Abstract: Remote sensing data can be used for quickly obtaining real-time forest attribute information. Forest aboveground biomass (AGB) estimated by remote sensing data is usually spatially continuous and highly accurate. Due to the complex terrain and special climate, it is difficult to obtain forest attribute information of alpine regions compared with low-latitude or low-altitude forest ecosystems. Therefore, optical remote sensing has become an important means to estimate forest attributes. Based on MODIS satellite images and plot survey data, we estimated forest AGB in the Tibetan Plateau by a random forest model (RF), and further explored the AGB of major tree species by K-nearest neighbor algorithm (KNN). Moreover, we evaluated the prediction accuracy of AGB at different scales. We also analyzed the importance of predictive variables. The results showed that: (1) The AGB estimation model we developed had good performance at both pixel (R²=0.82, RMSE=64.93 t·hm^-2) and landscape scales (t=0.15, P=0.88). (2) Spatially, forest AGB gradually decreased from southeast to northwest with an average AGB of 181.28±104.54 t·hm^-2. The AGB of forests below 1000 m asl was the highest, with a value of 237.66 t·hm^-2. The AGB of Abies fabri, Picea asperata, and Pinus yunnanensis were relatively high, with values of 214.86, 216.14, and 172.24 t·hm^-2, respectively. (3) Geographical location and climate were the more important variables in estimating AGB. Our results contribute to improving the understanding of forest resources on the Tibetan Plateau and the prediction accuracy of carbon dynamics in China.

Key words: machine learning, aboveground biomass, forest ecosystem, remote sensing information.