Monthly monitoring data from the HKEPD at 10 stations (Fig. 1) in the coastal waters off Hong Kong (113.8–114.5∘ E, 22.1–22.6∘ N) were chosen for formal analysis. Specifically, the data in use include vertical profiles of DO, temperature, salinity, dissolved inorganic nitrogen (DIN), and chlorophyll a (Chl a) concentrations measured in the water columns during 1994–2018 as well as total organic carbon (TOC) and total nitrogen (TN) measured in the sediments during 1998–2018. The survey stations can be divided into three subregions: (1) the northwestern subregion, including stations NM5 (with water depth of 20 m), NM6 (5 m), and NM8 (8 m); (2) the southern subregion, including stations SM20 (7 m), SM17 (12 m), SM18 (21 m), and SM19 (24 m); and (3) the eastern subregion, including stations MM8 (31 m), MM13 (28 m), and MM14 (25 m). Water samples were collected from the surface (1 m b.s.f.), middle (half of the depth at each station), and bottom (1 m above the sediments) layers, respectively. Details on the sampling procedures and measurements were described by Xu et al. (2010).

In addition, the monthly data of wind speeds and directions used for analysis were estimated using the daily wind observations during 1994–2018 provided by the Waglan Island automatic weather station (Fig. 1) of the Hong Kong Observatory. It should be noted that the duration of southwestern winds was defined as the number of their occurrence in days during summer. As for the freshwater inputs from the Pearl River, the monthly data during 1994–2018 were calculated using the discharge data obtained from three major hydrological stations (i.e., Gaoyao, Shijiao, and Boluo) of the Pearl River Water Resources Commission of the Ministry of Water Resources.

Several metrics, including the cross-sectional area and the layer thickness of low oxygen (DO <4 mg L-1), oxygen deficiency (DO <3 mg L-1), and hypoxia (DO <2 mg L-1) as well as the mean and minimum DO concentrations in the bottom waters, were used to depict the oxygen conditions in the region. Firstly, the observed DO profiles were interpolated with the “natural neighbor” method through MATLAB along the three subregions with a grid resolution of 600 m (distance) × 0.3 m (depth). The total areas of DO below 4, 3, and 2 mg L-1 were then calculated as the cross-sectional areas of low oxygen, oxygen deficiency, and hypoxia, respectively. The associated layer thickness was defined as the averaged thickness of the grids with DO below the corresponding levels (i.e., 4, 3, and 2 mg L-1). Regarding the island between stations NM8 and SM20, the spatial interpolations were performed directly with all the observed data and then the areas covered by the island were masked out roughly based on its size (Figs. 2, 3, and A1 in the Appendix), as the topographic data of the island were not available. Such treatment has little influence on the estimation of vertical low-oxygen areas because low-oxygen conditions were seldom found in stations NM8 and SM20. Moreover, the same treatment procedure was applied to the data for each month over 25 years to generate an interpolation set for every month, making it consistent when investigating the interannual variations in low-oxygen conditions.

In order to investigate the main variation (interannual changes) of oxygen conditions, we have introduced an indicator integrating the above metrics (except the hypoxic area and thickness, Table A1) through the PCA (principal component analysis) technique, which can reduce the dimensionality of a dataset to make it more interpretable with minimum information loss (Cadima et al., 2016). The two metrics related to hypoxia were excluded from PCA because the occurrence of hypoxia was relatively rare and its interannual variation was not as significant as that of low oxygen and oxygen deficiency. The results of PCA analysis (Table A2) showed that the first component explained most of the variance (86.40 %) for the six input variables, while the remaining components explained less variance (13.60 %). The first component was highly correlated with the interannual variations of the cross-sectional areas (with a correlation coefficient r of 0.96, p<0.01) and the thickness (r=0.96, p<0.01) of low oxygen as well as the bottom DO concentrations (r=-0.90, p<0.01), and it was thereafter referred to as “low-oxygen index” (LOI, Eq. 1) to describe the interannual severity of low-oxygen conditions comprehensively. 1LOI=-0.40×DOmean-0.39×DOmin+0.42×Area4+0.41×Area3+0.42×Thickness4+0.41×Thickness3, where DOmean and DOmin represent the mean and the minimum DO concentrations in the bottom waters, respectively; Area4 (Area3) and Thickness4 (Thickness3) represent the cross-sectional area and the thickness of low oxygen (oxygen deficiency), respectively.

