Inversion of river discharge from remotely sensed river widths: a critical assessent at three-thousand global river gauges

Accurately estimating river discharge from satellite-derived river hydraulic variables (e.g., width, height, and slope) is the overarching goal of the remote sensing of discharge (RSQ) community. Numerous past studies have developed and intercompared different RSQ algorithms to determine favorable RSQ conditions, yet relatively few have focused on tailoring RSQ to a wide range of river forms globally. As the RSQ community is now ready to expand to the global scale given advances in computing power, sensors, and the upcoming launch of the Surface Water and Ocean Topography (SWOT) satellite mission, a much broader geographic view of RSQ accuracy should be prioritized towards “better generalizability” instead of “higher accuracy at limited places”. To help close this gap, we extracted multi-temporal river widths from 357,389 Landsat scenes at >3,000 river reaches globally, and used them to estimate discharge using the Bayesian AMHG-Manning (BAM) algorithm and the geomorphologically-enhanced variant (geoBAM). Our daily discharge inversions (1984–2019) using the ‘off the shelf’ algorithmic parameters exhibited acceptable performance (positive Kling-Gupta Efficiency, KGE) at 27% of the gauges for BAM and 39% of the gauges for geoBAM, amounting to ~1,000 successful inversions. Exploratory analyses to discover controls on accuracy revealed that the inversions are most sensitive to a channel shape parameter b, among six factors assessed. By introducing richer prior knowledge on the discharge seasonality and intra-seasonal variability, 1400–2000 successful inversions were derived, and further constraining the factors to their optimal ranges led to a median KGE of 0.33 for >600 gauges, up from -0.10 for the entire set, which highlights the promising potential for global RSQ. On top of the encouraging results, we present two contrasting cases that demonstrate how the relative effectiveness between RS observations and prior discharge work together to influence the inversion, pointing toward the need to better characterize the effectiveness between RS and priors for spatio-temporally explicit RSQ improvements. Overall, our critical assessment of the BAM/geoBAM algorithms shows promise for global RSQ and reveals a lower bound of SWOT discharge accuracy. It also highlights the value of assessing RSQ at global scales as we move into a new era upon SWOT’s launch.

借助卫星反演的河流水力参数（如河道宽度、水位与河道坡度）精准估算河流流量，是遥感测流（RSQ）领域的核心目标。过往诸多研究已开发并对比了多款RSQ算法，以确定适宜的测流条件，但针对全球多样河型定制RSQ方法的相关工作仍相对较少。随着计算能力、传感器技术的进步，以及地表水与海洋地形（SWOT）卫星任务即将发射，RSQ领域现已具备向全球尺度拓展的条件，此时应优先以“更强泛化能力”而非“局部区域更高精度”为目标，拓展RSQ精度的全球覆盖范围。为填补这一研究空白，我们从全球超过3000个河道河段的357389景陆地卫星（Landsat）影像中提取了多时相河道宽度，并利用贝叶斯AMHG-曼宁（BAM）算法及其地貌增强型变体（geoBAM）开展流量估算。我们基于“现成”算法参数开展的1984–2019年日流量反演结果显示，BAM算法在27%的水文测站达到可接受的性能（克林-古普塔效率Kling-Gupta Efficiency, KGE为正值），geoBAM算法则在39%的测站达到该标准，总计约1000次成功反演。针对反演精度影响因素的探索性分析表明，在评估的六项因子中，反演结果对河道形状参数b最为敏感。通过引入更丰富的流量季节节律与季内变异性先验知识，成功反演次数提升至1400–2000次；进一步将各因子约束至最优区间后，超过600个测站的KGE中位数从全数据集的-0.10提升至0.33，凸显了全球尺度RSQ应用的可观潜力。除上述令人鼓舞的结果外，我们还展示了两个对比案例，阐明了遥感观测与先验流量信息的相对有效性如何共同影响反演结果，这也指向了需进一步表征遥感观测与先验信息的有效性差异，以实现时空显式的RSQ方法优化。总体而言，我们对BAM/geoBAM算法的批判性评估显示其具备全球尺度RSQ应用的潜力，并给出了SWOT卫星测流精度的下限。同时，随着SWOT卫星任务开启遥感测流的新纪元，本研究也凸显了在全球尺度开展RSQ方法评估的重要价值。