Hybrid Deep Learning–Geostatistical Mapping of Forest Aboveground Biomass in Lishui, China

TL;DR

This study improves forest biomass mapping in Lishui, China, by combining deep learning with geostatistical methods to better account for spatial patterns.

Contribution

A novel hybrid CNN-Transformer and geostatistical Kriging approach is introduced for more accurate AGB mapping.

Findings

The CNN-Transformer-CK model achieved an R2 of 0.72 and RMSE of 12.18 t/ha in predicting AGB.

Sentinel-2 Band 8 and Band 12 were identified as the most influential predictors for AGB.

The hybrid model outperformed traditional CNN-Transformer and Random Forest models in accuracy.

Abstract

Forest aboveground biomass (AGB) is a key indicator of forest productivity and carbon sequestration, yet many remote sensing AGB models overlook spatial autocorrelation in plot observations and model residuals. This study proposes a hybrid framework that combines a CNN-Transformer (Convolutional Neural Network-Transformer) model with geostatistical Kriging of residuals to improve regional AGB mapping in Lishui City, Zhejiang Province, China. Using 398 forest plots and multi-source predictors derived from Sentinel-2 imagery, ALOS-2 PALSAR-2 SAR data, and ALOS 12.5 m DEM, relevant variables were screened using Random Forest importance ranking. The most influential predictors included Sentinel-2 Band 8 and Band 12, EVI, PC1, mean77, HH/HV, ARVI, NDVI, RVI, and elevation. Ten-fold cross-validation showed that the CNN-Transformer-CK model had the highest accuracy in predicting forest AGB, with a validation R2 of 0.72 and RMSE of 12.18 t/ha, followed by the CNN-Transformer model (R2 = 0.69, RMSE = 12.22 t/ha) and RF (R2 = 0.59 and RMSE = 14.31 t/ha). The proposed approach supports wall-to-wall AGB mapping for forest management and conservation planning.

Figures