Lake Taihu is the third largest freshwater lake in China (Figure 1), with a water area of 2,338 km² and an average depth of 1.9 m (Wang and Dou 1998). It is located in a subtropical area that is warm and wet in summer but cold and dry in winter. Lake Taihu is found on the lower reach of the Yangtze River, and the area around it is one of the most developed regions in China. Excessive human activities have intensified eutrophication and cyanobacterial blooms since the 1980s (Qin et al., 2007). The lake supplies water to approximately 10 million residents of surrounding cities, including Wuxi, Suzhou, and Huzhou. Thus, the water quality of Lake Taihu is essential for local human activities and needs, such as drinking, tourism, fishing, and shipping. Lake Taihu is usually divided into seven subregions: Zhushan Bay (ZSB), Meiliang Bay (MLB), Gonghu Bay (GHB), West, Center, South, and East. Some areas in the MLB, GHB, and East regions are frequently covered by macrophytes, affecting the retrieval of Chla and SDD (Shi et al., 2017), and thus, these areas are not included in the analysis.

Monthly/seasonal surveys from 2007 to 2015 were conducted by the Taihu Lake Laboratory Ecosystem Research (TLLER) Station to collect water quality parameters (Min et al., 2019) (Figure 1). A total of 847 data samples were collected here after excluding the outliers, e.g., stations covered with cyanobacterial scums and macrophytes. These data included DO, water temperature (WTR), Chla, and SDD. The WTR and DO at each station were measured using a well-calibrated YSI probe (Yellow Springs, OH 45387 United States) (Table 1). A standard 30 cm diameter Secchi disk was used to measure SDD. At each station, water samples at the surface layer (0.5 m) were collected and stored in pre-cleaned 1 L high-density polyethylene bottles. The water samples were strained through glass fiber filters (0.70 μm pore size, Whatman GF/F), and Chla concentration was spectrophotometrically determined using a Shimadzu UV2700 spectrophotometer after the extraction of pigments by using 90% ethanol (Jeffrey and Humphrey 1975).

In addition, daily mean temperature (°C) and wind speed (m s⁻¹) at Dongshan meteorological station near Lake Taihu (Figure 1) from 2002 to 2021 were downloaded from the National Meteorological Information Center, China (http://data.cma.cn). These data were further aggregated into monthly and annual mean values from 2002 to 2021.

Two types of satellite data were used in this study: 1) MODIS Aqua Level 1 A data for retrieving Chla and SDD in Lake Taihu, and 2) MODIS land surface temperature (LST) (MYD11A1) products.

Three machine learning models, namely, random forest (RF), eXtreme gradient boosting (XGB), and support vector regression (SVR), which have been used for retrieving water quality parameters (Sagan et al., 2020; Cao et al., 2022c), were utilized for retrieving DO in Lake Taihu. Given the various predicting mechanism among the three models, which one exhibited the best performance remained unknown. First, we used 864 in situ data to train and validate the models for comparing and selecting the optimal one. Then, we examined the performance of the optimal model in retrieving DO from 58 matched MODIS samples.

Compared with the use of WTR and Chla in DO estimates in previous studies (Kim et al., 2020; Guo et al., 2021), the input to the models included WTR, Chla, and SDD. The output variable was DO concentrations. WTR alters thermal properties and affects the solubility of DO (Jankowski et al., 2006), while Chla reflects primary productivity, which is closely related to photosynthesis and respiration. SDD is a crucial factor for quantifying light attenuation in a lake column, possibly regulating the lake mixing and vertical distribution of DO (Zhang et al., 2015; Liu et al., 2020). It has also been utilized to estimate carbon dioxide in lakes (Qi et al., 2020). The field data suggested that these variables exhibited considerable correlations with DO during different seasons in Lake Taihu (Figure 2). Following our experiments, all three models trained with the three inputs outperformed the models with two inputs (i.e., WTR and DO).

We randomly chose approximately 70% (n = 499) of the matchups for training, and the remaining 30% of data (n = 265) were used to test model performance. All the input and output data were log-transformed and standardized using the mean and standard deviation within the 0–1 range before training the models. The hyperparameters of the RF, XGB, and SVR models were determined using a grid search method.

