Multi-method non-parametric assessment of river water-quality trends and change-points in the Danshui River basin, metropolitan Taipei (2010–2026)

The Danshui River drains metropolitan Taipei, where two decades of expanding sewer connection and riparian constructed wetlands have targeted river pollution. Using a continuous monthly record of 1,921 samples from 11 Ministry of Environment stations (2010–2026), we assessed trends and change-points in six measured water-quality variables and the River Pollution Index (RPI), a composite index calculated from four of them, through a multi-method non-parametric design: the Mann–Kendall and seasonal Mann–Kendall tests, a variance-corrected modified test, Theil–Sen slopes, the Pettitt and sequential-Mann–Kendall change-point tests, Innovative Trend Analysis, and a synchronized-permutation Monte Carlo test of field significance. Suspended solids, ammonia, and five-day biochemical oxygen demand declined at 9, 8, and 6 of 11 stations, each field-significant (Monte Carlo p ≤ 0.006); the basin-mean River Pollution Index fell from 3.83 to 3.20 and the severely polluted sample fraction from 11.1% to 2.5% between the early (2010–2012) and late (2023–2025) sub-periods, the incomplete 2026 record excluded. These declines persisted on the 8 complete-record stations, so they are not an artifact of network turnover. Dissolved oxygen did not improve, even as water temperature rose significantly (+0.15 °C yr - ¹), so oxygen conditions merely held against a warming that lowers solubility. Against this improvement, the coliform group rose at most stations, its change-points scattered across 2013–2024 rather than forming a single step; the largest treatment plant’s effluent-coliform log-removal collapsed after 2024, and river coliform was not associated with rainfall, implicating under-disinfected point-source effluent. The basin’s management priority has shifted from organic and nutrient control toward microbial risk and disinfection.

Rivers that drain large cities carry the concentrated weight of dense population, industry, and impervious land, and they rank among the most degraded freshwater systems anywhere. The pressures identified two decades ago (Vörösmarty et al., 2000; Meybeck, 2003; Dudgeon et al., 2006) have since been quantified with far better data. Roughly half of the wastewater produced worldwide still reaches the environment untreated (Jones et al., 2021); once surface-water quality is counted alongside quantity, the share of the world’s population under severe water scarcity in an average month rises from about 30% to 40% (van Vliet et al., 2021), and 56–66% of people are projected to face clean-water scarcity for at least 1 month a year by the end of the century (Jones et al., 2024). Freshwater vertebrate populations have fallen more than twice as steeply as terrestrial or marine ones (Tickner et al., 2020); few large rivers still run unimpounded (Grill et al., 2019; Best, 2019); and a synthesis of 965 case studies found river water quality deteriorating under droughts and heatwaves in two-thirds of cases and under long-term warming in more than half (van Vliet et al., 2023). Even where treatment is already extensive, the fecal indicators are the constituent that most often breaches quality thresholds, and they do so mainly in low-flow months (Jones et al., 2022). To manage such a river is first to measure it well, and then to read the measurements with tools able to separate a genuine trend from the strong seasonal and hydrological noise that urban watercourses generate.

The Danshui River and its tributaries drain the Taipei basin, the political and economic heart of Taiwan and one of the most densely settled regions on earth. Decades of domestic, industrial, and agricultural discharge left these rivers heavily polluted and prompted a sustained program of repair: households were connected to public sewers, black water was intercepted, and, because sewer coverage was at first incomplete, treatment wetlands were built within the river levees. Constructed wetlands have five decades of service as low-cost systems for reducing organic and nutrient loads (Vymazal, 2011), and in Taiwan they have been deployed specifically for non-point-source control and river purification (Kao et al., 2001; Wu et al., 2010). Together these works turned the river corridor into a site of active ecological restoration, with constructed wetlands as its most visible instrument, and raised a question that only a long record can answer: how far that restoration has improved the river’s function, and which pollutant burdens it has left in place. An analysis of this basin over 1993–2010 concluded that water quality had begun to recover as sewer connection widened and the wetlands came into service, using the Mann–Kendall and seasonal Mann–Kendall tests on dissolved oxygen, five-day biochemical oxygen demand (BOD 5), suspended solids, ammonia, and fecal-indicator bacteria (Cheng et al., 2011).

A decade and more has passed since that assessment, and the intervening years saw the major sewerage works reach maturity. Whether the recovery through 2010 continued, stalled, or turned is an open and policy-relevant question. The non-parametric toolkit for answering it has itself matured. Rank-based tests such as Mann–Kendall remain the standard for water-quality series because they tolerate non-normal, censored, and missing values (Hirsch et al., 1991; Van Belle and Hughes, 1984), and both their statistical foundations and their practical selection are well set out (Helsel et al., 2020; Machiwal and Jha, 2008). Their limitations, too, are now better understood and better remedied. Serial correlation, which inflates the apparent significance of a trend, can be corrected in the test variance (Hamed and Rao, 1998; Yue et al., 2002b); an abrupt shift in a record can be located by a change-point test (Pettitt, 1979); and Innovative Trend Analysis can show whether the low, middle, and high parts of a distribution move together or apart—something no single rank statistic reveals (Şen, 2012; Kisi and Ay, 2014). Composite indices, in turn, condense several variables into one policy-legible measure (Sutadian et al., 2016). Methods of this kind have traced long-term change in waters as different as the Ebro (Bouza-Deaño et al., 2008), the Han (Chang, 2008), the Euphrates and Tigris (Odemis et al., 2010), and Eymir Lake (Yenilmez et al., 2011).

This study carries the earlier Danshui assessment forward in time and asks four questions of the 2010–2026 record. First, has the recovery documented through 2010 continued, stalled, or reversed, and for which constituents? Second, when did the changes occur, and did chemically distinct variables shift together, as a basin-scale intervention would produce, or separately? Third, have two decades of wetland-based restoration and sewer connection improved the river’s functional condition in every respect, or have they left some pollutant burden in place? Fourth, are the apparent trends real basin-wide signals or artifacts of seasonality, serial correlation, network turnover, and the multiplicity of testing 11 stations at once? The record is read with a convergent set of rank-based and change-point tests, described in the next section, so that each question is answered by more than one line of evidence; the methodological contribution is to show what such a design adds to the single-test assessments that dominate the urban-river literature.

Study area

The Danshui River basin drains about 2,726 km² of northern Taiwan and empties into the Taiwan Strait. Its main watercourses are the Danshui main stem and its three principal tributaries, the Dahan Creek, the Xindian Creek, and the Keelung River, with the Jingmei Creek entering the Xindian from the east; all thread through the Taipei and New Taipei conurbation. The Dahan and the Xindian meet at Jiangzicui to form the main stem, which receives the Keelung River at Guandu before reaching the sea (Figure 1). The riparian corridor long received large volumes of untreated wastewater; repair since the 2000s has combined trunk-sewer construction, household connection, and treatment wetlands set within the floodplain levees (Cheng et al., 2011).