As the low-oxygen conditions within Hong Kong waters were jointly affected by physical and biogeochemical processes, we attempted to quantify the relative contributions of multiple relevant factors including wind, freshwater, water temperature, and nutrients to interannual variability and long-term trends of the oxygen conditions through multiple regression. As for the selection of the wind variable in use, the daily wind data were processed into monthly average wind speed (WS), southwestern wind duration (SWWD), southwestern wind cumulative stress (SWCS), and southeastern wind cumulative stress (SECS) in summer (June–August) to examine the effect of wind speed and direction (Fig. A2). Then, a suite of multiple regressions was carried out to fit the LOI for each wind-related variable. As shown in Table A3, the fitting effect of LOI was better when using WS, which also has the highest correlation with LOI among the wind-related variables, revealing that WS explained the highest interannual variation in LOI among the wind-related factors. Therefore, WS was eventually adopted to be the wind-related input variable in the multiple regression with freshwater discharge (flow), the monthly spatial average of bottom temperature (T), and surface DIN in the summer. The resulting regression coefficients were then standardized by multiplying the ratio between the standard deviation of each input variable (e.g., WS) and the standard deviation of LOI to evaluate their interannual contributions (Eq. 2). 2Csti=Ci×SDiSDLOI, where Csti and Ci represent the regression coefficients of WS, flow, T, and DIN after and before standardization, respectively; SDi represents the standard deviation of WS, flow, T, and DIN; SDLOI represents the standard deviation of LOI.

In addition, the dataset was randomly split into a training dataset (70 %) and a testing dataset (30 %) in order to provide a more robust data fitting with estimates on the uncertainties arising from different data selections. Consequently, over 480 700 combinations of training and testing datasets were generated randomly from this splitting process and were used to build up a variety of regression models. Coefficient of determination (R2) was used to measure the fitting effect in training and testing datasets. Of all the established models, the fitting effect of training datasets (e.g., Rtrain2) and coefficients of the four variables were similar, but the predictive skills in the testing dataset (e.g., Rtest2) varied over a large range (Fig. A3, Table A4). Besides, a larger standard deviation occurred in coefficients in cases with worse testing effects. To provide a more robust estimation for the fitting, only those with R2 over or equal to 0.6 both for the training and testing datasets were selected to quantify the impact of each input variable according to their regression coefficients on average (Fig. A4). Furthermore, based on the selected models (with R2≥ 0.6 for both datasets), we also set up four sensitive experiments in which the long-term trend of each input variable was removed and only interannual fluctuations were retained. The LOI was then re-calculated in each scenario and its change relative to the original LOI was used to assess the impact of each variable on the long-term oxygen trend.

Despite the large DO fluctuations according to year, a clear deoxygenation trend could be observed in summer over the past 25 years, showing a long-term decline in the DO concentrations associated with increases in the areas and occurrences affected by low oxygen (Fig. 4). More specifically, before 2000 the spatially averaged DO concentrations in the bottom waters exceeded 4 mg L-1 and low-oxygen conditions were seldom observed (Fig. 4a), while the DO minimums were all above 2 mg L-1 (i.e., no hypoxic events occurred). However, since 2000 the occurrences of low oxygen and hypoxia have become more frequent, with a significant growth in the LOI and its related metrics, confirming the exacerbation of low-oxygen conditions in the Hong Kong waters.