The well-validated machine learning model was used to retrieve DO from cloud-free MODIS R_rc and LST images. The annual and monthly mean DO values from 2002 to 2021 were further aggregated from the daily DO series. The mean DO values for each subregion were calculated from the clipped images by using specific boundaries. Pearson correlation was utilized to explain the relations between two variables, i.e., air temperature and DO. The correlation was significant at p < 0.05. We used the determination coefficient (R ²), RMSE, MAPE, median symmetric accuracy (MdSA), and the symmetric signed percentage bias (SSPB) (Morley et al., 2018) to evaluate the performance of the models (Eqs. 3–5). All statistics were collected in Python 3.8 environment. M A P E = 100 × 1 N ∑ i = 1 N E i − M i x i , (3) M d S A = 100 × 10 ζ − 1 , ζ = M e d i a n log 10 E i M i (4) S S P B = 100 × s g n ( Q ) ( 10 Q − 1 ) ,   Q = M e d i a n ( log 10 E i / M i ) (5) where N is the number of data pairs; the subscript i denotes individual data points; and E and M represent measured and estimated values, respectively.

The field dataset from 2007 to 2015 (Table 2) indicated that Lake Taihu had a mean DO of 9.21 ± 2.13 mg L⁻¹ (mean ± standard deviation) ranging from 2.3 mg L⁻¹ to 16.9 mg L⁻¹. Seasonal variation in DO was apparent, and exhibited lower in winter than in summer, which was highest in January (11.86 ± 1.17 mg L⁻¹) and lowest in July (7.69 ± 2.02 mg L⁻¹). The relations of DO to the related indicators during different seasons are illustrated in Figure 2. Generally, the seasonal distribution of DO was reversed with that of WTR and similar to that of SDD and Chla. DO exhibited a significantly negative correlation with WTR (p < 0.05) except in summer, which showed a slightly positive correlation (r = 0.17, p < 0.05). The positive relation in summer might resulted from the contributions of WTR and other factors. SDD was positively correlated with DO but only significant during summer (r = 0.13, p < 0.05). In terms of Chla, we found it had a significant positive correlation with DO, except during autumn. This analysis indicated that WTR, SDD, and Chla presented significant relations to the seasonal variation of DO, and the effect of SDD on DO during summer is peculiar. The relations provided the foundation for estimating the DO concentration through the above three parameters.

The performance of the XGB, RF and SVR models on the 265 in situ samples is presented in Figure 3. Notably, the inputs of these samples (i.e., WTR, Chla, and SDD) are in situ measurements. The three machine learning models performed satisfactorily with the <1.1 mg L⁻¹ RMSE and <10% MAPE. Moreover, these data pairs were distributed evenly along the unity for DO ranging from 2 mg L⁻¹ to 15 mg L⁻¹ in Lake Taihu and did not present evident deviations. Among the three models, XGB (R ² = 0.77, RMSE = 0.98 mg L⁻¹, MAPE = 7.9%) slightly outperformed RF (R ² = 0.77, RMSE = 0.99 mg L⁻¹, MAPE = 8.2%) and SVR (R ² = 0.77, RMSE = 1.1 mg L⁻¹, MAPE = 9.4%).

The XGB model was further examined on the 58 MODIS-derived WTR, SDD, and Chla points to determine its integrity in estimating DO from satellite images (4.8–16.0 mg L⁻¹) (Figure 4). MODIS-derived DO performed a satisfactory consistency with the measured values (RMSE = 1.28 mg L⁻¹, MAPE = 9.9%). It should be noted that a slight underestimation for points with DO more than >15 mg L⁻¹ was observed. The XGB model was inferred to be robust and suitable for mapping DO in Lake Taihu from MODIS images.

The well-validated XGB model was utilized to generate the DO series in Lake Taihu from 2002 to 2021. Overall, Lake Taihu had an average DO of 9.3 ± 1.8 mg L⁻¹ over the past 20 years (Figure 5A). DO in the northern three bays (9.8 ± 1.9 mg L⁻¹) was higher than in the central (9.2 ± 1.7 mg L⁻¹) and southern (9.1 ± 1.8 mg L⁻¹) regions. Notably, DO covered with macrophytes were omitted here. We also generated the WTR, Chla, and SDD in Lake Taihu since 2002 (Figures 5B-D). The spatial pattern of DO was similar with Chla and inversely related to that of SDD. Meanwhile, it did not exhibit specific relations with WTR.

