Estimation of above-ground carbon stock of Dendrocalamus giganteus using ANUSPLIN-interpolated GEDI and ICESat-2/ATLAS parameters

TL;DR

This study improves the estimation of carbon storage in giant bamboo forests using advanced interpolation and machine learning techniques with satellite data.

Contribution

The study introduces ANUSPLIN interpolation and XGBoost for more accurate regional carbon stock estimation using GEDI and ICESat-2/ATLAS data.

Findings

ANUSPLIN interpolation with specific spline degrees achieved highest accuracy for GEDI and ICESat-2/ATLAS parameters.

XGBoost outperformed LightGBM and RFR in estimating above-ground carbon stock with higher precision and lower error.

ANUSPLIN interpolation outperformed co-kriging in spatial interpolation performance.

Abstract

GEDI and ICESat-2/ATLAS have significant limitations in estimating forest structure parameters. This study aims to enhance the estimation accuracy of above-ground carbon storage (AGC) of Dendrocalamus giganteus by applying the ANUSPLIN interpolation technique for spatial expansion. The results indicate that: (1) When the spline degree of ANUSPLIN interpolation was set to 4 for GEDI parameters (cover, pai, sensitivity) and ICESat-2/ATLAS parameters (toc_roughness, h_median_canopy_abs, h_canopy_abs, h_max_canopy_abs), the model’s accuracy was highest. The spline degree 2 for the digital_elevation_model parameter (GEDI) and the asr parameter (ICESat-2/ATLAS) yielded optimal results. (2) The interpolation accuracy and performance of ANUSPLIN outperformed that of co-kriging (CK). (3) The Extreme Gradient Boosting (XGBoost) model (Coefficient of Determination, R2 = 0.93; Root Mean Square Error, RMSE = 5.89 Mg/ha; Overall Estimation Accuracy, P = 85.84%; relative RMSE, rRMSE = 14.16%) outperformed the Light Gradient Boosting Machine (LightGBM) (R2 = 0.52, RMSE = 14.61 Mg/ha, P = 64.84%, rRMSE = 35.16%) and Random Forest Regression (RFR) (R2 = 0.90, RMSE = 8.23 Mg/ha, P = 79.79%, rRMSE = 20.21%), achieving relative improvements of 78.85%, 59.67%, 32.36%, and 59.71% over LightGBM, and 3.33%, 28.41%, 7.58%, and 29.94% over RFR, respectively. This study demonstrates the feasibility of using ANUSPLIN interpolation for satellite LiDAR data from GEDI and ICESat-2/ATLAS. The approach offers a new perspective on spatial interpolation of satellite LiDAR data at a regional scale, providing a valuable reference for cost-effective, high-precision estimation of forest structural parameters.

Figures