The Use of a Modified GOPCE Method for Forest and Non-Forest Discrimination 1 The Use of a Modified GOPCE Method for Forest and Non-Forest Discrimination Junjun Yin, Zheng-Shu Zhou, Wooil M. Moon , Ruijin Jin, and Peter A. Caccetta Abstract—This study focuses on the development and evaluation of the GOPCE (Generalized Optimization of Polarimetric Contrast Enhancement) model to discriminate between forested and non-forested areas. The main objective is to investigate the performance of the GOPCE method for forest mapping, and to assess the potential of different polarimetric parameters for forest representation. We make two modifications to the original GOPCE method. Firstly, by comparing behaviors of different polarimetric parameters, the GOPCE model is modified. Then linear discriminant analysis is employed for further optimization of the target contrast. Forest/non-forest discrimination results are demonstrated on L-band fully polarimetric ALOS-1/PALSAR data acquired over a pilot study area in northeastern Tasmania, Australia, where the main forest type is eucalypt forests. Two other forest classification approaches (i.e., SVM and CVA) are also tested for comparison. The final results obtained from the modified GOPCE model with the generalized Fisher criterion can improve the forest/non-forest discrimination accuracy. Index Terms—Fisher criterion, forest and non-forest discrimination, generalized optimization of polarimetric contrast enhancement (GOPCE), radar polarimetry. I. INTRODUCTION orests are an important component of biological sinks for global carbon sequestration, and play an important role in mitigating greenhouse gas emissions from other sectors and land management practices. Forest cover can be monitored by many different instruments. One of the major objectives of monitoring forest is to estimate and locate the spatial extent of changes including reforestation, deforestation, afforestation and degradation. Radar remote sensing is an approach researched, developed and planned for some time [1][2] that is increasingly used for various Earth observation tasks as new sensors become more readily available. The space-borne synthetic aperture radar (SAR) sensors measure different properties of land cover as compared with optical imaging systems that have been extensively used for forest monitoring, and offer alternative means of measurement where geographical coverage or cloud cover may be a problem for optical instruments or complementary information where discrimination may be limited for either technology used in isolation. This study focuses on the application of fully polarimetric SAR to the discrimination problem between forested and non-forested areas. Compared to the ambient agricultural areas and grasslands, forest areas usually have larger backscattered energy and possess distinctly different polarimetric scattering properties and textural patterns. By exploiting these differences, several forest/non-forest classifiers were proposed in the literature, such as the classifiers based on the support vector machine (SVM) [3][4], and the directed discriminant technique-based canonical variate analysis (CVA) method [5]. Other approaches can also be found for the investigation of physical scattering mechanisms (PSMs) of forests. Widely accepted techniques are the ones using target decomposition-based methods [6][7][8]. In Antropov et al. (2011) [7], a new volume scattering model was proposed for forest representation, and then the forest/non-forest classification was carried out using a simple rule-based approach. However, no significant difference can be found between this method and the Freeman-Durden decomposition [6] when the proposed method was applied to L-band ALOS-1/PALSAR data. An important conclusion from Antropov et al. (2011) [7] was that, in boreal forest, the double-bounce scattering contributes little to the total backscattered signal. Scattering parameters from the Cloude-Pottier decomposition [9] have also been studied for forest classification [1][10]. It was reported by Park et al. [8] that polarization entropy and alpha angle are good indicators for describing the random volume scattering and averaged scattering process, and that both parameters are hardly affected by terrain slopes. In addition, anisotropy is also a useful parameter to evaluate the azimuthally symmetric scattering, which is the dominant scattering symmetry type for forested areas. In our previous study [10], polarization entropy, alpha angle, and anisotropy were investigated for forest/non-forest classification based on the Fisher discriminant analysis [11]. The classification accuracy was comparable to that of the CVA (linear discriminant) method [5]. However, in Yin et al. (2013) [10], the backscattered power contrast between the forest and non-forest classes was not optimized. Thus, in this study, the generalized optimization of the polarimetric contrast enhancement (GOPCE) method [12] is adopted to further optimize the discriminant distance. The GOPCE method has not previously been applied to forest/non-forest discrimination. We discuss further developments including two new modifications which can improve the classification accuracy. Verification is performed on the ALOS- 1/PALSAR data acquired of a pilot site in Tasmania, Australia. Results are compared with those of the SVM [1] and CVA [5] methods, showing the effectiveness of the F This work was supported in part by NNSFC under Grant 41171317 and in part by the Research Foundation of Tsinghua University under Grant 20111080968, China. Junjun Yin and Ruijin Jin are with the Department of Electronic Engineering, Tsinghua University, Beijing 100084, China (e-mail: yinjj07@gmail.com; jinruijin19900312@163.com). Zheng-Shu Zhou and Peter A. Caccetta are with the Division of Computational Informatics, CSIRO Floreat, WA 6014, Australia (e- mail: zheng-shu.zhou@csiro.au; peter.caccetta@csiro.au). Wooil M. Moon is with the Faculty of Environment, Earth and Resources, University of Manitoba, Winnipeg R3T 2N2, Canada (e- mail: Wooil.Moon@umanitoba.ca).