Data source and quality assurance

The monthly records come from the national environmental water-quality monitoring database of Taiwan’s Ministry of Environment, the successor to the Environmental Protection Administration. Eleven stations within the Danshui basin were used, listed here from upstream to downstream along each river: six on the Keelung River (KL6 Nanhu, KL5 Chengmei, KL4 Minquan, KL3 Dazhi, KL2 Zhongshan, and KL1 Bailing Bridges), two on the Jingmei Creek (JM1 Wanshou and JM2 Bao Bridges), one on the Xindian Creek (XD1 Huajiang Bridge, just above the Jiangzicui confluence), and two on the Danshui main stem (TS2 Zhongxiao and TS1 Chongyang Bridges). The Keelung and main-stem codes are numbered from downstream (KL1, TS1) to upstream; the Jingmei codes from upstream (JM1) to downstream; the figures and the supplementary tables list stations from upstream to downstream along each river, and station attributes appear in Supplementary Table S1. Sampling is monthly, a frequency adequate for basin-scale trend and status assessment (Naddeo et al., 2007), and the record runs continuously from 2010 to 2026. The monitoring network was not static: two stations (XD1, JM2) were retired in 2022 and one (TS2, Zhongxiao Bridge) was added the same year, so eight stations carry an unbroken 2010–2026 record while three do not. Because a shifting network can bias any basin-wide aggregate, every headline trend is re-tested on the fixed eight-station subset with complete records (the robustness analysis); TS2, whose 49 samples span only 4 years, is never allowed to drive a basin conclusion. The record for 2026 covers January–May only; being restricted to the cooler, better-oxygenated dry-season months, it is excluded from every annual-mean, sub-period, and trajectory statistic, and is shown in the figures solely to mark where the series ends.

Several period-specific database exports were merged. Overlapping and replicate entries were removed by de-duplicating on station and sampling date, which left 1,921 unique samples. Entries flagged “‐‐” were treated as missing, and values reported below a detection limit—“<4.0”, for instance—were set to half that limit, the usual substitution for rank-based work (Gilbert, 1987). Dissolved oxygen (DO) was taken from the electrode method, the titration value serving only where the electrode reading was absent. As an independent check on the parse, the four River Pollution Index (RPI) sub-indices were recomputed from DO, BOD 5, suspended solids (SS), and ammonia nitrogen (NH 3 -N) and the index re-derived; the reconstruction matched the database value exactly for 95.7% of samples, with a mean absolute difference of 0.03 index units, which confirms that columns and units were read correctly. Water temperature, recorded alongside every sample, was retained to interpret the dissolved-oxygen record (the Discussion).

To move beyond the time series toward the physical drivers that interpreting these trends requires, four ancillary datasets were assembled for the same basin and period: monthly treatment volumes for the basin’s municipal plants (Bali, Danshui, Linkou, Sanying, and Taipei-University water-resource centers, 2016–2025); monthly influent and effluent water quality—including the coliform group—for the Bali plant (the basin’s dominant discharge, 2022–2026) and the Dihua reclamation plant (2025–2026); daily rainfall from 42 gauges across the catchment (2007–2026), aggregated to monthly basin means; and the Ministry’s household sewer-connection and sewage-treatment-rate statistics (2024). These datasets support the covariate and mechanism analyses reported below (the analysis of drivers; the discussion of microbial persistence).

River Pollution Index

The River Pollution Index is Taiwan’s standard composite measure of organic river pollution. Each of four variables—DO, BOD 5, SS, and NH 3 -N—receives a sub-index of 1, 3, 6, or 10 according to fixed concentration bands, and the RPI is their arithmetic mean (Equation 1),

each S denoting the sub-index of the named variable. A reach is classed as unpolluted or negligibly polluted at RPI ≤ 2.0, lightly polluted at 2.0 < RPI ≤ 3.0, moderately polluted at 3.0 < RPI ≤ 6.0, and severely polluted above 6.0. The RPI is thus a calculated index, derived for each sample from its measured DO, BOD 5, SS, and NH 3 -N values, and not a quantity measured directly in the water; we refer to it throughout as a derived index rather than as a seventh measured variable. Because its four components are also analyzed individually, the trend and field-significance statistics reported for the RPI integrate, and do not independently corroborate, those of its components. Because the index folds oxygen, organic matter, solids, and nutrients into a single interpretable number, we use it as the summary measure of status alongside the individual variables, in keeping with standard water-quality assessment practice (Chapman, 1996).

Trend and change-point methods

Six measured variables (DO, BOD 5, chemical oxygen demand (COD), SS, NH 3 -N, and the coliform group) and the calculated RPI were analyzed; for each station and each of these seven series a chronological monthly record was assembled over 2010–2026.

The earlier study of this basin (Cheng et al., 2011) rested on the Mann–Kendall (MK) and seasonal Mann–Kendall (SMK) tests with Sen’s slope, which tell whether a monotonic trend exists and how steep it is. We keep those tests for comparability and add others that answer three questions they cannot. Does an apparent trend survive the serial correlation and the multiple-comparison structure of a multi-station monthly network? The modified Mann–Kendall (mMK) test, which corrects the variance for autocorrelation, and the Monte Carlo field-significance test, which asks whether the count of significant stations exceeds chance, address this. When did a change occur? A monotonic test gives direction but not date, so we add the Pettitt and sequential Mann–Kendall change-point tests. Do the low, middle, and high concentrations move differently? Innovative Trend Analysis (ITA) resolves what a single rank statistic conceals. The composite RPI and COD were added as integrative, policy-facing measures. These were not decorative additions: as the Discussion shows, the field-significance test reclassified dissolved oxygen from apparently improving to indistinguishable from chance, the change-point tests showed that the coliform rise—unlike the coherent chemical shift—has no single basin-wide onset, and ITA separated source removal from dilution in the ammonia record. We stress at the outset that these tests are not independent lines of evidence: MK, SMK, mMK, and ITA are all rank-based and draw on the same data, so their agreement is a check on robustness to method choice, not a multiplication of proof. The one test that guards against a genuinely separate failure mode—false positives from multiplicity and spatial covariance—is the Monte Carlo field-significance procedure, and it is the arbiter we privilege where tests disagree.

Mann–Kendall test

The rank-based MK test (Mann, 1945; Kendall, 1938), set out at length by Kendall (1975), tests the null of no monotonic trend against a two-sided alternative. For n observations ordered in time the statistic is given by Equation 2,

with the sign function defined in Equation 3,

and a variance, corrected for tied ranks, given by Equation 4,

where the correction sums over the g groups of tied ranks. The standardized statistic is given by Equation 5.

The MK test is distribution-free, tolerant of outliers and of gaps, and thus the natural choice for water-quality series (Van Belle and Hughes, 1984); a positive Z marks an increasing trend and a negative Z a decreasing one, and significance is reported at α = 0.05, 0.01, and 0.001.

Seasonal Mann–Kendall test

To strip out the pronounced within-year cycle of urban river quality, the SMK test (Hirsch et al., 1982; Hirsch and Slack, 1984) computes MK statistics within each of the 12 calendar months and combines them, so that every observation is compared only with others from the same season.

Modified Mann–Kendall test

Positive serial correlation inflates the false-positive rate of the ordinary test. The modification of Hamed and Rao (1998) rescales the variance of the MK statistic by the significant autocorrelation of the ranked series, and we report it as a check on the ordinary result (Yue et al., 2002a). Related corrections adjust the effective sample size (Yue and Wang, 2004) or allow for long-term persistence (Hamed, 2008).

