Observando os Rios: Citizen Science Monitoring Water Quality in Brazil

Study area

This study analyzes 187 water bodies, including major rivers and their tributaries, across 50 municipalities in three mesobasins located in the state of São Paulo, all within the Atlantic Forest biome: Alto Tietê, Sorocaba/Médio Tietê, and Piracicaba-Capivari-Jundiaí (PCJ). Collectively, these basins encompass an area of 31,782 km² and provide water to approximately 28 million people. They face significant anthropogenic pressures, including urbanization, industrialization, and intensive agriculture (CBH-AT et al. 2020) (Figure 1).

Figure 1

Ecologically, these mesobasins contain critical remnants of the Atlantic Forest, which are threatened by deforestation and pollution. From a socioeconomic perspective, they are characterized by industrial, urban, and agricultural activities, high population density, and social inequality, all of which exacerbate the impacts on water resources (Ribeiro et al. 2009).

The basins were selected for their significance in water management and their representativeness within the OoR program. Alto Tietê encompasses an area of 5,775 km² and includes 39 municipalities, among which is the São Paulo Metropolitan Region, home to 21 million inhabitants. The region boasts 36% native vegetation and features 14 reservoirs; however, it is also impacted by domestic pollution and water crises (Javeline et al. 2019). Sorocaba / Médio Tietê covers an area of 11,829 km² and comprises 37 municipalities with a population of 2 million inhabitants. The region is affected by industrial and agricultural pollution, as well as deficiencies in sanitation (ANA 2021). The PCJ region encompasses an area of 14,178 km², comprises 71 municipalities, and features a forest cover of 13.5%. It is also home to the Cantareira System and is notable for its participatory governance facilitated by river basin committees (Colombo and Joly 2010).

Data collection

Table 1 presents the list of parameters and their respective scoring criteria used in the construction of the OoR WQI.

OoR WQI data were collected between 2002 and 2023 by trained volunteers at fixed, georeferenced sites, including rivers, streams, lakes, and reservoirs, with varying sampling frequencies. The WQI comprises 14 parameters: 8 analytical (turbidity, total coliforms, DO, BOD, pH, ${\mathrm{\text{NO}}}_3^{–}$, ${\mathrm{\text{PO}}}_4^{\mathrm{\text{3–}}}$, and settleable solids) and 6 organoleptic/biological indicators (floating trash, odor, fish, red larvae, transparent/dark larvae, and foams). Analytical parameters measured with LaMotte kits adhered to the manufacturer’s standardized protocol, as outlined in Supplemental File 1: Table S1.

The additional procedures employed by the program were based on an internally developed methodology, also described in Table S1, and validated in accordance with the criteria established by Rocha et al. (2004). For each sampling, volunteers assigned a score from 1 to 3 to each parameter based on established criteria. A score of 0 was recorded when a parameter was not assessed.

The OoR WQI classification and scale conversion range from 14 to 42 points, while CETESB’s WQI spans 0 to 100. Both indices utilize five qualitative categories, ranging from Excellent to Terrible, but they differ in their value ranges and color codes. For standardization purposes, we adopted the Viridis color palette, which is designed to be accessible for individuals with visual impairments (Figure 2).

Figure 2

Four conversion approaches were implemented to facilitate comparison.

1. Direct Linear Conversion - This method assumes that the OoR WQI can be linearly rescaled from its original range (14 to 42) to the CETESB scale (0 to 100), without considering the methodological differences between the indices.

   ${\text{OoR}}_{\text{CET}}\hspace{0.17em}\text{=}\hspace{0.17em}\frac{\left({\text{IQA}}_{\text{OoR}}\hspace{0.17em}–\hspace{0.17em} 14\right)}{28}\hspace{0.17em}*\hspace{0.17em}100$

   Where:

   • • IQA_OoR is the original index value, which ranges from 14 to 42.

   • • The range of the OoR scale is 28 (42 – 14).

   • • The result is a rescaled value on a scale of 0 to 100.

2. Partial Conversion Using Compatible Parameters – This method utilizes only the eight analytical parameters that are common to both indices (OoR and CETESB). The possible score for these eight parameters in the OoR index ranges from a minimum of 8 to a maximum of 24, assuming each parameter is scored on a scale from 1 to 3.