To further quantify the intensity of long-term deoxygenation in summer, linear regressions were performed for the DO concentrations in different layers and in different subregions and also for the areal extents of low-oxygen conditions during 1994–2018 (Fig. 5). As shown, apparent declining trends were found for the DO series both at the surface (although not significant, Fig. 5a) and the bottom (Fig. 5b). For the bottom waters, the averaged DO concentrations displayed a decreasing pattern with a rate of 0.03 mg L-1 yr-1 (equivalent to approximately 0.7 % of the climatological DO mean at the bottom), while the DO minimums showed a more significant decline with a rate of 0.08 mg L-1 yr-1 (∼3.5 % of the climatological mean of the bottom DO minimums). It was also noted that the intensity of deoxygenation varied between subregions (Fig. 5c–h). As for the bottom DO concentrations, the most significant decrease was found in the eastern subregion (with a deoxygenation rate of 0.05 mg L-1 yr-1, Fig. 5h), while the most significant decline in the DO minimum appeared in the southern subregion (with a rate of 0.08 mg L-1 yr-1, Fig. 5f). Likewise, significant increasing trends were also found for the areas of low oxygen and oxygen deficiency (Fig. 5i–j), showing an annual growth rate at 1.95×104 m2 and 4.75×103 m2, respectively. Regarding the changes in LOI, it had a growth rate of 0.20  yr-1, which corresponds to an increasing rate of 1.99×104 m2 in the low-oxygen area and a declining rate of 0.07 mg L-1 in the DO minimum.

Furthermore, the long-term oxygen changes varied between different months of the summer season as well (Fig. 6). It could be seen that the decreasing magnitudes of the averaged DO concentration were close to each other for all the summer months, while the decline in the DO minimum was most pronounced in July (with a decreasing rate of 0.10 mg L-1 yr-1, Fig. 6c), followed by that in August (0.06 mg L-1 yr-1, Fig. 6e). In fact, the long-term changes in the DO minimum had different patterns in July and August. For July, the DO minimum generally showed a consecutive decrease over the past 25 years (Fig. 6d). While in August, the DO minimum experienced a rapid decline at a rate of 0.14 mg L-1 yr-1 during 1994–2011, which was higher than that in July during the same period (0.11 mg L-1 yr-1), but subsequently has undergone a recovery from the hypoxic conditions since 2012 (Fig. 6f). Along with such distinctive intraseasonal patterns, an interesting phenomenon was also noticed: hypoxic events were present mostly in August prior to 2012 (e.g., in 2007 and 2010–2011; no hypoxia was found in July during the same period) but since 2012 only in July instead (e.g., in 2014 and 2016–2017), as shown in Fig. 6. This finding implied a potential shift in the onset of hypoxia generation from August to July, i.e., an earlier timing for the arrival of the summertime hypoxia. Accordingly, distinct changes were found for the areas affected by hypoxia in the two periods around 2012. The hypoxic area estimated in July increased from 0 during 1994–2011 to (5.42 ± 8.77) ×103 m2 during 2012–2018, whereas the hypoxic area in August decreased from (0.89 ± 2.82) ×104 m2 to 0.

As shown above, significant interannual variabilities were observed in the spatial extent (e.g., cross-sectional area) and intensity of oxygen conditions (e.g., the mean bottom DO concentrations). Such variabilities were largely influenced by multiple physical and biogeochemical factors, including wind forcing, freshwater discharge, water temperature, and nutrient loads. These processes act jointly to affect density stratification (Yu et al., 2015), water residence time (Li et al., 2021), and temporal and spatial distributions of eutrophication parameters (Cui et al., 2018). As described in Sect. 2.2, four important influential factors (i.e., WS, flow, T, and DIN) were used to predict the interannual variations in LOI with multiple regression models, in which there have been 56 010 cases (∼12 % of the total, Fig. 7) with R2≥0.6 both in the training dataset (mean R2 of 0.64) and the testing dataset (mean R2 of 0.70). The standardized coefficients (mean ± SD) for these well-performing regression cases were given as follows: 3LOI=-(0.39±0.12)×WS-(0.14±0.12)×flow-(0.11±0.08)×T+(0.49±0.12)×DIN.