Theil–Sen slope

The rate of change was estimated by the Theil–Sen estimator (Sen, 1968; Gilbert, 1987; Equation 6),

the median of all pairwise slopes and thus insensitive to outliers; pairing the Mann–Kendall test with Sen’s slope is standard practice in trend studies (Gocic and Trajkovic, 2013). Slopes are given per year.

Change-point tests

The Pettitt test (Pettitt, 1979) locates a single abrupt shift in a series and tests its significance. In terms of the ranks r of the observations, its statistic, given by Equation 7,

attains K = max|U t | at the change-point, whose significance follows from K. It is paired with the sequential Mann–Kendall procedure, whose forward and backward rank statistics cross near the onset of a trend and so estimate the timing graphically; such tests are now routine for probing the stationarity of hydrological and water-quality records (Villarini et al., 2009).

Innovative trend analysis

ITA (Şen, 2012, Şen, 2017) divides a series into two equal chronological halves, sorts each in ascending order, and plots the second half against the first. Points above the 1:1 line indicate an increasing tendency and points below a decreasing one; the method thus reveals whether low, middle, and high values behave alike, which a single rank test cannot. An ITA slope was computed from the difference in half-series means. We use ITA descriptively, to corroborate the rank tests rather than as an independent significance test; the method has been applied widely to precipitation and water-quality records (Şen, 2014; Wu and Qian, 2017).

Field significance (Monte Carlo)

Testing 11 stations for each of the six measured variables and the derived RPI amounts to 77 hypothesis tests, so at α = 0.05 roughly 4 “significant” stations would appear by chance alone, and the stations co-vary in space. To ask whether the observed count of significant stations for a variable is more than chance would give, we used a synchronized-permutation Monte Carlo procedure: a single random re-ordering of the monthly index was applied to every station—preserving the correlation among stations while destroying any trend—the MK test was recomputed at each station, and the number with p < 0.05 was tallied; 2,000 repetitions yield the null distribution of that count, and the field-significance p-value is the share of permutations whose count reaches or exceeds the observed value. A 95% confidence interval for each basin-median Theil–Sen slope was obtained by bootstrapping stations (5,000 resamples). Both procedures used a fixed seed and are reproducible.

Robustness, segmentation, and covariates

Three further procedures address the specific fragilities of a long urban record. First, to separate real basin change from the effect of network turnover, every headline trend was recomputed on the 8 stations with complete 2010–2026 records and the significant-station counts were compared with the full network (the robustness analysis). Second, where a change-point test flags an abrupt shift—as it does for the coliform group—reporting a single monotonic Theil–Sen slope through the discontinuity is inappropriate, because a step violates the monotonic-trend and homogeneity assumptions of the MK test; the coliform record is therefore described by its per-station change-point dates and its pre- and post-shift levels rather than by one basin slope. Third, to probe drivers rather than mere timing, the river series were associated with the ancillary data by Spearman correlation: monthly basin coliform against monthly rainfall, and annual river NH 3 -N, BOD 5, and RPI against the Bali plant’s annual effluent volume. These correlations are read as suggestive associations, not proof of cause, given the modest annual sample sizes.

Analyses were run in Python 3.11 with the pymannkendall (Hussain and Mahmud, 2019), pandas, numpy, scipy, and statsmodels libraries; the change-point, ITA, segmentation, and covariate routines were implemented directly. Equation 2–7 were checked against the primary sources cited beside them, and the rank-test, slope, and change-point routines will be made available by the authors, without undue reservation, so that the implementation itself can be inspected. Every reference cited here was checked against its publisher or DOI record.

Water-quality status and distributions

Table 1 sets out the six measured variables and the derived RPI for the early (2010–2012) and late (2023–2025) sub-periods—3 complete years at each end, the incomplete 2026 excluded—and for the full record, and Supplementary Figure S1 shows their annual distributions. Across 2010–2025 the basin-median values were 4.4 mg L - ¹ for DO, 2.8 mg L - ¹ for BOD 5, 12.9 mg L - ¹ for COD, 15.3 mg L - ¹ for SS, 1.39 mg L - ¹ for NH 3 -N, 7.7 × 10 4 colony-forming units (CFU) per 100 mL for the coliform group, and 3.5 for RPI. From the early to the late sub-period the medians fell for BOD 5 (2.9 to 2.4 mg L - ¹), SS (19.8 to 15.2 mg L - ¹), NH 3 -N (1.71 to 0.95 mg L - ¹), and RPI (3.75 to 3.25), while the coliform group rose (the microbial counter-trend). Dissolved oxygen did not improve between the sub-periods (4.5 to 4.2 mg L - ¹) and traced a U within the record, sinking to an annual median near 3.2 mg L - ¹ around 2019–2020 before recovering (Supplementary Figure S1). Excluding the partial 2026 record removes an artificial dry-season upswing at the end of the series; a late-period statistic that included it would be biased toward the cooler, better-oxygenated months and would overstate the improvement.

Monotonic trends

Mann–Kendall statistics for every station and variable appear in Supplementary Table S2, with the corresponding Theil–Sen slope magnitudes in the companion Supplementary Table S3, and are mapped in Figure 2; the basin-wide tally is in Supplementary Table S4. The dominant signals are declines in suspended solids, ammonia, and BOD 5. Suspended solids fell significantly (p < 0.05) at 9 of 11 stations, NH 3 -N at 8, and BOD 5 at 6, and none of the three rose significantly at more than 1 station. The River Pollution Index fell significantly at 10 of 11 stations and rose at none, but since the RPI is a calculated index, the mean of the DO, BOD 5, SS, and NH 3 -N sub-indices rather than a measured quantity, its decline integrates the trends in those components rather than corroborating them, and we treat it as a summary. Dissolved oxygen rose significantly at only 2 stations, both on the Jingmei Creek, with weak and non-significant positive trends elsewhere; whether that 2-station count is more than chance is taken up in the field-significance test. COD gave a mixed picture, 3 stations rising and 2 falling. The seasonal Mann–Kendall test, which removes the monthly cycle, returned concordant results—RPI decreasing at 9 stations, SS and NH 3 -N at 9 and 8 (Supplementary Table S4)—so the trends are not an artifact of seasonality. The modified Mann–Kendall test preserved the sign of every trend and, at the strongly trending Keelung stations, the significance of the RPI, SS, and NH 3 -N results, so serial correlation does not account for them.

Improvement was concentrated in space. All 6 Keelung River stations show strongly significant declines in RPI, SS, and NH 3 -N; the steepest are NH 3 -N at Bailing Bridge (Z = −4.84, p < 0.001) and RPI at Zhongshan Bridge (Z = −4.20, p < 0.001). The cleaner Jingmei Creek supplied the only significant increases in DO, and the 2 stations retired in 2022 (XD1, JM2) contributed shorter records.

Trend magnitudes and Innovative Trend Analysis

The Theil–Sen slopes (Supplementary Table S3) put rates to these trends. The basin-median slopes were +0.040 mg L - ¹ yr - ¹ for DO, −0.054 for BOD 5, −0.013 for COD, −0.585 for SS, −0.081 for NH 3 -N, and −0.080 index units yr - ¹ for RPI. Held over the 17-year window, the RPI slope amounts to a basin-median gain of about 1.4 index units and the SS slope to roughly 10 mg L - ¹—the largest absolute change of any variable, and one concentrated on the Keelung River, where station slopes run from −0.4 to −0.7 mg L - ¹ yr - ¹.

