Differences in estimation of above-ground carbon stocks in forests by UAV hyperspectral underground sample plots

<bold>Objective</bold> The deployment of ground sample plots represents a fundamental aspect of remote sensing-based forest parameter inversion.The configuration and dimensions of these plots play a pivotal role in determining the precision and dependability of forest parameter retrievals derived from remote sensing data.Moreover，they substantially affect the effectiveness of forest resource assessments and monitoring initiatives.Consequently，choosing an optimal plot shape and size becomes crucial，as it ensures that remote sensing data effectively mirror actual forest conditions while simultaneously improving the efficiency of monitoring operations. <bold>Method</bold> Taking the north subtropical coniferous forests and broad-leaved forests as the research objects，sample plots （25.82 m × 25.82 m）are selected through typical sampling.Within these designated plots，all trees with a diameter of at least 5.0 cm were comprehensively surveyed.The biomass of individual trees was determined through the standing tree biomass models specific to the existing tree species.Subsequently，the total biomass of the sample plots was calculated，and the carbon stock of the plot was derived based on the established carbon measurement parameters.Simultaneously，unmanned aerial vehicle（UAV）based hyperspectral data of the ground sample plots were acquired to obtain high-resolution spectral information.With the aid of RTK position data，a detailed planimetric map of the trees in each plot was developed.To explore the effect of plot structure，the sample plots were further subdivided into smaller units of various shapes（e.g.，circular and square）and sizes.A series of preprocessing techniques were applied，including multiple scattering correction（MSC），standard normal variate transformation（SNV），polynomial smoothing （Savitzky-Golay smoothing），derivative computation，and discrete wavelet transform （DWT）to process the hyperspectral data.A comprehensive range of feature factors，including vegetation indices and texture features，was extracted.Subsequently，the optimal subset of feature variables for different forest types with varying land characteristics was identified using random forest importance ranking.Then random forest（RF）and extreme gradient boost （XGBoost）algorithms were employed to construct estimation models of aboveground carbon stocks on the coniferous forest，broad-leaved forest and all forests coniferous and broad-leaved forests）.This research aims to determine the most effective layout for sample plots and to investigate how plot characteristics influence aboveground carbon stock modeling in the forest ecosystems. <bold>Result</bold> The vegetation index factors processed using Savitzky-Golay smoothing-derivative，DWT，and SNV transformations were found to be more suitable for carbon stocks model.In comparison to standard square sample plots，circular sample plots demonstrated superior performance in constructing carbon stocks models for coniferous and broad-leaved forests.Notably，the models based on sample plots with a radius of 12.91 meters exhibited significantly higher accuracy than those based on plots of other sizes.The optimal test set evaluation for carbon stocks models yielded the following results for coniferous forests：<italic>R</italic>²_test=0.78，RMSE_test=10.15 t/hm²，rRMSE_test=18.6%；and for broad-leaved forests：<italic>R</italic>²_test=0.77，RMSE_test=5.77 t/hm²，rRMSE_test=10.95%，respectively. <bold>Conclusion</bold> The integration of circular sample plots with a radius of 12.91 meters and the XGBoost algorithm ensures the reliability and validity of UAV-based hyperspectral forest carbon stocks estimation for coniferous and broad-leaved forests，offering valuable insights for sustainable forest management and carbon sequestration research.