ABSTRACT

Quantifying mangrove soil organic carbon (SOC) is key to better understanding the global carbon cycle, a critical phenomenon in reducing greenhouse gas emissions. However, it is challenging to have a large sample size in soil carbon measurements and analysis due to the high costs associated with them. In the current research, we propose a novel hybridized artificial intelligence model based on the categorical boosting regression (CBR) and the particle swarm optimization (PSO) algorithm for feature selection, namely, the CBR-PSO model for estimating mangrove SOC. We integrated multisensor optical (Sentinel-2) and synthetic aperture radar (Sentinel-1 and ALOS-2 PALSAR-2) remote sensing data to construct and verify the proposed model, drawing upon a survey in 85 soil cores at 100 cm depth in the Red River Delta, Vietnam. The CBR-PSO model estimated the mangrove SOC ranging from 44.74 to 91.92 Mg ha⁻¹ (average = 68.76 Mg ha⁻¹) with satisfactory accuracy (coefficient of determination (R²) = 0.809 and root-mean-square error (RMSE) = 9.30 Mg ha⁻¹). We also compared the proposed model’s capability with four machine learning techniques, i.e. support vector regression (SVR), random forest regression (RFR), extreme gradient boosting regression (XGBR), and XGBR-PSO models. We show that multimodal and multisensor earth observation dataset combined with the CBR-PSO model can significantly improve the estimates of mangrove SOC. Our findings contribute novel and advanced machine learning approaches for robustness of SOC estimation using open-source software. Our novel framework, which is automated, fast, and reliable, developed in this study can be easily applicable to other mangrove ecosystems across the world, thus providing insights for a voluntary blue carbon offset marketplace for sustainable mangrove conservation.

摘要

量化红树林土壤有机碳 (SOC) 是更好地了解全球碳循环的关键，这是减少温室气体排放的关键现象。然而，由于与之相关的高成本，在土壤碳测量和分析中拥有大样本量具有挑战性。在目前的研究中，我们提出了一种基于分类提升回归 (CBR) 和粒子群优化 (PSO) 算法进行特征选择的新型混合人工智能模型，即用于估计红树林 SOC 的 CBR-PSO 模型。我们整合了多传感器光学 (Sentinel-2) 和合成孔径雷达 (Sentinel-1 和 ALOS-2 PALSAR-2) 遥感数据来构建和验证所提出的模型，利用对 100 厘米深度的 85 个土壤核心的调查越南红河三角洲。^-1（平均值 = 68.76 Mg ha ^-1），准确度令人满意（决定系数 ( R ² ) = 0.809 且均方根误差 (RMSE) = 9.30 Mg ha ^-1））。我们还将所提出模型的能力与四种机器学习技术进行了比较，即支持向量回归 (SVR)、随机森林回归 (RFR)、极端梯度提升回归 (XGBR) 和 XGBR-PSO 模型。我们表明，多模式和多传感器地球观测数据集与 CBR-PSO 模型相结合可以显着提高红树林 SOC 的估计。我们的发现为使用开源软件进行 SOC 估计的稳健性提供了新颖和先进的机器学习方法。我们在本研究中开发的自动化、快速且可靠的新型框架可以轻松应用于世界各地的其他红树林生态系统，从而为可持续红树林保护的自愿蓝碳抵消市场提供见解。