Innovative Trend Analysis of the Keelung River stations (Supplementary Figure S2) confirmed the rank tests and added detail. The NH 3 -N and RPI points lie well below the 1:1 line across the whole range of values, marking a general decline, while the DO points sit close to the line, as its weak monotonic signal would predict. For NH 3 -N the gap below the line widens toward the high values, so the most ammonia-rich samples fell proportionally more than the cleaner ones—the pattern expected when concentrated wastewater is removed rather than merely diluted.

Change-point analysis

Timing was consistent for the oxygen-demand and nutrient variables. The Pettitt test placed a change-point at 2021 for DO, BOD 5, NH 3 -N, and RPI (Supplementary Table S4), and suspended solids changed earlier, in 2015. For COD and, tellingly, the coliform group, the per-station change-points scatter across the record (2013–2024) rather than cluster, so no single basin-wide onset can be assigned to either; the coliform behavior is taken up in the discussion of microbial persistence. Figure 3a shows the Keelung River annual-mean RPI falling from about 3.8 in the early sub-period to about 3.0 in 2023–2025, approaching the boundary between the moderately and lightly polluted bands (the incomplete 2026 point is plotted but excluded from this mean). The sequential Mann–Kendall statistic for NH 3 -N (Figure 3b) tells the same story: the progressive series turns down and crosses the −1.96 bound in the early 2020s. The DO record at Chengmei Bridge (Figure 3c), smoothed using locally weighted scatterplot smoothing (LOWESS), traces the mid-series dip and the recovery that follows 2020.

Spatial pattern of the River Pollution Index

Supplementary Figure S3 maps the RPI in space and time. The station-by-year panel greens after 2021—that is, the RPI falls—most plainly along the Keelung River. The rate panel confirms that 10 of 11 stations improved, only the Xindian Creek station (XD1, retired in 2022) lacking a significant RPI trend. The severely polluted share of valid basin samples fell from 11.1% in 2010–2012 to 2.5% in 2023–2025, and the unpolluted share rose from 20.4% to 25.9%; the basin-mean RPI fell from 3.83 to 3.20 between those windows. The station added downstream in 2022 (TS2, Zhongxiao Bridge) shows a steep RPI decline over its short record (−0.57 yr - ¹), but that rate rests on 49 samples across 4 years, so it is excluded from the basin trend and the improvement reported here in no way depends on it (the robustness analysis).

A microbial counter-trend

One variable ran against the tide. The coliform group rose significantly at a majority of stations—7 of 11 in the full network, and 5 of 8 complete-record stations—among them 5 of 6 Keelung stations, from Nanhu Bridge down to Zhongshan Bridge (all but the lowest, Bailing Bridge), the Xindian station at Huajiang Bridge, and Chongyang Bridge on the lower main stem, and fell significantly at none (Supplementary Table S2), a verdict the seasonal test repeats. The rise is real but gradual, and its shape matters. The per-station change-points scatter across 2013–2024 (2013–2022 for the 7 significantly increasing stations; Figure 4b) rather than clustering, so—unlike the coherent 2020–2021 chemical shift—the coliform signal has no single basin-wide onset; a single monotonic Theil–Sen slope fitted through the whole record (which would read about +3,000 CFU (100 mL) - ¹ yr - ¹) forces a straight line through a stepwise, heterogeneous series and violates the monotonic-trend assumption, so we do not use it to describe the trend. Described by level rather than by one slope, the basin geometric-mean coliform concentration rose from about 7.9 × 10 4 CFU (100 mL) - ¹ over 2010–2022 to about 1.6 × 10 5 CFU (100 mL) - ¹ in 2023–2024 (Figure 4a). The reaches that shed their oxygen-demanding, solid, and nitrogenous load thus held, or increased, their fecal-indicator load; the discussion of microbial persistence examines why.

Field significance (Monte Carlo)

Because 77 station–variable tests were run (66 on the six measured variables and 11 on the derived RPI), some individually significant results will be accidents of chance. The synchronized-permutation Monte Carlo test (Table 2) shows that under the null of no trend only about 0.5 stations per variable reach p < 0.05, with a 95th percentile of 2 to 3. Measured against that null, the declines in RPI (10 stations), SS (9), NH 3 -N (8), and BOD 5 (6), and the rise in the coliform group (7), are all field-significant (Monte Carlo p ≤ 0.006), and these five survive a conservative Bonferroni correction for the seven series tested (α = 0.05/7 ≈ 0.007). COD is marginal at best—5 stations, p = 0.021—and its significant stations split between increasing and decreasing, so it represents no coherent basin change. Dissolved oxygen is the clear exception: its 2 significant stations fall within the null expectation (Monte Carlo p = 0.11), so DO shows no field-significant basin-wide trend over 2010–2026.

The bootstrap confidence intervals for the basin-median slope excluded zero for SS, NH 3 -N, RPI, and the coliform group, and two apparent tensions repay a moment’s attention. The interval for BOD 5 just included zero even though BOD 5 is field-significant, so its basin-median rate is loosely constrained. Conversely the DO interval narrowly excluded zero—a small but consistent positive median slope of +0.040 mg L - ¹ yr - ¹—although DO is not field-significant, because DO edges upward at many stations yet strongly enough at none to lift the count of significant tests above chance. We therefore read DO as a faint, uncertain drift rather than an established trend, and where the field-significance test and the slope interval disagree we follow the former.

Robustness to station turnover

Because the network changed in 2022, the basin trends were recomputed on the 8 stations with unbroken 2010–2026 records (Supplementary Table S5). The picture does not change. Ammonia declines significantly at all 8 of these stations, suspended solids and RPI at 7 of the 8, and BOD 5 at 4—fractions that echo the full-network counts of 8, 9, 10, and 6 of 11. The coliform group still rises at 5 of the 8. The retired Xindian and Jingmei stations and the short-record TS2 therefore neither create nor inflate the basin signal; the improvement is carried by the stations measured throughout, and the bootstrap intervals from the fixed set overlap those from the full network. No conclusion in this paper depends on the stations that came or went.

Ancillary evidence on drivers

Three external datasets bear on interpretation. First, water temperature at the 8 fixed stations rose significantly over 2010–2025 (Theil–Sen +0.15 °C yr - ¹, p = 0.015; Supplementary Figure S4a). Because oxygen solubility falls as water warms, this warming works against dissolved oxygen; that DO nonetheless held level (p = 0.65; Supplementary Figure S4b) means the oxygen regime resisted a warming headwind rather than merely stagnating. Second, basin coliform showed no association with monthly rainfall (Spearman ρ = −0.01, p = 0.89; Supplementary Table S6), which argues against diffuse, runoff-driven contamination and points toward point sources. Third, and consistent with that, the influent–effluent records of the basin’s treatment plants (Supplementary Table S7, Figure 5) show that the Bali plant—treating some 1.3 million m³ d - ¹, the basin’s dominant discharge—operates at primary level and, after 2024, saw its effluent-coliform log-removal collapse from about 4.0 to about 1.5, so its effluent geometric-mean coliform rose roughly 200-fold to 10 5 –10 6 CFU (100 mL) - ¹; the Dihua reclamation plant, which disinfects, held its effluent near 5 × 10² CFU (100 mL) - ¹ over the same interval. River NH 3 -N, BOD 5, and RPI fell as annual treated volume rose (ρ ≈ −0.44 to −0.58), a suggestive but not statistically decisive association (Supplementary Table S6). These are correlational readings, offered as mechanism-consistent evidence, not proof of cause.

