State Key Laboratory of Earth Surface Processes and Resource Ecology, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China; Southern Marine Science and Engineering Guangdong Laboratory (Guangdong), Guangdong 511458, China.
State Key Laboratory of Lake Science and Environment, Nanjing Institute of Geography and Limnology, Chinese Academy of Sciences, Nanjing 210008, China.
Water Res. 2022 May 15;215:118241. doi: 10.1016/j.watres.2022.118241. Epub 2022 Mar 1.
Information regarding water clarity at large spatiotemporal scales is critical for understanding comprehensive changes in the water quality and status of ecosystems. Previous studies have suggested that satellite observation is an effective means of obtaining such information. However, a reliable model for accurately mapping the water clarity of global lakes (reservoirs) is still lacking due to the high optical complexity of lake waters. In this study, by using gated recurrent units (GRU) layers instead of full-connected layers from Artificial Neural Networks (ANNs) to capture the efficient sequence information of in-situ datasets, we propose a novel and transferrable hybrid deep-learning-based recurrent model (DGRN) to map the water clarity of global lakes with Landsat 8 Operational Land Imager (OLI) images. We trained and further validated the model using 1260 pairs of in-situ measured water clarity and surface reflectance of Landsat 8 OLI images with Google Earth Engine. The model was subsequently utilized to construct the global pattern of temporal and spatial changes in water clarity (lake area>10 km) from 2014 to 2020. The results show that the model can estimate water clarity with good performance (R = 0.84, MAE = 0.55, RMSE = 0.83, MAPE = 45.13%). The multi-year average of water clarity for global lakes (16,475 lakes) ranged from 0.0004 to 9.51 m, with an average value of 1.88 ± 1.24 m. Compared to the lake area, elevation, discharge, residence time, and the ratio of area to depth, water depth was the most important factor that determined the global spatial distribution pattern of water clarity. Water clarity of 15,840 global lakes between 2014 and 2020 remained stable (P ≥ 0.05); while there was a significant increase in 243 lakes (P < 0.05) and a decrease in 392 lakes (P < 0.05). However, water clarity in 2020 (COVID-19 period) showed a significant increase in most global lakes, especially in China and Canada, suggesting that the worldwide lockdown strategy due to COVID-19 might have improved water quality, espically water clarity, dueto the apparent reduction of anthropogenic activities.
关于大时空尺度水清澈度的信息对于了解水质和生态系统状况的综合变化至关重要。先前的研究表明，卫星观测是获取此类信息的有效手段。然而，由于湖泊水的高光学复杂性，仍然缺乏一种可靠的模型来准确绘制全球湖泊（水库）的水清澈度。在这项研究中，我们通过使用门控循环单元（GRU）层代替人工神经网络（ANNs）中的全连接层来捕获原位数据集的有效序列信息，提出了一种新颖且可转移的基于混合深度学习的循环模型（DGRN），以利用陆地卫星 8 操作陆地成像仪（OLI）图像绘制全球湖泊的水清澈度。我们使用谷歌地球引擎中的 1260 对原位测量的水清澈度和陆地卫星 8 OLI 图像的表面反射率对模型进行了训练和进一步验证。随后，该模型用于构建 2014 年至 2020 年期间水清澈度的时空变化全球格局（湖泊面积>10 km）。结果表明，该模型可以很好地估计水清澈度（R = 0.84，MAE = 0.55，RMSE = 0.83，MAPE = 45.13%）。全球湖泊（16475 个湖泊）多年平均水清澈度范围为 0.0004 至 9.51 m，平均值为 1.88 ± 1.24 m。与湖泊面积、海拔、流量、停留时间和面积与深度比相比，水深是决定全球水清澈度空间分布格局的最重要因素。2014 年至 2020 年间，全球 15840 个湖泊的水清澈度保持稳定（P ≥ 0.05）；而有 243 个湖泊（P < 0.05）的水清澈度显著增加，有 392 个湖泊（P < 0.05）的水清澈度显著下降。然而，由于 COVID-19 期间人为活动明显减少，2020 年（COVID-19 期间）全球大部分湖泊的水清澈度显著增加，特别是在中国和加拿大。
