Combining 3D Radiative Transfer Model and Convolutional Neural Network to Accurately Estimate Forest Canopy Cover From Very High-Resolution Satellite Images
Decai Jin, Jianbo Qi, Huaguo Huang, Linyuan Li
Abstract
Forest canopy cover (FCC) plays an important role in many ecological, hydrological and forestry applications. For large-scale applications, FCC is usually estimated from remotely sensed data by inverting radiative transfer models (RTMs) or using data-driven regressions. In this study, we proposed a hybrid model, which combines 3D RTMs and transfer learning-based convolutional neural network (T-CNN), to estimate FCC from very high-resolution satellite images (e.g., Chinese GaoFen-2, 1 m resolution with 4 bands). Unlike common hybrid models that are purely trained with simulation data, T-CNN combines simulation data-based pre-training and actual data-based transfer learning, which is a widely used technique in artificial intelligence for fine-tuning models. The performance of T-CNN was compared with a random forest (RF) model and two general CNN models, including CNN trained with actual dataset only (Data-CNN) and CNN trained with RTM simulation data only (RTM-CNN). Results on the independent validation dataset (not used in training stage) showed that T-CNN had higher accuracy (RMSE = 0.121, R2 = 0.83), compared with RF (RMSE = 0.26, R2 = 0.61), Data-CNN (RMSE = 0.142, R2 = 0.81), and RTM-CNN (RMSE = 0.144, R2 = 0.73), which indicates that T-CNN has a strong transferability. Tests on different training sizes showed that T-CNN (0.084<RMSE<0.108) provided constantly better performances than RF (0.116<RMSE<0.122) and Data-CNN (0.103<RMSE<0.128), which demonstrates the potential of T-CNN as an alternative to RTM-based inversion and data-driven regressions to estimate FCC, especially when training data is imbalanced and inadequate.