Modelling and mapping growing stock volume in small-scale forest plantations using Sentinel-2 remotely sensed data

Abstract

Preserving and managing plantation forests requires detailed information on growing stock volume (GSV) at different spatial scales. In this study, we demonstrated the potential of integrating field inventory and remotely sensed data to enhance large-scale predictions of GSV in small-scale forest plantations in the Southern Highlands of Tanzania. A total of 32 small-scale forest plantation farms/stands were selected within the area of interest (AOI). In each stand, one to three circular plots of 10 m radius were established, making a total of 63 field plots within the AOI. Within each field plot, the diameter at breast height of all trees were measured and for three sample trees the total tree height was measured. The GSV was then computed for each plot. Sentinel-2 remotely sensed data covering the AOI were downloaded and processed using the Google Earth Engine platform. For each field plot, remotely sensed predictor variables were extracted in the buffer corresponding with the field plot area. Three sets of predictor variables, comprising band values, vegetation indices and texture variables, were extracted. Statistical models incorporating GSV and the remotely sensed predictor variables were developed using ordinary least square (OLS) and Random Forest regression. The results showed that the OLS linear model fitted using texture-based variables was the best among the parametric models with a relative root mean square error (RMSE_r) of 41.28%. Likewise, the Random Forest model fitted using the combination of all predictor variables was best among the nonparametric models with a RMSE_r of 37.8%. Generally, the Random Forest model was the best model, which was used to produce cell-wise predictions across the entire AOI. The mean GSV for the entire AOI predicted using a Random Forest was 152 m³ ha⁻¹ and the standard deviation was 53.9 m³ ha⁻¹. The estimated mean GSV was similar to the field-measured mean GSV of 160.7 m³ ha⁻¹. The standard deviation was relatively lower compared to the field-based value, which was 76.47 m³ ha⁻¹. Thus, our study demonstrates that Sentinel-2 remotely sensed data can be used to develop a cost-effective method for GSV estimation in the small-scale plantations of Tanzania.