   ${\mathrm{\text{OoR}}}_{\mathrm{\text{CET}}\_8}\hspace{0.17em}=\hspace{0.17em}\frac{\left({\mathrm{\text{IQA}}}_{\mathrm{\text{OoR}}\_8\mathrm{\text{par}}}\hspace{0.17em}–\hspace{0.17em} 8\right)}{16}\hspace{0.17em}*\hspace{0.17em}100$

   Where:

   • • IQA_OoR_8par represents the total score derived from the eight compatible parameters.

   • • The possible score range is 16 (24 – 8).

   • • The result converts the partial score to a scale of 0 to 100.

3. Weighted Conversion by Parameter – This approach adjusts the scores of the eight compatible parameters, weighting them according to the weights defined in CETESB’s WQI. The weights are proportionally redistributed, as the parameter is not utilized in OoR. Each parameter, denoted as i, is normalized on a scale from 0 (poor) to 1 (excellent).

   ${\mathrm{\text{OoR}}}_{{\mathrm{\text{CET}}}_8\mathrm{\text{*}}}\hspace{0.17em}=\hspace{0.17em}\left(\Sigma \left(\frac{\left({\mathrm{\text{P}}}_{\mathrm{\text{i}}}\hspace{0.17em}–\hspace{0.17em}1\right)}{2}\hspace{0.17em}\times \hspace{0.17em}{\mathrm{\text{w}}}_{\mathrm{\text{i}}}\right)\right)\hspace{0.17em}*\hspace{0.17em}100$

   Where:

   • • P_i represents the score for parameter i in the OoR, with values ranging from 1 to 3.

   • • w_i is the adjusted relative weight for parameter i, based on CETESB’s WQI weights.

   • • The weighted sum is converted to a scale of 0 to 100.

4. Adjustment for Incomplete Sampling – When not all parameters are recorded (e.g., in cases of partial or faulty sampling), this formula adjusts the conversion by considering only the parameters that were effectively measured.

   ${\mathrm{\text{OoR}}}_{{\mathrm{\text{CET}}}_{\mathrm{\text{n}}}}\hspace{0.17em}=\hspace{0.17em}\frac{\left({\mathrm{\text{S}}}_{\mathrm{\text{n}}}\hspace{0.17em}–\hspace{0.17em}\mathrm{\text{n}}\right)}{2n}\hspace{0.17em}*\hspace{0.17em}100$

   Where:

   • • S_n represents the sum of the scores for n available parameters, each of which can take values ranging from 1 to 3.

   • • n represents the number of parameters considered in the sampling process.

   • • The total possible range is represented by the expression 2n, which can be calculated as (3n – n).

This conversion does not apply to the weighted approach (OoR_CET_8*).

These conversions enabled both qualitative and quantitative comparisons. Because OoR and CETESB classifications rely on discrete categories, strict one-to-one matching can overstate disagreements near class boundaries. For this reason, we assessed both exact agreement and agreement within ±1 category. This tolerance, also adopted by Quinlivan et al. (2020), accounts for the inherent uncertainty of interval-based classifications and prevents minor boundary effects from being misinterpreted as substantive discrepancies.

Statistical analyses

We compared OoR converted indices (OoR_CET, OoR_CET_8, OoR_CET_8*) across mesobasins using the Kruskal–Wallis test, followed by pairwise post-hoc comparisons with Benjamini–Hochberg false discovery rate (BH-FDR) correction (α = 0.05). Quantitative agreement between annual means of the converted OoR WQIs and CETESB’s WQI was assessed with Welch’s two-sample t-tests; years with n < 2 per group were excluded. Qualitative agreement among five WQI classes (Excellent–Terrible) was evaluated with Cohen’s Kappa and interpreted using standard benchmarks.