As denoted by the regression coefficients, wind forcing has exerted a significant impact on the interannual changes in LOI, with a relative contribution of 39 ± 12 % to the LOI variability explained. Its importance could also be evidenced by the significant negative correlation between WS and LOI (r=-0.67, p<0.01, Fig. 8), suggesting that calm winds were beneficial to low-oxygen conditions. In most cases (e.g., in Modaomen Bay in the PRE and the northern Gulf of Mexico), strong winds could break down stratification in the water column (Rabalais et al., 1991; Feng et al., 2012), which was conducive to water mixing and atmospheric reoxygenation (Rabalais et al., 1991). However, the weak correlation between WS and Δρ (Fig. 8) indicated that the wind forcing may control hypoxia through other alternative mechanisms. Actually, weak winds in combination with flow convergence induced by wind-driven circulation could contribute to long water residence time and nutrient accumulation in the eastern PRE and thus favor the phytoplankton blooms (Li et al., 2021). This could be supported by the significant negative correlation between WS and Chl a (r=-0.62, p<0.01). By contrast, the wind direction showed less significant effect on the interannual variability in low-oxygen conditions, as suggested by the comparatively poor performance in the LOI fitting and the weaker correlations of the wind direction-related variables with LOI (Table A3). It was noted that the monthly average wind direction in summer was generally southerly with small changes (mostly varying between 150 and 200∘, Fig. A2). Overall, our results indicated that the wind speed played a more important role in regulating the low-oxygen conditions in the coastal waters off Hong Kong from an interannual perspective, although the wind direction could significantly influence the short-term generation and development of low-oxygen conditions by modulating the Pearl River plume and material fluxes (Yin et al., 2004; Li et al., 2021). With respect to the DIN concentrations, they played a vital role in determining the interannual variabilities of the oxygen conditions, with a contribution of up to 49 ± 12%. It has been widely recognized that eutrophication stimulated by anthropogenic nutrient inputs could provide a large quantity of depositing detritus and subsequently lead to substantial oxygen depletion and occurrence of low-oxygen events (Rabalais et al., 2010; Fennel and Testa, 2018); for example, in the northern Gulf of Mexico (Feng et al., 2012; Forrest et al., 2011) and Chesapeake Bay (Wang et al., 2015), the interannual hypoxic areas in summer were directly regulated by the nutrient levels. A similar situation was found in the PRE (Li et al., 2020) and Hong Kong waters, as confirmed by the significant positive correlation between DIN and LOI (r=0.65, p<0.01). Collectively, DIN and WS were identified as the two key factors controlling the interannual changes in low-oxygen conditions.

Compared to WS and DIN, the freshwater discharges (flow) had a much smaller contribution (∼14 ± 12 %) to the variations in LOI. Generally, large freshwater inputs tend to enhance the intensity of water stratification and facilitate the generation of hypoxia (Rabalais et al., 1991). However, we found a negative correlation between flow and LOI (r=-0.45, p<0.05, Fig. 8), implying that the effect of freshwater discharges on low-oxygen conditions might involve more complex mechanisms and act through indirect pathways. Due to their long distance from the river outlets of the Pearl River, the coastal waters off Hong Kong were relatively less influenced by terrestrial inputs (Yu et al., 2021) and the effect of freshwater discharge and its carrying of organic matter in this area was not as significant as that in other subregions (e.g., the upper reach of Lingdingyang Bay and the western PRE). Nevertheless, freshwater discharge in combination with the wind-driven circulation could significantly affect the water residence time (Sun et al., 2014) and nutrient accumulation in the Hong Kong waters (Li et al., 2021). Specifically, the weakened discharge could prolong the retention of nutrients and thereby stimulate local production of organic matter in the region (Li et al., 2021), which ultimately promoted oxygen depletion. Regarding the water temperature, previous studies showed that it could exert significant influence on coastal hypoxia largely by regulating water stratification intensity, oxygen solubility, and microbial respiration rate (Breitburg et al., 2018). However, our results showed that the contribution of water temperature (T) to the LOI changes (∼11 ± 8 %) was not significant in the Hong Kong waters, as revealed by its weak correlation with LOI as well. Given the fact that the Hong Kong waters are a region heavily affected by human activities, the effect of temperature (e.g., global warming) might be more significant in the region with a larger geographic scale. Overall, the role of temperature and freshwater discharges in regulating the interannual oxygen variability in the Hong Kong waters appeared to be secondary.

