基于水体光学类型的全球悬浮物浓度动态遥感反演研究

Research on Global Suspended Matter Concentration Dynamics Remote Sensing Inversion Based on Water Optical Types

摘要: 内陆湖泊及近岸水体的光学性质极端复杂，悬浮颗粒物(SPM)浓度跨度可达数个数量级，导致传统的单一通用反演算法难以兼顾全动态范围内的监测精度。针对这一难题，本研究基于涵盖全球多源的实测数据集(N = 3339)，构建了包含22类典型水体的光学分类(OWT)体系，并系统评估了11种主流算法(涵盖经验、半分析及机器学习模型)在原始、全局率定(CAL)及分类独立率定(CLUS)模式下的性能差异。研究量化修正了“分类率定绝对优于全局率定”的传统认知，揭示了CLUS策略在样本稀缺的极端水体中存在严重的过拟合风险(高方差)。基于“偏差–方差权衡”理论，本研究提出了一种混合优选策略，阐明了最佳反演机制随水体浑浊度增加呈现出由“半分析/经验法”向“机器学习算法”演变的规律。在此基础上，构建了集成分类识别与无缝反演的动态切换反演系统。独立验证结果表明，该系统有效融合了各子模型的局部优势，相比表现最好的单一全局算法，其均方根误差(RMSE)降低约25%，平均偏差(Bias)近乎为0，有效解决了低浓度区的估算偏差与高浓度区的信号饱和问题，为全球内陆及近岸水体的高精度、无缝SPM监测提供了一套稳定的解决方案。

Abstract: The extreme optical complexity and wide dynamic range of Suspended Particulate Matter (SPM) in inland and coastal waters pose significant challenges for traditional “one-size-fits-all” retrieval algorithms. To address this, this study establishes an Optical Water Type (OWT) framework comprising 22 distinct classes based on a comprehensive global in-situ dataset (N = 3339). We systematically evaluated the performance of 11 mainstream algorithms (including empirical, semi-analytical, and machine learning models) under three calibration modes: original, global (CAL), and class-specific (CLUS). The results quantitatively revise the conventional wisdom that “class-specific calibration is inherently superior”, revealing that the CLUS strategy suffers from high variance (overfitting) in data-scarce extreme waters. Drawing on the “Bias-Variance Tradeoff” theory, we propose a hybrid optimization strategy and identify a clear paradigm shift in the optimal retrieval mechanism—transitioning from semi-analytical/empirical methods to machine learning algorithms as turbidity increases. Consequently, a dynamic switching retrieval system integrating classification and smooth blending was constructed. Independent validation demonstrates that this system effectively synthesizes the local advantages of different models. Compared to the best single global algorithm, the dynamic system reduces the Root Mean Square Error (RMSE) by approximately 25% and achieves a near-zero bias. It successfully mitigates estimation bias in low-concentration waters and signal saturation in high-concentration waters, providing a robust solution for high-precision, seamless SPM monitoring across global scales.