A recovery that held

The central finding is that the improvement reported for the Danshui basin through 2010 (Cheng et al., 2011) carried on over 2010–2026. Suspended solids, ammonia, and BOD 5 fell across the majority of the network, and the steepest declines were on the Keelung River, historically the basin’s dirtiest tributary. The ordinary, seasonal, and modified Mann–Kendall tests, Innovative Trend Analysis, and the Monte Carlo field-significance test agree—tests that make different assumptions but draw on the same data—so the trends are robust to the choice of test; this is agreement of method, not independence of evidence, and it settles nothing about cause. The steep, widening fall of the highest NH 3 -N values in the ITA plots is what one would expect if concentrated wastewater were removed rather than diluted. The interval also spans the later stages of the basin’s sewerage-expansion and household-connection program and the continued operation of its treatment wetlands (Cheng et al., 2011; Vymazal, 2011; Wu et al., 2010); but the record is observational, and we report the correspondence as a coincidence in time rather than attribute the trends to any one measure.

Dissolved oxygen alone did not improve in a way the field test can defend (the field-significance test), and the apparent tension between a flat DO and a falling RPI dissolves once index construction and temperature are taken into account. The RPI is the arithmetic mean of four sub-indices; its decline here is carried by the pronounced drops in BOD 5, SS, and NH 3 -N, and a mean of four terms falls whenever three fall even if the fourth holds. A stationary DO sub-index alongside three declining ones therefore lowers the RPI rather than contradicting it—the two are arithmetically compatible, not at odds. Temperature sharpens the point. Water temperature rose significantly across the record (Theil–Sen +0.15 °C yr - ¹, p = 0.015; the analysis of drivers), and because oxygen solubility falls as water warms, a DO concentration that merely held level represents oxygen conditions that gained a little ground against a falling saturation ceiling; the percentage-of-saturation series is flat to slightly declining rather than improving. Dissolved oxygen in this warm, tidally influenced lowland river answers to temperature, respiration, and tidal exchange as much as to organic loading, so the removal of BOD 5 need not translate into a proportional oxygen gain. The U-shaped path, with a trough around 2019–2020 and a recovery that coincides with the change-point in the other variables, is consistent with this reading. The drivers behind each variable’s trajectory, DO included, are weighed in the discussion of drivers below.

A shared change around 2020–2021

That several variables turned at once is itself informative. For BOD 5, NH 3 -N, and RPI the Pettitt test placed the change-point at 2021, and the sequential Mann–Kendall statistic marked a similar onset. A change shared across chemically distinct variables is easier to reconcile with a common, basin-scale influence than with noise peculiar to each. One candidate is the crossing of a sewer-connection threshold that would remove a large increment of untreated load in a short span; the reduced commercial and travel activity of 2020–2022 is another. Observational data can fix the timing of the change but not its cause, and to choose among these accounts would require the sewer-completion and discharge records that lie outside this dataset. That the low RPI persisted through 2025–2026 does at least suggest a structural shift rather than a passing one.

Drivers of the variable-specific trends

The convergent design fixes the direction, size, and timing of each trend. It does not say why. The evidence already in hand (the timing of the change-points, the spatial pattern of the significant stations, the shape of the ITA plots, and the ancillary records) does allow the candidate drivers to be ranked variable by variable, and that ranking is the business of this section. Three cautions apply. The record is observational, the ancillary series are short, and several drivers operated at once; what follows orders plausible mechanisms and stops short of attribution.

Ammonia and BOD 5

Domestic sewage dominates both loads in this basin. A source apportionment for the Danshui showed ammonium to be the main form of dissolved inorganic nitrogen in the urban reaches, and showed that most ammonium enters not through the treatment plants but through the rain-water pumping stations that lift the city’s storm-drain flow, and the sewage mixed into it, over the levees (Kuo et al., 2017). Household connection acts on exactly that pathway. Each household joined to the separate sewer is one whose black and gray water no longer reaches a storm drain. By 2024 Taipei City had brought public-sewer connection to 88% of its population, and New Taipei City, which holds the upstream Keelung valley, to 72%, with 1.24 million New Taipei households connected (Ministry of the Interior sewerage statistics; the data sources). Three features of the record fit source removal, and fit it better than dilution or in-stream processing. The ITA plots show the largest fall in the most ammonia-rich samples, which is the mark of a concentrated source removed rather than of a load spread thinner (the trend magnitudes and ITA analysis). The 2021 change-point is shared by BOD 5, NH 3 -N, and RPI, as a threshold in connection would produce; dilution would have moved SS and COD as well. River NH 3 -N and BOD 5 fell as the treated volume at the Bali plant rose (ρ ≈ −0.44 to −0.58; Supplementary Table S6), that volume standing in for the sewage diverted from the river. In-stream nitrification adds little. Ammonium dominates the dissolved inorganic nitrogen of the lower Danshui, which Lee et al. (2014) attributed to low oxygen and short residence time, so ammonia leaves this river by transport rather than by oxidation; the decline records less ammonia entering, not more ammonia consumed. The riparian wetlands remove ammonia efficiently within their own footprint (Vymazal, 2007). Their treated volume is a small fraction of the basin’s sewage flow, however, and their contribution to the station record is second-order beside interception.

COD

COD did not follow BOD 5, and the divergence is itself a finding. Between the sub-periods the basin-median BOD 5 fell from 2.9 to 2.4 mg L - ¹ while the median COD rose from 12.0 to 13.0 mg L - ¹ (Table 1), so the ratio of the two medians fell from about 0.24 to 0.18. That ratio indexes biodegradability. Untreated domestic sewage typically runs between 0.3 and 0.8, whereas treated effluent and urban runoff carry organic matter that is largely refractory (Tchobanoglous et al., 2014). Interception therefore removes the biodegradable fraction first and leaves the slowly degradable residue behind: treated effluent, road and roof runoff, and industrial and commercial discharges. The significant COD stations split between rising (Chengmei, Wanshou, and Huajiang Bridges) and falling (the two main-stem stations), and the change-points scatter (Supplementary Table S4), which is how a variable fed by several sources on different schedules would behave. COD is the one variable for which the study cannot name a dominant driver. It is also the one for which source-resolved data, on industrial discharges and stormwater quality above all, would repay the effort of collection.

