No abstract

The largest area of tropical rainforests in China is on Hainan Island, and it is an important part of the world’s tropical rainforests. The structure of the tropical rainforests in Hainan is complex, the biomass density is high, and conducting ground surveys is difficult, costly, and time-consuming. Remote sensing is a good monitoring method for biomass estimation. However, the saturation phenomenon of such data from different satellite sensors results in low forest biomass estimation accuracy in tropical rainforests with high biomass density. Based on environmental information, the biomass of permanent sample plots, and forest age, this study established a tropical rainforest database for Hainan. Forest age and 14 types of environmental information, combined with an enhanced vegetation index (EVI), were introduced to establish a tropical rainforest biomass estimation model for remote sensing that can overcome the saturation phenomenon present when using remote sensing data. The fitting determination coefficient of the model was 0.694. The remote sensing estimate of relative bias was 2.29%, and the relative root mean square error was 35.41%. The tropical rainforest biomass in Hainan Island is mainly distributed in the central mountainous and southern areas. The tropical rainforests in the northern and coastal areas have been severely damaged by tourism and real estate development. Particularly in low-altitude areas, large areas of tropical rainforest have been replaced by economic forests. Furthermore, the tropical rainforest areas in some cities and counties have decreased, affecting the increase in tropical rainforest biomass. On Hainan Island, there were few tropical rainforests in areas with high rainfall. Therefore, afforestation in these areas could maximize the ecological benefits of tropical rainforests. To further strengthen the protection, there is an urgent need to establish a feasible, reliable, and effective tropical rainforest loss assessment system using quantitative scientific methodologies.

The accurate estimation of carbon stocks in natural and plantation forests is a prerequisite for the realization of carbon peaking and neutrality. In this study, the potential of optical Sentinel-2A data and a digital elevation model (DEM) to estimate the spatial variation of carbon stocks was investigated in a mountainous warm temperate region in central Japan. Four types of image preprocessing techniques and datasets were used: spectral reflectance, DEM-based topography indices, vegetation indices, and spectral band-based textures. A random forest model combined with 103 field plots as well as remote sensing image parameters was applied to predict and map the 2160 ha University of Tokyo Chiba Forest. Structural equation modeling was used to evaluate the factors driving the spatial distribution of forest carbon stocks. Our study shows that the Sentinel-2A data in combination with topography indices, vegetation indices, and shortwave-infrared (SWIR)-band-based textures resulted in the highest estimation accuracy. The spatial distribution of carbon stocks was successfully mapped, and stand-age- and forest-type-level variations were identified. The SWIR-2-band and topography indices were the most important variables for modeling, while the forest stand age and curvature were the most important determinants of the spatial distribution of carbon stock density. These findings will contribute to more accurate mapping of carbon stocks and improved quantification in different forest types and stand ages.

The accurate classification of forest types is critical for sustainable forest management. In this study, a novel multiscale global graph convolutional neural network (MSG-GCN) was compared with random forest (RF), U-Net, and U-Net++ models in terms of the classification of natural mixed forest (NMX), natural broadleaved forest (NBL), and conifer plantation (CP) using very high-resolution aerial photographs from the University of Tokyo Chiba Forest in central Japan. Our MSG-GCN architecture is novel in the following respects: The convolutional kernel scale of the encoder is unlike those of other models; local attention replaces the conventional U-Net++ skip connection; a multiscale graph convolutional neural block is embedded into the end layer of the encoder module; and various decoding layers are spliced to preserve high- and low-level feature information and to improve the decision capacity for boundary cells. The MSG-GCN achieved higher classification accuracy than other state-of-the-art (SOTA) methods. The classification accuracy in terms of NMX was lower compared with NBL and CP. The RF method produced severe salt-and-pepper noise. The U-Net and U-Net++ methods frequently produced error patches and the edges between different forest types were rough and blurred. In contrast, the MSG-GCN method had fewer misclassification patches and showed clear edges between different forest types. Most areas misclassified by MSG-GCN were on edges, while misclassification patches were randomly distributed in internal areas for U-Net and U-Net++. We made full use of artificial intelligence and very high-resolution remote sensing data to create accurate maps to aid forest management and facilitate efficient and accurate forest resource inventory taking in Japan.