To assess temporal patterns in monitoring effort and data consistency, we applied nonparametric trend tests (Mann–Kendall) and estimated monotonic slopes using the Theil–Sen estimator (α = 0.05). Trends in completeness were modeled with a bias-corrected binomial generalized linear model (GLM), where the response was the probability that a sample contained all 14 parameters; model significance was evaluated by Wald tests. For parameter-specific “success rates” (annual proportion of events with a successful measurement), we likewise tested monotonic trends via Mann–Kendall/Theil–Sen. Unless noted otherwise, two-sided p-values are reported and multiple comparisons were adjusted with BH-FDR. All analyses were conducted in R (v4.3.x), using base stats and the packages rcompanion (post-hoc and BH-FDR utilities), irr Cohen’s Kappa (Sim and Wright, 2005; McHugh, 2012), and Kendall (Mann–Kendall and Theil–Sen).

Data consistency

The quality of OoR data was assessed based on the total number of samples, the success rate of parameters, and completeness, which is defined as the proportion of parameters recorded per sample. In this context, the success rate refers to the annual proportion of monitoring events in which each water quality parameter was successfully measured within a given mesobasin. This indicator reflects the operational completeness of the dataset rather than the analytical accuracy of results.

Socio-environmental analysis

The socio-environmental analysis examined the following aspects:

Comparison of Observando os Rios’s Conversions

Differences Between the OoR_CET, OoR_CET_8, and OoR_CET_8* In the Alto Tietê mesobasin, the Kruskal-Wallis test revealed statistically significant differences among the medians of the OoR_CET, OoR_CET_8, and OoR_CET_8* (H = 245.5; p = 3.454 × 10^–54). Post-hoc comparisons confirmed significant differences between all pairs: OoR_CET vs. OoR_CET_8 (p = 3.347 × 10^–31), OoR_CET versus OoR_CET_8* (p = 9.428 × 10^–7), and OoR_CET_8 versus OoR_CET_8* (p = 9.968 × 10^–47).

Quantitative water quality index (Observando os Rios versus São Paulo State Environmental Agency)

Annual means and standard deviations were calculated for the OoR conversions (OoR_CET, OoR_CET_8, OoR_CET_8*) and CETESB’s WQI on a scale of 0 to 100 (Supplemental File 2: Table S2).

The WQI analysis encompassed the Alto Tietê, Sorocaba/Médio Tietê, and Piracicaba-Capivari-Jundiaí (PCJ) mesobasins from 2003 to 2022, with a total of 2,692, 2,304, and 402 samples collected, respectively. Welch’s t-test was employed to assess statistical differences between the converted indices and CETESB’s WQI (p < 0.05).

The weighted conversion (OoR_CET_8*) demonstrated the highest similarity to CETESB’s WQI, with non-significant differences (p < 0.05) observed in several years—particularly in the PCJ basin (2010, 2012, 2013, 2016–2020), Sorocaba/Médio Tietê (2004–2006, 2008, 2010–2014, 2018–2022), and Alto Tietê (2004–2008, 2013, 2022). This convergence reflects the weight adjustments made in OoR_CET_8*, which align it more closely with CETESB’s methodology. Years with low sampling, such as 2015 and 2022 in the PCJ basin (n = 1), limit the robustness of the findings, whereas larger sample sizes (e.g., 2007: 449 in Alto Tietê and Sorocaba/Médio Tietê) enhance reliability.

Qualitative water quality index (Observando os Rios versus São Paulo State Environmental Agency)

The WQI classifications from OoR (original and converted) and CETESB, categorized from Excellent to Terrible across the mesobasins, are presented (Figure 3).

Figure 3

The comparative analysis between the Observando os Rios (OoR) classifications and the official CETESB WQI revealed substantial regional variation in agreement rates across the analyzed meso-basins (Figure 4).

Figure 4

In the Alto Tietê meso-basin, the original OoR classification achieved the highest level of concordance, with 90% exact agreement relative to CETESB, outperforming all conversion-based approaches (OoR_CET: 60%; OoR_CET_8: 40%; OoR_CET_8*: 56%). When a tolerance of ±1 category was considered, all methods reached 100% agreement, indicating strong qualitative consistency between citizen and official assessments.

In Sorocaba/Médio Tietê, the original OoR classification showed only 10% exact agreement with CETESB, representing the lowest convergence among the three basins. The conversion-based indices improved performance substantially, with OoR_CET and OoR_CET_8* reaching 70% exact agreement and OoR_CET_8 reaching 55%. With the ±1-category tolerance, all methods—including the original index—achieved at least 95% agreement, indicating that discrepancies were generally limited to a single qualitative class.