Suspended solids

Solids answer to sediment supply and hydraulics as much as to pollution, and their record differs from the dissolved variables in two ways: the steepest declines fall on the Keelung River, and the change-point (2015) precedes the 2021 chemical shift by 6 years (Supplementary Table S4). Two drivers stand out. The first is sediment supply. The Yuanshanzi flood-diversion tunnel, commissioned in 2005 in the upper Keelung catchment, diverts up to 1,310 m³ s - ¹ of flood flow directly to the sea; at the Wudu gauge below it the annual sediment load fell from about 0.59 to 0.17 Mt yr - ¹, and the typhoon sediment pulses that once dominated the load were cut off (Lee et al., 2024). A structure commissioned 5 years before the record begins cannot by itself produce a decline within it. It does two other things. It removes the typhoon-driven pulses that governed the interannual SS variance in the earlier study, and it explains why the Keelung could fall while the Dahan-weighted network of Cheng et al. (2011), on a tributary with no such diversion, rose. The second driver is the removal of raw sewage, which carries its own suspended solids, together with settling in the treatment wetlands and the low-gradient tidal reaches. Interannual scatter remains large, because a single typhoon sampled by chance in the monthly cycle can add tens of milligrams per liter to a year’s mean (the limitations), and no single event accounts for the 2015 change-point. The absence of typhoon landfalls on Taiwan between August 2019 and September 2023 would have held solids down in those years without any change in management; the SS record partly tracks the storm climate, and a trend test cannot tell the two apart.

Dissolved oxygen

Four influences set the DO record, and their signs differ. Warming lowers the ceiling. For water between 20 and 25 °C, the 2.25 °C rise implied by the Theil–Sen rate over 2010–2025 lowers the freshwater saturation concentration by about 0.35 mg L - ¹, or 4% (Benson and Krause, 1984), which is of the same order as the change in the DO median between the sub-periods (4.5 to 4.2 mg L - ¹; Table 1). Load reduction raises the floor, but by less than the ammonia figures suggest. The 0.76 mg L - ¹ fall in median NH 3 -N corresponds, at about 4.6 g of oxygen per gram of ammonia nitrogen, to a potential nitrogenous oxygen demand of some 3.5 mg L - ¹ (Tchobanoglous et al., 2014), seven times the 0.5 mg L - ¹ fall in BOD 5. That demand was largely never exerted in the river, as the persistence of ammonium in the lower reaches shows (Lee et al., 2014), so removing it returned little oxygen. Low flow deepens the trough. The tidal reaches of the Danshui system turn hypoxic in low-flow periods, when dilution and reaeration both weaken (Liu et al., 2005), and the DO minimum of 2019–2020 falls within an unusually dry spell: no typhoon made landfall on Taiwan between August 2019 and September 2023, and the failure of the 2020 typhoon season together with a weak 2020 Mei-yu produced, from mid-2020 to spring 2021, the most severe drought Taiwan had seen in 56 years (Huang et al., 2025). The recovery that follows coincides with the return of wetter years and with the 2021 change-point in ammonia and BOD 5; monthly data cannot separate the two. Sediment oxygen demand from the organic deposits that decades of raw discharge left on the bed, and tidal exchange with the estuary, probably buffer these reaches against any rapid response. The net result is a DO series that resists rather than rises, which is what the field-significance test reports.

The coliform group

The next section treats the microbial counter-trend at length; only the mechanism belongs here. The treatment train that removes organic matter, solids, and ammonia, primary treatment at Bali and constructed wetlands within the levees, removes fecal bacteria only where a disinfection step follows it, and at Bali that step failed after 2024 (Supplementary Table S7). Reduced freshwater discharge raises coliform concentrations in the Danshui estuary by weakening dilution (Liu and Chan, 2015), while rainfall mobilizes overflow and runoff loads; the two effects of a wet month run in opposite directions, and their near cancellation is consistent with the null rainfall correlation (ρ = −0.01). The growing, incompletely sewered population of the upstream Keelung valley (Supplementary Note S1) supplies an expanding source exactly where the rise concentrates. One consequence for the index deserves to be stated plainly. The RPI is built from DO, BOD 5, SS, and NH 3 -N and carries no microbial term, so the official measure records the basin as improving while its fecal risk rises; the index measures what the first-generation program was designed to fix, and nothing else.

Why the Keelung River improved most

The spatial pattern follows from the drivers above. The Keelung reaches began the period as the basin’s dirtiest, so they had the most to gain from interception; all six Keelung stations lie within Taipei City, where connection is highest, even though their upstream catchment in Xizhi lags; and they alone benefit from the upstream flood diversion that cut their sediment supply. The Jingmei Creek, cleaner at the outset, had less ammonia and BOD 5 to lose and returned the only significant DO gains. The two stations retired in 2022 (XD1 on the Xindian Creek and JM2 on the Jingmei Creek) contribute shorter records and, at Huajiang Bridge, a rising COD and coliform signal from the lowest reach of the Xindian that the network no longer follows.

Chemical recovery, microbial persistence

The coliform counter-trend is the study’s chief caveat, and because it opposes every other variable it deserves a closer, data-led reading. Three features of the record shape the interpretation. First, the rise is real in the rank tests—significant at 7 of 11 stations, and at 5 of 8 complete-record stations (the microbial counter-trend and the robustness analysis)—so it is not an artifact of network turnover. Second, and against any single-cause reading, the per-station change-points scatter across 2013–2024 rather than cluster (Figure 4b): several stations begin rising in the mid-2010s, well before any recent event, so the coliform series has no single basin-wide onset of the kind seen in the coherent 2020–2021 chemical shift. Third, on top of that gradual climb the basin geometric mean is elevated in 2023–2024 (about 1.6 × 10 5 against 7.9 × 10 4 CFU (100 mL) - ¹ over 2010–2022), a recent bump whose partial coincidence with administrative change in the national laboratory system (below) means an analytical contribution to those two years cannot be excluded—though it plainly cannot explain the station-by-station rises that began years earlier.

The recent 2023–2024 elevation does warrant caution about method. Taiwan’s environmental-analysis system was reorganized in exactly this window: the Environmental Protection Administration was restructured into the Ministry of Environment in August 2023, and its Environmental Analysis Laboratory was folded into the new National Environmental Research Academy, while two coliform determinations—the multiple-tube fermentation method (NIEA W505) and the enzyme-substrate method (NIEA E215)—remained concurrently valid rather than one superseding the other. A change of analyzing laboratory under annual contracting, or a switch between these two methods, can move reported counts without any change in the river, and the public record does not let us exclude such a shift for 2023–2024. We therefore treat the magnitude of that recent bump with reserve and recommend it be reconciled against the Ministry’s method and contracting records before it is read as wholly environmental. This reservation is bounded, however: it cannot account for the coliform trend as a whole, because most stations’ change-points precede 2023 by several years (Figure 4b). The bulk of the rise is not a 2023 analytical artifact.

The timing tells equally against a pandemic explanation. Taiwan imposed its only nationwide movement restriction—a Level-3 alert, with partial work-from-home—for some 10 weeks, from mid-May to late-July 2021; yet many stations were already rising before 2020, and the 2020–2021 window itself shows only a modest coliform rise. A 10-week shift of domestic wastewater from workplaces to homes is an unlikely author of a multi-year, staggered increase. Wastewater-based surveillance did capture the pandemic’s community signal in sewage elsewhere (Peccia et al., 2020), but that signal is viral RNA rather than the coliform group, and its timing does not match the record here. The demographic record tells against a pandemic pulse as well, though not in the way a glance at Taipei City alone would imply. The city the river passes downstream lost population over the study period (falling below 2.5 million in 2024), but the Keelung River’s upstream catchment grew: Xizhi, the New Taipei district in the Keelung valley just above the Taipei reaches, rose from about 188,000 residents in 2010 to about 212,000 in 2026—a 13% gain, its registered households climbing by roughly 31%—even as sewer connection there lagged the pace of development (New Taipei City and Ministry of the Interior household-registration statistics; Supplementary Note S1, Supplementary Figure S5). Incompletely collected domestic sewage from this growing upstream population is therefore a demographic source situated exactly where the coliform increase concentrates, consistent with the point-source, rainfall-independent reading developed below rather than a basin-wide pandemic effect.

