Landsat-based spatiotemporal estimation of subtropical forest aboveground carbon storage using machine learning...

Area of Science:

• Forestry Science
• Remote Sensing Applications
• Machine Learning in Ecology

Background:

• Forest aboveground carbon storage (AGC) is vital for ecosystem assessment.
• Estimating regional forest AGC faces technical challenges.
• Remote sensing and machine learning offer solutions for accurate AGC monitoring.

Purpose of the Study:

• To model and analyze the spatiotemporal dynamics of forest AGC in Lishui City over 30 years (1989-2019).
• To compare the performance of Backpropagation Neural Network (BPNN), Random Forest (RF), and Categorical Boosting (CatBoost) for AGC estimation.
• To identify key variables for accurate forest AGC modeling using remote sensing data.

Main Methods:

• Utilized Landsat remote sensing images for Lishui City.
• Employed BPNN, RF, and CatBoost machine learning algorithms to model forest AGC.
• Evaluated model performance using R-squared and Root Mean Squared Error (RMSE).
• Analyzed texture information from 9x9 and 11x11 windows as predictor variables.

Main Results:

• Texture information is a significant variable for forest AGC estimation.
• All tested machine learning models accurately estimated forest AGC.
• CatBoost demonstrated superior performance with R² of 0.95 (training) and 0.83 (testing).
• Forest AGC is concentrated in central and southwestern Lishui City, showing a significant increase from 1989 to 2019.

Conclusions:

• The CatBoost algorithm is highly effective for estimating AGC in subtropical forests.
• This study provides optimized machine learning models and hyperparameters for subtropical forest AGC assessment.
• The findings offer crucial reference data for enhancing carbon sequestration in Lishui City's subtropical forests.