Machine learning and geostatistical approaches for estimating aboveground biomass in Chinese subtropical forests

Abstract

Background: Aboveground biomass (AGB) is a fundamental indicator of forest ecosystem productivity and health and hence plays an essential role in evaluating forest carbon reserves and supporting the development of targeted forest management plans. Methods: Here, we proposed a random forest/co-kriging framework that integrates the strengths of machine learning and geostatistical approaches to improve the mapping accuracies of AGB in northern Guangdong province of China. We used Landsat time-series observations, Advanced Land Observing Satellite (ALOS) Phased Array L-band Synthetic Aperture Radar (PALSAR) data, and National Forest Inventory (NFI) plot measurements, to generate the forest AGB maps at three time points (1992, 2002, and 2010) showing the spatio-temporal dynamics of AGB in the subtropical forests in Guangdong, China. Results: The proposed model provided excellent performance for mapping AGB using spectral, textural, and topographical variables, and the radar backscatter coefficients. The root mean square error of the plot-level AGB validation was between 15.62 and 53.78 (t/ha), the mean absolute error ranged from 6.54 to 32.32 t/ha, and the relative improvement over the random forest algorithm was between 3.8% and 17.7%. The highest coefficient of determination (0.81) and the lowest mean absolute error (6.54 t/ha) were observed in the 1992 AGB map. The spectral saturation effect was minimized by adding the PALSAR data to the modeling variable set in 2010. By adding elevation as a covariable, the co-kriging outperformed the ordinary kriging method for the prediction of the AGB residuals, because co-kriging resulted in better interpolation results in the valleys and plains of the study area. Conclusions: Validation of the three AGB maps with an independent dataset indicated that the random forest/co-kriging performed best for AGB prediction, followed by random forest coupled with ordinary kriging (random forest/ordinary kriging), and the random forest model. The proposed random forest/co-kriging framework provides an accurate and reliable method for AGB mapping in subtropical forest regions with complex topography. The resulting AGB maps are suitable for the targeted development of forest management actions to promote carbon sequestration and sustainable forest management in the context of climate change.

Automated Machine Learning Driven Stacked Ensemble Modeling for Forest Aboveground Biomass Prediction Using Multitemporal Sentinel-2 Data

“…The type of modelling algorithm substantially affects the precision and accuracy of the predictions. As per literature, there are several studies that use different ML models for predicting forest AGB [7], [58], [59] and provide a comparative assessment of the models in order to identify the best modelling algorithm for AGB prediction. Apart from the range of ML models that can be used, each model has associated hyperparameters that require tuning.…”

“…Differently, stacking involves heterogeneous weak learners that are combined using a metamodel with a parallel learning strategy. Ensemble models have been used for AGB modelling and prediction by various studies in literature [8], [58], [59]. However, developing a stacked ensemble with manual trial and error approaches to model AGB is an inefficient and time-consuming task.…”

The Influence of Widow Size on Remote Sensing-Based Prediction of Forest Structural Variables

“…However, the traditional methods involving forest inventories that require labor-intensive eld surveys are quite time consuming and expensive (Kwak et al 2007;Joshi et al 2015;García et al 2018). Remote sensing data has been used as an auxiliary data source in forest inventory since early days of remote sensing (Kayitakire et al 2006;Su et al 2020). Stand structural variables such as crown canopy, species composition, stem density and basal area can be estimated on the forest stand level using remote sensing data.…”