Optimal Wavelength Selection on Hyperspectral Data With Fused Lasso for Biomass Estimation of Tropical Rain Forest

Abstract

Commission VIII, WG VIII/7KEY WORDS: Biomass, Hyperspectral, Forest Management, Fused Lasso, GLCM ABSTRACT:Above-ground biomass prediction of tropical rain forest using remote sensing data is of paramount importance to continuous largearea forest monitoring. Hyperspectral data can provide rich spectral information for the biomass prediction; however, the prediction accuracy is affected by a small-sample-size problem, which widely exists as overfitting in using high dimensional data where the number of training samples is smaller than the dimensionality of the samples due to limitation of require time, cost, and human resources for field surveys. A common approach to addressing this problem is reducing the dimensionality of dataset. Also, acquired hyperspectral data usually have low signal-to-noise ratio due to a narrow bandwidth and local or global shifts of peaks due to instrumental instability or small differences in considering practical measurement conditions. In this work, we propose a methodology based on fused lasso regression that select optimal bands for the biomass prediction model with encouraging sparsity and grouping, which solves the small-sample-size problem by the dimensionality reduction from the sparsity and the noise and peak shift problem by the grouping. The prediction model provided higher accuracy with root-mean-square error (RMSE) of 66.16 t/ha in the crossvalidation than other methods; multiple linear analysis, partial least squares regression, and lasso regression. Furthermore, fusion of spectral and spatial information derived from texture index increased the prediction accuracy with RMSE of 62.62 t/ha. This analysis proves efficiency of fused lasso and image texture in biomass estimation of tropical forests.

Estimating Biomass of Native Grass Grown under Complex Management Treatments Using WorldView-3 Spectral Derivatives

“…Our results are consistent with those of Zhang et al [62], who noted that the models derived from a combination of spectrum and texture models of the Chinese high-resolution remote sensing satellite Gaofen-1, increased the estimation accuracies of Populus euphratica forest when compared with the performance of reflectance or texture models. In another similar study, Takayama and Iwasaki [63] showed that the combination of the spatial and spectral information from spectral responses and texture models optimally improved the estimation accuracies of tropical vegetation biomass from a RMSE of 66.16 t/ha to a RMSE of 62.62 t/ha in Hampangen, Central Kalimantan, Indonesia, based on WV-3 satellite data. Kelsey and Neff [53] also demonstrated that texture models improved the estimation of vegetation biomass at the San Juan National Forest in southwest Colorado, USA, from a RMSE of 56.4 to a RMSE of 45.6, based on Landsat data.…”

Remotely sensed normalized difference red-edge index for rangeland biomass estimation

“…Our results are consistent with those of Zhang et al (2015), who noted that the models derived from a combination of spectrum and texture models of the Chinese high-resolution remote sensing satellite Gaofen-1, increased the estimation accuracies of Populus euphratica forest when compared with the performance of reflectance or texture models. In another similar study, Takayama and Iwasaki (2016) showed that the combination of the spatial and spectral information from spectral responses and texture models optimally improved the estimation accuracies of tropical vegetation biomass from a RMSE of 66.16 t/ha to a RMSE of 62.62 t/ha in Hampangen, Central Kalimantan, Indonesia, based on WV-3 satellite data. Kelsey and Neff (2014) also demonstrated that texture models improved the estimation of vegetation biomass at the San Juan National Forest in southwest Colorado, USA, from a RMSE of 56.4 to a RMSE of 45.6 and based on Landsat data.…”

Rice Seed Purity Identification Technology Using Hyperspectral Image with LASSO Logistic Regression Model

“…According to the number of feature wavelengths and recognition accuracy, the optimisation range of is set to 0.0001–0.2 in this study. We used cross-validation to determine the relatively small range of [ 63 , 64 ]. Thereafter, the number of feature bands and the accuracies for different values were analysed, and the optimal value was selected.…”