Wavelet transform applied to EO-1 hyperspectral data for forest LAI and crown closure mapping

Abstract

A comparison of the performance of three feature extraction methods was made for mapping forest crown closure (CC) and leaf area index (LAI) with EO-1 Hyperion data. The methods are band selection (SB), principal component analysis (PCA) and wavelet transform (WT). Hyperion data were acquired on October 9, 2001. A total of 38 field measurements of CC and LAI were collected on August 10–11, 2001, at Blodgett Forest Research Station, University of California at Berkeley, USA. The analysis method consists of (1) conducting atmospheric correction with High Accuracy Atmospheric Correction for Hyperspectral Data (HATCH) to retrieve surface reflectance, (2) extracting features with the three methods: SB, PCA and WT, (3) establishing multivariate regression prediction models, (4) predicting and mapping pixel-based CC and LAI values, and (5) validating the CC and LAI mapped results with photo-interpreted CC and LAI values. The experimental results indicate that the energy features extracted by the WT method are the most effective for mapping forest CC and LAI (mapped accuracy (MA) for CC=84.90%, LAI MA=75.39%), followed by the PCA method (CC MA=77.42%, LAI MA=52.36%). The SB method performed the worst (CC MA=57.77%, LAI MA=50.87%).