The annual variations in DO from 2002 to 2021 are mapped in Figure 6, and the corresponding statistics are illustrated in Figure 7. The interannual DO variations of Lake Taihu were divided into two stages (Figure 7A): 1) significantly increased from 2002 to 2009, with a slope of 0.16 mg L⁻¹ (R ² = 0.63, p < 0.05); and 2) slightly declined from 2011 to 2021, with a slope of −0.04 mg L⁻¹ (R ² = 0.30, p < 0.05). The annual variations in DO of different subregions also presented similar trends (Figure 7B). The annual mean DO was different in various stages. For example, ZSB had the highest DO before 2010, while DO in MLB was the highest after 2010. DO in central and southern Lake Taihu was lowest all the time. We observed that annual mean air temperature decreased from 2002 to 2010 (slope = −0.07 °C, R ² = 0.20, p < 0.05) and exhibited a dramatic increase since 2010 (slope = 0.21 °C, R ² = 0.83, p < 0.05, Figure 7C). The annual air temperature was significantly negatively correlated with DO in Lake Taihu (R ² = 0.21, p < 0.05). In addition, wind speed showed a continuous decline and was not significantly correlated with DO (Figure 7D).

Figure 8 presents the monthly mean DO in Lake Taihu from 2002 to 2021. We found that DO in winter was higher than that in summer. DO was highest in January (11.9 mg L⁻¹) and lowest in August (7.7 mg L⁻¹) (Figure 9A). This finding was consistent with the aforementioned analysis based on the field dataset (Table 1). The monthly variations in DO were negatively correlated with air temperature (R ² = 0.80, p < 0.05), while wind speed was not significantly associated with it (p > 0.05, Figures 9C,D).

A two-step model was developed to estimate DO from MODIS images in Lake Taihu (Figures 3, 4). The model was established on the theoretical relations between thermal and optical properties and DO in lakes (Kim et al., 2020; Guo et al., 2021). Such an idea for estimating non-optically active matters have been employed in other studies (IOCCG 2018; Chen et al., 2019; Qi et al., 2020). The current work developed a machine learning model to determine the relations between DO and the aforementioned three indicators rather than the regressions (Kim et al., 2020). In general, their relations in eutrophic lakes are complicated and exhibit spatiotemporal heterogeneity (Figure 2) (Wetzel 2001; Jankowski et al., 2006; Breitburg et al., 2018). The input variables had various contributions to the DO in different seasons (Figure 2) and regions for Lake Taihu. The properties of water in norther regions were mainly influenced by algal while central and south regions were turbid. Compared with traditional regressions such as linear/non-linear regression and step-wise regression, machine learning models are particularly efficient for solving complicated non-linear regression (Sagan et al., 2020). The model was developed using 864 samples that spanned different seasons across 8 years in Lake Taihu, suggesting that the model was suitable for most cases in Lake Taihu.

Although machine learning models demonstrate the nature of a “black box,” the relative contributions of input variables to DO prediction can be useful for understanding the mechanism of a model. We calculated the decrease in the accuracy score of the models for each variable to interpret the contribution of each variable to Chla (Cao et al., 2022a). The decrease in accuracy score was defined as the difference between the baseline metric from permutating the feature column, which was implemented in the scikit-learn package of Python. Figure 10 reveals that WTR makes the highest contribution to DO estimation while Chla and SDD have low contributions. Thus, the model still estimated satisfactory DO in Lake Taihu even though the retrievals of Chla and SDD suffered from fair uncertainty (−30%) (Figure 4). The retrieval of Chla in turbid waters frequently depends on the red edge band near 700–710 nm (Gitelson 1992; Gilerson et al., 2010; Gurlin et al., 2011), which is not equipped with MODIS instrument.

