Abstract:

Efficient monitoring of the spatiotemporal distribution of ammonia nitrogen (NH₃-N) in river networks is crucial for managing regional water pollution and ensuring the health of the ecological environment. This study developed a machine learning-based remote sensing inversion model, tailored for large water bodies with significant variations in NH₃-N concentrations, using 204 in-situ NH₃-N data (ranging from 0.026 to 6.210 mg·L^-1) and eight high-quality Sentinel-2 MSI images collected in Guangzhou in 2019. The results indicated that the existing models for NH₃-N inversion exhibited low accuracy when applied to water bodies in Guangzhou, whereas models with multi-feature inputs demonstrated relatively better predictive capabilities. To address this, principal component analysis (PCA) was employed for feature dimensionality reduction (BC-FDR) after evaluating over 16000 Sentinel-2 band combinations. This approach was integrated with three machine learning models: Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Regression (SVR), to construct an optimized band-feature machine learning method for NH₃-N inversion. The BC-FDR_XGBoost model performed the best (r_c²=0.6872, σ_RMSEc=0.617 mg·L^-1, σ_MAEc=0.385 mg·L^-1, n=102; r_v²=0.5436, σ_RMSEv=0.438 mg·L^-1, σ_MAEv=0.362 mg·L^-1, n=44). Additionally, independent validation using 58 in-situ data points of NH₃-N (n=33) and trend analysis (n=25) further confirmed the high accuracy of the BC-FDR_XGBoost model (r²=0.5315, σ_RMSE=0.459 mg·L^-1, σ_MAE=0.287 mg·L^-1). The satellite-derived NH₃-N estimates showed strong consistency with the in situ data in both spatiotemporal distribution and trend analyses. The average NH₃-N concentration in the Guangzhou river network in 2019 met the Class III water quality standard, with significantly higher concentrations observed during the dry season than during the wet season (0.795 mg·L^-1 vs. 0.552 mg·L^-1). During the wet season, NH₃-N concentrations were generally lower in the northern and southern regions, and relatively higher in the central area. In contrast, during the dry season, only the Nansha district and certain sections of the mainstream exhibited relatively low NH₃-N concentrations. This study provides a reference for the development of city-scale, large-area NH₃-N remote-sensing inversion models, contributing to the assessment and management of regional water environments.

Key words: ammonia nitrogen model, river networks, feature optimization, machine learning, Sentinel-2 imagery, Guangzhou

摘要：

高效监测河网水系氨氮（NH₃-N）的时空分布对区域水体污染防控治理和生态环境健康发展具有重要意义。基于2019年在广州市收集的204个NH₃-N实测数据（0.026－6.210 mg·L^-1）和8景高质量Sentinel-2 MSI遥感影像，发展了适用于大范围水域、NH₃-N质量浓度差异显著的机器学习遥感反演模型。结果显示，已有的NH₃-N反演模型应用于广州市水体时精度受限，但多特征输入的模型预测能力相对较好。在检索16000多种Sentinel-2波段组合的基础上，利用主成分分析方法进行了特征降维（BC-FDR），并结合极端梯度提升（XGBoost）、随机森林（RF）、支持向量回归（SVR）3种机器学习模型，构建了波段特征优化的机器学习NH₃-N反演方法。其中BC-FDR_XGBoost模型表现最佳（r_c²=0.6872，σ_RMSEc=0.617 mg·L^-1，σ_MAEc=0.385 mg·L^-1，n=102；r_v²=0.5436，σ_RMSEv=0.438 mg·L^-1，σ_MAEv=0.362 mg·L^-1，n=44）。另外基于58个实测数据进行了独立验证（n=33）和趋势检验（n=25），结果进一步表明，BC-FDR_XGBoost模型的精度较高（r²=0.5315，σ_RMSE=0.459 mg·L^-1，σ_MAE=0.287 mg·L^-1），卫星遥感反演结果与实测数据在时空分布和变化趋势上具有良好的一致性。2019年，广州市河网水系NH₃-N质量浓度平均为Ⅲ类水质等级，枯水期（0.795 mg·L^-1）显著高于丰水期（0.552 mg·L^-1）。空间上，丰水期NH₃-N质量浓度整体呈南北部低、中部相对较高的特点；枯水期仅南沙区及部分干流NH₃-N相对较低。该研究为建立城市尺度大区域范围水体NH₃-N遥感反演模型提供了参考，有助于区域水环境的评价和治理。

关键词: 氨氮模型, 河网水系, 特征优化, 机器学习, Sentinel-2影像, 广州