Beyond any measurement artifact, a real upward pressure on fecal indicators is plausible on mechanism and fits the chemical results. Constructed wetlands and conventional secondary treatment cut oxygen demand, solids, and ammonia far more reliably than they remove fecal bacteria. Wetlands take out nutrients and organic matter efficiently, and their design and long-term performance are by now well characterized (Vymazal, 2007; García et al., 2010; Kadlec and Wallace, 2009; Wu et al., 2015); but the reported log-reductions of fecal indicators are modest and highly variable unless a dedicated disinfection step is added (Vymazal, 2011; Wu et al., 2016). As household connection to the sewer network rose over the study period—the same process that lowered in-stream BOD 5 and NH 3 -N—a growing volume of collected sewage was routed to treatment plants and their outfalls; where that effluent is treated for organic and nutrient load but not disinfected, fecal indicators can persist at and below the discharge points even as dissolved pollutants fall. The basin’s own treatment records make this concrete (Supplementary Table S7; Figure 5). The Bali plant, handling about 1.3 million m³ d - ¹ at primary level, saw its effluent-coliform log-removal collapse from roughly 4.0 in 2022–2023 to roughly 1.5 in 2024–2025—its effluent geometric mean climbing some 200-fold to 10 5 –10 6 CFU (100 mL) - ¹—so that even while removing about a third of incoming BOD 5 it returned a heavy fecal load to the estuary; because influent counts were roughly stable, the collapse in log-removal reflects a real loss of disinfection performance, not a change of scale in measurement. The contrast with the Dihua reclamation plant, which disinfects and held its effluent near 5 × 10² CFU (100 mL) - ¹, shows the missing ingredient is disinfection, not treatment as such. This fits the general parting of the coliform group from BOD 5 across the network. Basin coliform, moreover, showed no association with monthly rainfall (ρ = −0.01; the analysis of drivers), which argues against diffuse, runoff-driven contamination as the principal author and points to continuous point-source discharge; combined-sewer overflows and illicit connections may still add episodic loads during the basin’s frequent convective and typhoon rain (Tornevi et al., 2014), and a warming climate is expected to lift waterborne-pathogen concentrations further (Hofstra, 2011), but the dominant signal here is a persistent, under-disinfected effluent load. The practical lesson is unchanged and, if anything, sharper: having largely won against organic, solid, and nutrient pollution, the next phase of Danshui management must both verify the microbial record and close the fecal pathways—most obviously by adding disinfection to the major discharges—that treatment has so far left open.

A natural alternative to the effluent explanation is wildlife, since the Danshui wetlands hold nationally important waterbird assemblages and avian feces are a recognized diffuse source of fecal indicators. We tested this against the Tamsui River Bird Count, a basin-wide synchronized census running since 2016 (Fu et al., 2021; Lin and Chiu, 2024a, b), and it does not survive the test (Supplementary Note S2; Supplementary Figure S6). Although the total winter count roughly doubled over 2017–2024, the increase is carried by landbirds—mynas, bulbuls, doves, and starlings—that do not deposit into the water column; among the species with a significant winter trend, only one waterbird (the black-crowned night heron) increased, whereas three water-associated species declined (common greenshank, the actively culled sacred ibis, and white wagtail) and the principal waterfowl and waders (spot-billed duck, common moorhen, and Kentish plover) merely fluctuated. Spatially, the census wetlands lie in the estuary and along the Dahan tributary rather than on the Keelung River, where the coliform rise concentrates, and a Keelung River transect was added only in 2023. A deliberately generous order-of-magnitude estimate puts the entire standing flock at roughly 10¹³ CFU day - ¹ of fecal coliform—two to three orders of magnitude below the Bali plant’s ~10 15 –10 16 CFU day - ¹. Waterbirds are therefore best read as a stable, spatially localized background source that lifts the baseline but cannot generate the observed rise, a reading consistent with the absence of any rainfall association and with the point-source attribution above.

Methodological note

The agreement among methods is a result in its own right. The ordinary and seasonal Mann–Kendall statistics agreed in sign for every station–variable pair, and the modified test rarely changed a verdict, so seasonal structure and serial correlation, real as they are, did not manufacture the trends. Innovative Trend Analysis added what the rank tests omit, the separate behavior of high and low values, and the change-point tests supplied the timing that a monotonic test leaves out. Where records were short—the station added in 2022—or discontinued, we confined the testing accordingly and declined to project steep annual rates from 4 years of data. To combine complementary non-parametric tools, rather than lean on one, is a design we would recommend for comparable urban-river work (Şen, 2017; Yue et al., 2002a).

Limitations

Four limitations qualify the findings. The mid-record revision of the network—two stations retired and one added in 2022—means it is not perfectly stationary; we met this by testing each station’s own continuous series and flagging the short records. Monthly sampling can alias short-lived events such as the sediment pulse of a typhoon, which contributes to the year-to-year scatter in the SS record. And the size of the coliform trend is sensitive both to a possible method change and to the extreme skew of bacterial counts, so we report it as a direction and a flag for inquiry rather than a precise rate. Finally, the 2026 data cover only January to May, so any 2026 annual figure is seasonally biased and is kept out of the formal inference; the rank-based tests run on the full monthly series and are untouched by the partial final year.

Comparison with the 1993–2010 assessment

This study was built as a direct temporal extension of the trend analysis of the same basin by Cheng et al. (2011), and the two together span more than three decades (Table 3). They differ in three ways that shape the comparison. In coverage, the earlier study weighted its 13 stations toward the Dahan Creek and the downstream Danshui, whereas the present network leans toward the Keelung River and adds the Jingmei Creek, so the monitored reaches overlap only in part and cannot be matched station for station. We therefore read the two studies as basin-scale narratives bracketing three decades, not as a station-matched before-and-after; every variable-by-variable comparison in Table 3 is qualitative, and a reversal of sign between the studies may reflect this change of spatial emphasis as much as any change in the basin itself. In method, the earlier study used the Mann–Kendall and seasonal Mann–Kendall tests with Sen’s slope, while the present one adds the autocorrelation-corrected modified test, the Pettitt and sequential-MK change-point tests, Innovative Trend Analysis, and a Monte Carlo field-significance test, and introduces the composite RPI and COD. Those additions are what let the present study date the improvement to 2020–2021 and tell a real basin-wide trend from a multiple-comparison artifact—distinctions the earlier design could not draw. Water-quality trends have elsewhere been tied to particular interventions and pressures—deep-tunnel operation (Ab Razak and Christensen, 2001), agroforestry and paired-watershed management (Udawatta et al., 2002), and basin-scale hydrological change (Abdul-Aziz and Burn, 2006)—yet attributing a trend to a cause demands controls that observational monitoring seldom supplies.

