Forest height estimation combining single-polarization tomographic and PolSAR data

• DOI: 10.1016/j.jag.2023.103532
• Publication Date: 2023-10-27T23:34:05Z
• Authors: Yihao Zhang; Xing Peng; Qinghua Xie; Yanan Du; Bing Zhang; Xiaomin Luo; Shaobo Zhao; Zhentao Hu; Xinwu Li
• Keywords: Environmental sciences; Forest height; Physical geography; Machine learning; 0211 other engineering and technologies; Polarimetric SAR; GE1-350; 02 engineering and technology; Tomographic SAR; LightGBM; Random forest; GB3-5030

Forest height is of great significance for forest resource management and carbon sink estimation. Tomographic synthetic aperture radar (TomoSAR) technology provides an effective means the accurate inversion this parameter. Several multi-polarization (SAR) images are generally required to obtain height. However, it common that only a small number single-polarization can be acquired, due complexity systems limitations observation cycles, there may one fully polarimetric image available. This impossible use TomoSAR estimate over wide area. Based on this, in study, we combined SAR (PolSAR) variables (SP-TomoSAR) features first time. The fusion was achieved through six machine learning methods: light gradient-boosting (lightGBM), random (RF), extreme gradient boosting (XGBoost), gradient-boosted decision tree (GBDT), k-nearest neighbor (KNN), support vector regression (SVR). To investigate advantages proposed method, amount SP-TomoSAR data with non-uniformly distributed baselines PolSAR were acquired tropical rainforest French Guiana. We then used H/A/alpha Freeman-Durden decomposition methods polarization applied Capon algorithm tomographic features. Four sets comparative experiments carried out, results confirmed combination achieve estimation height, result HV better than HH Moreover, after adding features, accuracy clearly improved, compared using which suggests provide important supplementary information SP-TomoSAR. In addition, among algorithms, RF has highest root mean square error (RMSE) 5.14 m R 0.83, while lightGBM significantly ahead others terms computational efficiency.