A new method for estimating forest stand carbon stock: Segmentation and modeling based on forest aboveground imagery

Abstract

Estimating forest stand carbon stock (SCS) is crucial for forest management and achieving carbon neutrality. However, traditional forest surveys are time-consuming, laborious and may posing the additional risk of harming the forest habitats. Remote sensing technology provides a faster and non-destructive method for SCS estimation. But these methods are hampered by issues like expensive equipment and complex data processing. To address these challenges, this study proposes a novel method to estimate SCS based on forest aboveground imagery (AGI) by integrating semantic segmentation techniques with SCS statistical modeling. Following the comparative analysis of three Fully Convolutional Network models (FCNs) for segmenting AGI, the optimized FCN8s semantic model was selected to segment AGI, capturing the pixel ratio of tree trunks within the imagery. Moreover, this study validated the enhanced precision and efficiency in segmentation models achieved through transfer learning. Subsequently, utilizing the foreground proportion of AGI derived by FCN8s as the independent variable and SCS as the dependent variable, ordinary least squares model and weight least squares model were developed. Through comparative analysis, the optimal SCS estimation model was determined. The experimental results demonstrated high segmentation accuracy with a validation set of 96.98 % for accuracy, 96.74 % for mean pixel accuracy, and 96.63 % for intersection over union. The optimal SCS model was the weight least squares model with R2 of 0.7989, RMSE of 34.41 Mg hm−2, TRE of 1.74 %, rRMSE of 0.137 %, MPE of 3.94 %, MPSE of 10.71 %. Overall, the proposed the method could estimate SCS in a low-cost and effective way.

Keywords

• Aboveground imagery
• Fully convolutional networks
• Weight least squares methods
• Stand Carbon Stock