Against that backdrop, three of the five shared variables agree and two reverse. Ammonia and BOD 5 fell in both periods, confirming a sustained, two-decade retreat of nutrient and organic pollution. Dissolved oxygen, which rose at 84.6% of stations through 2010 as the wetlands and sewers came online, shows no field-significant trend over 2010–2026: the earlier oxygen recovery appears to have largely run its course, leaving DO to fluctuate about a higher but steady level. The two reversals matter more. Suspended solids rose at 53.8% of stations in the earlier study—an increase its authors laid to typhoon-driven sediment from the Dahan Creek—but fall significantly at 9 of 11 stations in the present Keelung-weighted network, so the SS reversal is at least in part a change of spatial emphasis rather than of basin behavior. The fecal reversal is the most consequential: E. coli fell at 69.2% of stations through 2010, in step with the spread of sewerage, yet the coliform group rises at 7 of 11 stations over 2010–2026. The microbial gains of the sewer-connection era have stalled or turned, for the mix of reasons set out in the discussion of microbial persistence.

The implication is a change in the basin’s agenda. The first-generation strategy examined by Cheng et al. (2011)—riparian wetlands and household sewer connection—succeeded against the oxygen-demand, nutrient, and, in the reaches now watched, solids burdens, and the present record shows that success has held. The frontier has moved to the fecal-microbial dimension, which conventional organic-and-nutrient treatment does not resolve, and to the integrity of the microbial record itself. The comparison also makes a methodological point: upgrading the analytical toolkit between assessments turned a broad narrative of “improvement” into a dated, multiplicity-controlled account that tells apart the variables still improving, the one that has leveled off, and the one now moving the wrong way.

The Danshui in a global context

The Danshui trajectory has close parallels elsewhere, and the parallels sharpen its lesson. China’s national monitoring network recorded COD falling from 9.3 to 3.4 mg L⁻¹ between 2003 and 2017, with ammonium falling in step, as industrial, rural, and urban residential discharges were cut (Ma et al., 2020); the Danshui’s ammonia decline from 1.71 to 0.95 mg L⁻¹ between the sub-periods is of the same character and rests on the same lever, interception. In the United States, half a century of Clean Water Act investment lowered pollution concentrations substantially, with part of the decline traced to grants for municipal treatment plants, yet more than half of river miles still violate a standard (Keiser and Shapiro, 2019). British rivers tell the same divided story: sanitary quality downstream of towns improved over three decades, whereas sewer overflows, nitrate, and emerging contaminants remain unresolved, so that the official claim that water quality is “better than at any time since the end of the Industrial Revolution” holds only in part (Whelan et al., 2022). The Han River in Seoul (Chang, 2008), the Ebro (Bouza-Deaño et al., 2008), and Eymir Lake (Yenilmez et al., 2011) show, in turn, how much the reported direction of change can depend on the reach and the test chosen. What unites these cases with the Danshui is a sequence rather than a level. Interception and conventional secondary treatment remove the oxygen-demanding and nutrient load first, and the residual problem migrates to constituents the same treatment does not remove.

Global modeling reaches the same conclusion from the other direction. Meeting the Sustainable Development Goal target of halving untreated wastewater by 2030 would cut fecal-coliform loads by 17–43% across world regions, yet the fecal-coliform threshold would still be breached in many rivers, including in Western Europe, Japan, and the eastern United States, where treatment rates are already high, and mostly in low-flow months (Jones et al., 2022). Climate is the second point of convergence. Most rivers lose water quality under droughts, heatwaves, and long-term warming, through the reduced dilution and lower oxygen solubility identified here for the DO record (van Vliet et al., 2023), and the Danshui’s 2019–2020 oxygen trough, which fell within the typhoon-free dry spell that culminated in the 2020–2021 drought, is a small-scale instance of that pattern.

The Danshui adds three things to this literature. It shows the transition from sanitary to microbial limitation happening inside one continuous record, dated to 2020–2021 for the chemical variables and staggered across 2013–2024 for the microbial one. It shows that the official index used to steer policy, having no microbial term, registers the transition as unqualified improvement. And its data density, 1,921 monthly samples at 11 stations together with plant influent and effluent records, allows a degree of attribution that national and global studies working from sparse networks cannot reach. The Danshui is therefore not an exception to the global pattern but an early, well-instrumented example of it.

Over the continuous 2010–2026 record the Danshui River basin showed a broad, spatially consistent improvement in chemical water quality. Suspended solids, ammonia, and BOD 5 fell at a field-significant majority of stations—9, 8, and 6 of 11, with Monte Carlo p ≤ 0.006—and the composite River Pollution Index that summarizes them fell at 10 of 11. Between the early (2010–2012) and late (2023–2025) sub-periods—the incomplete, dry-season-only 2026 excluded—the basin-mean RPI dropped from 3.83 to 3.20, the severely polluted sample fraction from 11.1% to 2.5%, and the Keelung River annual-mean RPI from 3.81 to 3.00; these declines persisted, at the same strength, on the 8 stations with complete records, so they do not depend on the mid-record change of network. Dissolved oxygen, by contrast, showed no field-significant trend and did not improve, holding level even as the river warmed significantly (+0.15 °C yr - ¹) and thus resisting rather than gaining against a falling oxygen-saturation ceiling. The Pettitt and sequential Mann–Kendall tests dated a shared shift to about 2020–2021, and Innovative Trend Analysis showed that the most ammonia-rich samples fell the most. Read variable by variable, the trends answer to identifiable drivers: household sewer connection, which removed the sewage that once reached the river through storm drains and pumping stations, for ammonia and BOD 5; the 2005 upstream flood diversion and the interception of raw-sewage solids for suspended solids on the Keelung River; a warming, low-flow oxygen regime that offset the reduced load for DO; and a treatment train that removes organics and nutrients but, without disinfection, not fecal bacteria for the coliform group. One variable moved the other way: the coliform group rose at a majority of stations, its per-station change-points scattered across 2013–2024 rather than forming a single step, marking fecal contamination as the basin’s outstanding problem. Effluent records for the basin’s largest plant, whose disinfection collapsed after 2024, together with the absence of any rainfall association, point to persistent, under-disinfected point-source discharge rather than diffuse runoff or a pure analytical artifact. Over this interval the declines in oxygen-demanding, solid, and nutrient pollution ran alongside the basin’s continuing remediation; because the data are observational, we report that correspondence as association, and the microbial trend marks the pathway still to be closed. Methodologically, the agreement of the ordinary, seasonal, and modified Mann–Kendall tests with Innovative Trend Analysis, change-point detection, and a Monte Carlo field-significance test shows the worth of a convergent, multiplicity-guarded design for reading trends in the noisy record of an urban river.

For basin management, these results reframe the agenda. The organic-and-nutrient battle that justified the sewer-connection and constructed-wetland program is largely won and is holding, whereas fecal-microbial contamination is now the binding constraint on the river’s recovery. The most direct management lever is disinfection of the major municipal discharges—the point at which the basin’s coliform load is both largest and, on the present evidence, worsening—coupled with routine reconciliation of the bacterial monitoring record against changes in laboratory method and contractor, so that a genuine environmental signal is not confused with an analytical one. More broadly, the Danshui case illustrates a transition that many maturing urban basins can expect: once conventional secondary treatment and interception have removed the oxygen-demanding and nutrient load, the residual, management-relevant risk migrates to pathogens and to the disinfection step that conventional treatment leaves incomplete.