A Novel Feature Evaluation Method in Mapping Forest AGB by Fusing Multiple Evaluation Metrics Using PolSAR Data

For mapping forest above-ground biomass (AGB), a novel feature evaluation method, named integrating multiple metrics with collinearity considered (IMC), is proposed by fusing different types of metrics with considering information redundancy among features. The optimal feature set was obtained using IMC metric from various types of features extracted from L-band ALOS-2 PALSAR data in planted Chinese fir forest. To evaluate the effectiveness of the IMC metric, three common feature evaluation metrics, Pearson correlation coefficient, importance derived from random forest (RF), and the sensitivity index (SI), were also used to obtain the optimal feature set, respectively. Moreover, the combination effect among features was also analyzed to filter the global best feature set in mapping for forest AGB. The results showed that the accuracy of forest AGB estimation was obviously improved using the proposed IMC metric [0.45–0.67 for $R^{2}$ and from 24.51% to 32.16% for relative root mean square error (rRMSE)]. Furthermore, it is also implied that the combination effect among features has the great potential to improve the capability of forest AGB estimation.

I. Introduction

With the rapid advancement of remote sensing technology, Polarimetric Synthetic Aperture Radar (PolSAR) data have a great potential to map forest above-ground biomass (AGB) by reducing the limitations of cloud cover and solar illumination [1], [2]. Commonly, many alternative features extracted from PolSAR are widely used to construct the estimation models with ground measured samples. However, how to evaluate these features and obtain the optimal feature set is a crucial factor that directly impacts the accuracy of mapping forest AGB [3], [4].