Despite the satisfactory DO retrievals from MODIS images in Lake Taihu, several limitations must be improved. First, R_rc used for retrieving Chla and SDD in Lake Taihu did not remove the signals of aerosol contributions, which might limit the accuracy of DO estimations. For turbid waters, including Lake Taihu, the elevating water-leaving radiance suggested that R_rc can be utilized to retrieve water quality (Hu et al., 2004; Cao et al., 2020; Seegers et al., 2021). We also found that the XGB model underestimated DO slightly in the high range (Figure 4), possibly resulting from insufficient samples in the extremely high DO data (Stock 2022). It is anticipated to improve by adding more high DO values (Cao et al., 2020). The surface temperature products of MODIS had a resolution of 1 km, which was lower than Chla and SDD. The pixels were not changed, although we resampled it to 250 m. Since water is frequently homogenous, WTR between adjacent pixels may exhibit slight differences. In addition, DO in shallow lakes might change fast due to wind-induced reoxygenation and diurnal variations in air temperature. Thus, it would be efficient to improve the observations of DO in lakes through the Geostationary satellites (IOCCG 2012), such as Geostationary Ocean Color Imager (GOCI), GOCI-II, and Himawari-8. Our model was developed for MODIS instruments; however, the MODIS mission operation has exceeded its anticipated lifetime and is nearing its end. In the future, the model is expected to be extended to the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard SNPP and NOAA-20/21 and Ocean Land Color Instrument (OLCI) onboard Sentinel-3 for continuing observations (Cao et al., 2022b).

The factors that regulate DO in lakes include the physical processes induced by light, WTR, and lake mixing, and biochemical factors, such as the photosynthesis-induced increase of DO concentration, the respiration of aquatic organisms, the bacterial oxidation of organic matter, and the consumption of DO by other reduced inorganic substances (Zhang et al., 2015). In addition, some anthropogenic modifications of the environment, such as eutrophication (Müller et al., 2012), salinization, and hydrological management (Carpenter et al., 2011) can reduce DO in lakes.

We found that the temporal (interannual and seasonal) variations in the DO of Lake Taihu were related to the variability of air temperature (Figure 7). First, air temperature determined the solute of oxygen in lakes by influencing the WTR indirectly (Jane et al., 2021). The general narrative is that climate warming induces widespread deoxidation in waters (Jankowski et al., 2006; Perron et al., 2014; Zhang et al., 2015; Jane et al., 2021). In deep waters, surface temperature warming can further intensify thermal stratification, reducing water circulation and preventing deep water DO replenishment (Jankowski et al., 2006; North et al., 2014; Kraemer et al., 2015). In Lake Taihu, a shallow lake, such an effect should mostly occur during summer (Yang et al., 2018). In addition, wind speed in Lake Taihu over the past 20 years has declined (Figure 7D), suggesting that turbulence should be weakening (Macintyre 1993; Fernández Castro et al., 2021), possibly reducing DO replenishment.

The spatial distribution of DO in Lake Taihu is consistent with that of Chla (Figure 5). Higher Chla reflects high primary productivity that releases plenty of oxygen via photosynthesis. The higher water clarity in northern areas facilitated the growth of phytoplankton in lakes (Liu et al., 2020). In addition, wind direction might regulate the spatial variations in the DO of Lake Taihu. The wind direction for Lake Taihu was usually southwest each year, suggesting that the stronger mixing process induced higher DO concentration (Zhang et al., 2014).

This research successfully tracked the long-term DO variations of Lake Taihu, allowing us to reveal its trends and elucidate its potential driving factors. In the past, the monitoring of DO in lakes was mostly based on field surveys (Jane et al., 2021). Operational surveys have been conducted on some well-studied lakes, such as Lake Taihu and Lake Erie. However, most lakes cannot be well monitored, which should largely limit our understanding of aquatic health and ecology (Plisnier et al., 2022). Our approach provides a practical idea of employing satellite images to monitor DO in lakes. This idea can be easily extended to other lakes, although the model’s coefficients should be recalibrated using the local dataset. By contrast, the methodology can be utilized to estimate other non-optical water quality parameters, such as total nitrogen, total phosphorus, and the permanganate index. With the relations between OACs and meteorological data and the non-optical parameters, it would be possible to estimate more water qualities in lakes. The MODIS-derived DO of Lake Taihu in the past 2 decades demonstrated declining trends compared with the first decade of this century due to climate warming, which is consistent with the conclusion of previous studies (Jankowski et al., 2006; North et al., 2014; Jane et al., 2021). Climate warming was also regarded as a primary regulator that affected cyanobacterial blooms in Lake Taihu (Qin et al., 2019). The warming climate will continue in the future (Woolway and Merchant 2019), and it is crucial to formulate scientific strategies to prevent the negative ecological effects of deoxidation in lakes.