Forest height estimation combining single-polarization tomographic and PolSAR data

Abstract

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

Keywords

• Forest height
• Tomographic SAR
• Polarimetric SAR
• Machine learning
• Random forest
• LightGBM