Machine Learning Improves Forest Aboveground Biomass Estimations with Multimodal Remote Sensing Data

A recent study published in the Journal of Engineering has demonstrated the potential of machine learning algorithms in improving forest aboveground biomass estimations using multimodal remote sensing data. The research, conducted by scientists at the University of Connecticut, highlights the importance of variable selection methods, hyperparameter tuning of machine learning algorithms, and the integration of multimodal remote sensing data in improving large-area aboveground biomass prediction models.

Key Takeaways:

• The research found that the integration of multimodal remote sensing data, including airborne LiDAR, National Agriculture Imagery Program (NAIP) aerial imagery, and Sentinel-2 satellite images, with field-based forest inventory analysis (FIA) data, improved the efficiency of large-scale aboveground biomass modeling and carbon monitoring initiatives.
• The hyperparameter-tuned random forest (RF) model produced a root mean square error (RMSE) of 27.19 Mgha-1 and an R2 of 0.41, outperforming the evaluation metrics of support vector machine (SVM) and multiple linear regression (MLR) models.
• The study found that 68% of the variables used to build the best RF model were derived from the LiDAR height data.
• The linear ensemble model, developed using the predictions of all three models, yielded an R2 of 0.79, indicating a significant improvement in aboveground biomass predictions.

Statistics:

• The RMSE of the hyperparameter-tuned RF model was 27.19 Mgha-1.
• The R2 of the hyperparameter-tuned RF model was 0.41.
• The RMSE of the linear SVM model was 32.17 Mgha-1.
• The R2 of the linear SVM model was 0.10.
• The RMSE of the MLR model with eight explanatory variables was 22.59 Mgha-1.
• The R2 of the MLR model with eight explanatory variables was 0.22.
• The R2 of the linear ensemble model was 0.79.

Sources:

• Journal of Engineering. October 20, 2025; p 818.
• Forests, 2025;16(9):1430.