AUTHOR=Liu Wenjun , Xu Cong , Zhang Zhiming , De Boeck Hans , Wang Yanfen , Zhang Liankai , Xu Xiongwei , Zhang Chen , Chen Guiren , Xu Can TITLE=Machine learning-based grassland aboveground biomass estimation and its response to climate variation in Southwest China JOURNAL=Frontiers in Ecology and Evolution VOLUME=11 YEAR=2023 URL=https://www.frontiersin.org/journals/ecology-and-evolution/articles/10.3389/fevo.2023.1146850 DOI=10.3389/fevo.2023.1146850 ISSN=2296-701X ABSTRACT=

The demand for accurate estimation of aboveground biomass (AGB) at high spatial resolution is increasing in grassland-related research and management, especially for those regions with complex topography and fragmented landscapes, where grass and shrub are interspersed. In this study, based on 519 field AGB observations, integrating Synthetic Aperture Radar (SAR; Sentinel-1) and high-resolution (Sentinel-2) remote sensing images, environmental and topographical data, we estimated the AGB of mountain grassland in Southwest China (Yunnan Province and Guizhou Province) by using remote sensing algorithms ranging from traditional regression to cutting edge machine learning (ML) and deep learning (DL) models. Four models (i.e., multiple stepwise regression (MSR), random forest (RF), support vector machine (SVM) and convolutional neural network (CNN)) were developed and compared for AGB simulation purposes. The results indicated that the RF model performed the best among the four models (testing dataset: decision co-efficient (R2) was 0.80 for shrubland and 0.75 for grassland, respectively). Among all input variables in the RF model, the vegetation indices played the most important role in grassland AGB estimation, with 6 vegetation indices (EVI, EVI2, NDVI, NIRv, MSR and DVI) in the top 10 of input variables. For shrubland, however, topographical factors (elevation, 12.7% IncMSE (increase in mean squared error)) and SAR data (VH band, 11.3% IncMSE) were the variables which contributed the most in the AGB estimation model. By comparing the input variables to the RF model, we found that integrating SAR data has the potential to improve grassland AGB estimation, especially for shrubland (26.7% improvement in the estimation of shrubland AGB). Regional grassland AGB estimation showed a lower mean AGB in Yunnan Province (443.6 g/m2) than that in Guizhou Province (687.6 g/m2) in 2021. Moreover, the correlation between five consecutive years (2018–2022) of AGB data and climatic factors calculated by partial correlation analysis showed that regional AGB was positively related with mean annual precipitation in more than 70% of the grassland and 60% of the shrubland area, respectively. Also, we found a positive relationship with mean annual temperature in 62.8% of the grassland and 55.6% of the shrubland area, respectively. This study demonstrated that integrating SAR into grassland AGB estimation led to a remote sensing estimation model that greatly improved the accuracy of modeled mountain grassland AGB in southwest China, where the grassland consists of a complex mix of grass and shrubs.