To achieve sustainable forest management in tropical forests, it is necessary to monitor aboveground carbon stock (AGC) and to apply that information for improving management practices. Conventional methods to estimate AGC at large spatial scale (>1000km2) require tree inventory data as ground truthing at each year and region in which satellite images are taken. Because the costs for ground truthing is a major burden for monitoring AGC, we have developed a method to reduce the cost by (i) integrating tree inventory data from multiple regions across Borneo (483 plots; Komatsu unpublished) and (ii) using a machine learning model to estimate AGC. The large-scale model predicts AGC in a region using the information of surrounding regions, leading to the lower cost for ground truthing. However, the model accuracy depends on the availability of training data of the target region: prediction accuracy was 69.3 Mg ha–1 RMSE if more than 20 plots are available in a region whereas 81.4 Mg ha–1 RMSE if no data are available.

We examined whether the prediction accuracy of the large-scale model is improved by using drone-derived ground information instead of conducting labor-intensive field survey. We took RGB images using a drone (10 ha for each; Mavic-2 pro), which can take ground information with low costs, at 100-m height above 0.12 ha permanent plots in the Deramakot/Tangkulap, Malaysia (25 plots), and estimated canopy height and the ratio of dipterocarp canopy cover of the plots. AGC was estimated with linear regression with the metrics (R2 = 0.57), and the regression model was extrapolated to random points within the drone images (a total of 705 points), which were used as training data to improve model accuracy.

Accuracy to predict plot AGC in the Deramakot/Tangkulap was calculated with three different approaches simulating situations of different data availabilities. Overall model accuracy of the large-scale model was 63.2 Mg ha–1 RMSE (random cross validation, the accuracy when inventory data are available). When plot data in Deramakot/Tangkulap were excluded from training data, model accuracy declined (80.9 Mg ha–1 RMSE, accuracy when plot data are unavailable). Adding data derived from drone imagery as training data significantly improved model accuracy by 11.9 Mg ha–1 RMSE (the accuracy when plot data are unavailable but drone-derived data are available). This study demonstrated that the combination of drone and satellite imagery has a potential to minimize the cost for ground truthing.