3
Picea glehnii
Larix kaempferi
Abies sachalinensis
Natural Broadleaf
Legend
▲ Field Survey Points
※ Dots indicate top-leftpoint of each pixel
Result: Tree Species Composition in Subcompartments
19Copyright (C) Mitsubishi Research Institute, Inc.
Predicted Tree Species Distribution Map
Legend
registry
Predicted
Young Picea glehnii
Invasion of broadleaf
to larch plantation
mixed = below 70% dominancy
20Copyright (C) Mitsubishi Research Institute, Inc.
Predicted Maps
Canopy Cover Timber Volume Tree Height
21Copyright (C) Mitsubishi Research Institute, Inc.
Validation and Evaluation – Canopy Cover Map
Line-thinnednon-thinned
Line-thinnednon-thinned
Confirm prediction reflects forest conditions rightly by field surveyConfirm prediction reflects forest conditions rightly by field survey
Canopy Density
22Copyright (C) Mitsubishi Research Institute, Inc.
Larix trees with relatively higherstand age were observed
Young Picea glehnii& Broadleaf forest of low height
Validation and Evaluation – Tree Height Map
23Copyright (C) Mitsubishi Research Institute, Inc.
Conclusions
Present forest monitoring method from hyperspectral and SAR image
Integrate diverse data source with different property of targets To overcome high-dimension-small-sample-size problem resulting in over-fitting,
sparse regularization techs (LASSO & Sparse Discriminant Analysis) are adopted
3 advantages of sparse regularization
Generalization, Interpretability, Optimal Band Selection
Experimental simulations of satellite-based operation prove effectiveness
Advantage in prediction accuracy to several supervised methods Advantage of hyperspectral data to multispectral Prediction results reflect existing forest conditions rightly
24Copyright (C) Mitsubishi Research Institute, Inc.
Many thanks for your kind attention.
Questions ?
25Copyright (C) Mitsubishi Research Institute, Inc.
Supplementary Slides
26Copyright (C) Mitsubishi Research Institute, Inc.