The data over the past 25 years showed that the coastal waters off Hong Kong have experienced a notable long-term oxygen decline, especially for the DO minimum in the bottom waters. Based on the observed deoxygenation rate, the bottom DO minimum was expected to decrease by approximately 15 %–70 % in 5–20 years (reaching a level of 0.4–1.6 mg L-1) compared to the climatological mean of 1994–2018. The impacts of influential factors on the long-term deoxygenation trend were then evaluated using the regression models mentioned in Sect. 4.1 and quantified by the relative changes of LOI in the sensitive experiments (see details in Sect. 2.2) compared to the original one. It was noted that WS exhibited a decreasing trend of 0.03 m s-1 yr-1 (p<0.05, Fig. 9a) within the coastal regions off Hong Kong over the past 25 years, while a similar situation was also found in the Pearl River Basin (Zhang et al., 2019) and the northern South China Sea (Gao et al., 2020) due to the long-term climate changes (Xu et al., 2006; Zhang et al., 2009; Chen et al., 2020). Meanwhile, DIN showed an increasing trend with a rate of 0.01 mg L-1 yr-1 (p<0.01, Fig. 9e). The growth in DIN and decline in WS led to a 56 ± 10 % and 39 ± 14 % increase in LOI (Table 1), respectively, indicating that DIN and WS were the main driving factors for the long-term deoxygenation. On the other hand, significant long-term trends were also found for the freshwater discharges (with a decreasing rate of 4.19×102 m3 s-1 yr-1, Fig. 9b) and water temperature (with an increasing rate of 0.06 ∘C yr-1, Fig. 9d), but their impacts were relatively small, resulting in a 16 ± 14 % increase and 11 ± 9 % decrease in LOI, respectively.

Despite the different influences of the factors mentioned above, they were likely to exert synergetic impacts on the low-oxygen conditions by aggravating eutrophication as discussed earlier; it was observed that the long-term growth in Chl a (with a rate of 0.15 µg L-1 yr-1, Fig. 9f) matched well with the increase in LOI. Specifically, the significant increase in phytoplankton biomass was primarily due to the combined effects of more stable water-column conditions and longer residence time facilitated by the weakened wind forcing and river discharges, higher nutrient levels, and lower water turbidity (Fig. 9g) in recent years. Consequently, the elevated organic matter through phytoplankton primary production would lead to strong oxygen consumption, thereby contributing to an expansion of low-oxygen conditions in terms of areal extent and intensity.

Moreover, with massive algal fragments provided by primary production, the composition of organic matter in the coastal waters off Hong Kong has probably changed and would cause substantial changes in the timing of hypoxia generation. As noted in Sect. 3.2, the onset of hypoxia was observed to shift from August to July around 2012. To explore this issue, we first used the ratio of TOC to TN measured in the sediments to estimate the main source of organic matter, with values of 14–30 pointing to a terrestrial source and values of 4–10 indicating a marine source from in situ production (Bordovskiy, 1965; Meyers and Ishiwatari, 1993). It is clear that the TOC:TN showed a significant decreasing trend and was mostly below 10 from 2012 on (Fig. 9h). This implied a shift in the dominant source of organic matter from terrestrial inputs to local production (marine sourced). As such, oxygen consumption became faster because the marine-sourced organic matter was fresher and more active (Raymond and Bauer, 2001) and therefore the time required to reach hypoxia would be shortened. Furthermore, changes in the physical conditions provided sufficient time for a more thorough decomposition of organic matter in July, which left less organic matter for August and thus weakened the deoxygenation therein.

Similarly, the long-term oxygen changes in terms of the areal extents and arrival timing of hypoxia have also been found in other coastal systems. For example, in Chesapeake Bay, sea level rise and elevated freshwater discharges would lead to an approximately 10 %–30 % increase in hypoxic volume between the late 20th and the mid-21st centuries (Ni et al., 2019), while the increase in water temperature would cause hypoxia to develop 5–10 d earlier in ∼30 years (Ni et al., 2020). In the northern Gulf of Mexico, the growth in riverine nutrient inputs would result in an increase in the frequency of hypoxia occurrence by 37 % (Justić et al., 2003). While in the Hong Kong waters, low-oxygen conditions would develop into hypoxic conditions in two decades with larger areal extent and earlier arrival ascribed to the ongoing alterations in physical conditions and nutrients as mentioned earlier. This inference was based on the assumption that the external factors (e.g., wind speed, DIN, discharges) would change at the same rates as those in the past 25 years. Although the real situation would be more complicated and compounded by factors such as the implementation of management and nonlinear changes in climatic factors, our findings still served as an alarming signal that changes in wind and freshwater discharges could cancel out potential benefits of nutrient management. To this end, it is of great importance to conduct long-term and more intensive control of nutrient inputs in order to mitigate the low-oxygen conditions in the region.