Remote sensing has been widely applied to investigate fluvial processes, but depth retrievals face significant constraints in deep and turbid conditions. This study evaluates the potential for depth retrievals under such challenging conditions using NASA's Airborne Visible/Infrared Imaging Spectrometer-Next Generation (AVIRIS-NG) imagery. We employ interpretable machine learning to construct a hyperspectral regressor for water depth and explore the spectral characteristics of deep and turbid waters in Wax Lake Delta (WLD), Louisiana, USA. The reflectance spectra of WLD show minor effects from depth differences due to turbidity. Nevertheless, a Random Forest with Recursive Feature Elimination (RF-RFE) effectively generalizes high and low turbidity cases in a single model, achieving a $R^2$ of $0.94\pm 0.005$. Moreover, this model shows a maximum detectable depth of approximately 30 m, outperforming other methods. A spectral analysis using Shapley additive explanations (SHAP) points out the importance of learning various spectral bands and non-linear relationships between depth and reflectance. Specifically, the short blue and Near-InfraRed (NIR) bands, with high attenuation coefficients, play a crucial role. This finding highlights attenuation as the key process for deep-depth retrievals. The depth maps of WLD produced by this model accurately capture the spatial distribution of deep river and shallow delta regions. However, the high dependency on short blue and NIR bands leads to discontinuous areas due to the noise sensitivity of these bands. This result highlights a drawback of remote sensing using empirical models. Future research will focus on correcting such discontinuities by integrating data from multiple remote sensing sources.

中文翻译：

---

使用 AVIRIS-NG 影像在浑浊和深水区进行深度测绘：美国路易斯安那州蜡湖三角洲的研究


遥感已被广泛用于研究河流过程，但在深海和浑浊条件下，深度检索面临重大限制。本研究使用 NASA 的机载可见光/红外成像光谱仪-下一代 （AVIRIS-NG） 图像评估了在如此具有挑战性的条件下进行深度检索的潜力。我们采用可解释机器学习来构建水深的高光谱回归器，并探索美国路易斯安那州蜡湖三角洲 （WLD） 深水和浑浊水域的光谱特征。WLD 的反射光谱显示，由于浊度引起的深度差异的影响很小。尽管如此，递归特征消除随机森林 （RF-RFE） 在单个模型中有效地推广了高浊度和低浊度情况，实现了 $R^2$ 为 $0.94\pm \mathrm{0.005\$0.94}$.此外，该模型显示最大可探测深度约为 30 m，优于其他方法。使用 Shapley 加法解释 （SHAP） 的光谱分析指出了学习各种光谱波段以及深度和反射率之间的非线性关系的重要性。具体来说，具有高衰减系数的短蓝色和近红外 （NIR） 波段起着至关重要的作用。这一发现突出了衰减是深度检索的关键过程。该模型生成的 WLD 深度图准确地捕捉了深河和浅水三角洲区域的空间分布。然而，由于这些频段的噪声敏感性，对短蓝色和 NIR 频段的高度依赖性会导致区域不连续。 这一结果突出了使用经验模型进行遥感的缺点。未来的研究将侧重于通过整合来自多个遥感源的数据来纠正这种不连续性。