In the Piracicaba–Capivari–Jundiaí (PCJ) meso-basin, the original OoR classification recorded 65% exact agreement, whereas the conversions produced lower concordance (OoR_CET: 30%; OoR_CET_8: 40%; OoR_CET_8*: 50%). Even in this more heterogeneous system, the ±1 category agreement exceeded 95%, reflecting that deviations were generally within a single classification step.

Overall, these results indicate that while exact numerical equivalence between indices can vary across regions, the qualitative patterns and management implications derived from citizen-collected data remain strongly aligned with those obtained through official monitoring programs.

A detailed comparison between the OoR program and CETESB was conducted at co-monitored stretches of the Tietê River, located within the Alto Tietê mesobasin. Both institutions monitor corresponding points in this area. The evaluated CETESB sites include Biritiba-Mirim (ID2050), Mogi das Cruzes (ID2090), Suzano (ID3120), Itaquaquecetuba (ID3130), Guarulhos (ID4150), and São Paulo (ID4170/4180/4200), with the latter represented by the annual mean across the three São Paulo stations. The objective of this analysis was to assess the robustness of concordance in segments historically associated with poor water quality (Figure 5) and to explore potential systematic trends introduced by the OoR methodology. This focus is justified by the predominance of the “Regular” category in CETESB’s long-term averages for the mesobasins—a range that, due to its broad amplitude, encompasses 32.1% of the OoR scale.

Figure 5

Across these co-monitored sites, exact agreement between OoR and CETESB classifications ranged from 0% to 100%, with a mean concordance of approximately 33%. When a tolerance of ±1 category was considered, agreement increased substantially, varying from 20% to 100% and yielding an average of 71.4%, indicating that most discrepancies were confined to adjacent qualitative classes. Cohen’s Kappa values for these stretches ranged from 0.12 to 0.68, with a mean of 0.43, reflecting overall moderate agreement but with pronounced spatial variability, especially in highly urbanized segments.

Number of samples, success rates, and completeness

The analysis of water quality monitoring data from the Alto Tietê, Sorocaba/Médio Tietê, and PCJ river basins (2002–2023) revealed distinct patterns in monitoring effort and analytical consistency. Figure 6 synthesizes these patterns by presenting, for each mesobasin and year, (i) the proportion of samples that included 11, 12, 13, or 14 parameters and (ii) the total number of sampling events, represented by labeled circular markers.

Figure 6

Although analytical completeness improved in certain periods, the trajectories differed among basins. In Alto Tietê, completeness rose mainly from 2014 onward after a mid-2000s decline. In Sorocaba/Médio Tietê, it peaked between 2006 and 2013 before decreasing in recent years. In the PCJ basin, completeness was initially high, then declined, and later increased again, reflecting a fluctuating rather than monotonic pattern.

Sampling effort, however, exhibited strong year-to-year variability. In Alto Tietê, the number of collected samples ranged from a minimum of 1 in 2002 (with zero samples in 2011–2012) to a peak of 684 in 2015. Sorocaba/Médio Tietê reached its maximum monitoring intensity in 2007 (449 samples) but declined to 8 in 2022, followed by a partial recovery in 2023. In the PCJ basin sampling effort increased gradually until 2017 (92 samples), declined markedly between 2018 and 2022 (minimum of 2 samples), and rose again in 2023 (76 samples). A Mann–Kendall trend analysis confirmed a significant upward trend only in PCJ (τ = 0.31, p = 0.045; Theil–Sen = +2 samples yr^–1), suggesting gradual expansion of monitoring capacity in this basin.

A bias-corrected binomial regression indicated an annual increase of approximately 5% in the likelihood that a sample contained all 14 parameters (p < 0.001), confirming long-term improvements in completeness across the network. Despite the sharp decline in 2020 caused by the COVID-19 pandemic, the dataset shows consistent recovery in all basins by 2023, without long-term loss of monitoring capacity.

