Combining 3D Radiative Transfer Model and Convolutional Neural Network to Accurately Estimate Forest Canopy Cover From Very High-Resolution Satellite Images

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 14; pp. 10953 - 10963
• Main Authors: Jin, Decai, Qi, Jianbo, Huang, Huaguo, Li, Linyuan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial intelligence; Artificial neural networks; Biological system modeling; Canopies; Canopy; convolutional neural network (CNN); Convolutional neural networks; Data models; Datasets; FCC; forest canopy cover (FCC); Forestry; High resolution; Hydrology; Image resolution; Learning; Neural networks; Plant cover; Radiative transfer; Remote sensing; Resolution; Satellite imagery; Simulation; Solid modeling; Spaceborne remote sensing; Three dimensional models; Three-dimensional (3-D) radiative transfer model; Three-dimensional displays; Training; Transfer learning
• DOI: 10.1109/JSTARS.2021.3122509
• ISSN: 1939-1404

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 article, we proposed a hybrid model, which combines a 3-D RTM 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, R 2 = 0.83), compared with RF (RMSE = 0.26, R 2 = 0.61), Data-CNN (RMSE = 0.142, R 2 = 0.81), and RTM-CNN (RMSE = 0.144, R 2 = 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.