Success rates for individual parameters (Supplemental File 3: Table S3) were generally high, particularly for turbidity, DO, ${\mathrm{\text{PO}}}_4^{\mathrm{\text{3–}}}$, foams, debris, odor, fish, and settleable solids, which frequently approached 100%. ${\mathrm{\text{NO}}}_3^{–}$ showed near-complete success after 2009. In contrast, pH and biological indicators (red and transparent/dark worms) exhibited greater variability. Mann–Kendall tests detected a significant upward trend in pH success rates in Alto Tietê (τ = 0.49, p = 0.003) and Sorocaba/Médio Tietê (τ = 0.43, p = 0.008), while ${\mathrm{\text{PO}}}_4^{\mathrm{\text{3–}}}$ presented a slight but significant decline in both basins (τ ≈ –0.43, p < 0.03). Biological indicators showed no long-term trends and presented expected inconsistencies linked to visual identification and site conditions.

Importantly, the increased completeness over time did not produce a measurable effect on the agreement between OoR and CETESB classifications. This result is consistent with the mathematical structure of the OoR index: Missing analytical parameters are compensated internally within the WQI calculation, preventing structural bias. Therefore, the absence of a statistical correlation between completeness and concordance is interpreted as evidence of the effectiveness of this compensatory mechanism.

Monitoring network coverage community participation and engagement

The volume of sampling conducted by the OoR program in these basins, particularly in the past decade and in the years since the lifting of COVID-19 restrictions, frequently surpassed the number of samples collected by the state environmental agency (CETESB) in the Alto Tietê and Sorocaba/Médio Tietê mesobasins (Figure 7). In the PCJ region, however, CETESB maintained higher sampling volumes across the three evaluated periods.

Figure 7

This result reflects the broad reach of the program: While CETESB focuses on large rivers, the OoR network often includes small or neglected water bodies, located in areas that are difficult for official monitoring teams to access or of particular importance to the local community.

A total of 187 different water bodies were monitored across the three mesobasins (Supplemental File 4: Table S4), including reservoirs, urban streams, and small, hard-to-access tributaries. Several of these sites even lacked official names, being identified instead by local residents according to neighborhood or historical references—a reflection of the program’s strong territorial embeddedness. Figure 8 (panels 8a–c) illustrates this spatial diversity by showing the annual distribution of sampling events across municipalities: Highlighted colors correspond to the most representative locations in each mesobasin, while all others appear as “other municipalities.”

Figure 8

In the Alto Tietê basin (Figure 8a), São Paulo city alone accounted for nearly 70% of all monitoring events, including the program’s highest single-municipality peak in 2004. This predominance is expected, as São Paulo is both Brazil’s largest metropolis and the birthplace of the OoR program (2002), currently hosting 41 active volunteer groups. Municipalities such as Mogi das Cruzes and Suzano also contributed substantially to regional sampling efforts.

In the Sorocaba/Médio Tietê basin (Figure 8b), overall activity was lower, but cities such as Itu—home to SOS Mata Atlântica’s headquarters—stood out as key operational hubs, especially during the 2010s. Salto, known for its frequent foam pollution events on the Tietê River, and Ibiúna, where monitoring first began in this basin, also exhibited consistent engagement.

In the PCJ basin (Figure 8c), the monitoring structure shifted markedly after 2016 with the emergence of more organized groups in Campinas and Amparo, often supported by corporate partnerships. Earlier, in the 2000s, Mairiporã had been the only participating municipality, while from 2007 onward, Indaiatuba became important before later discontinuing activities. This trajectory reflects the dynamic participation patterns characteristic of the basin.

In certain municipalities, continuous monitoring of major rivers such as the Camanducaia and Anhumas in the PCJ basin (Amparo and Campinas) is carried out by volunteer teams formed through partnerships with local companies. A similar situation is observed in the Médio Tietê basin, where neighborhood associations, public schools, and universities (e.g., Salto, Itu, Tietê, and Cabreúva) manage regional monitoring activities—comparable in frequency to the work of public sanitation agencies.

Since its inception in the mid 1990s, expansion in the 2000s, and consolidation in the 2010s, OoR experienced fluctuations in volunteer engagement but has maintained continuous operation. Group sizes varied by mesobasin: Alto Tietê - average group size between 4 and 28 participants. Sorocaba/Médio Tietê - average of 6 to 58 participants per event. And PCJ - average of 2 to 19 participants.