Water Quality Index Modelling and Trend Analysis of a Tungabhadra River Stretch Using CPCB Monitoring Data

This dissertation converts the Central Pollution Control Board's (CPCB) monthly monitoring records for a stretch of the Tungabhadra river in Karnataka — the reach that passes the industrial towns of Harihar and Hospet — into a Water Quality Index (WQI), a defensible trend analysis, and a small predictive model. The Tungabhadra supplies drinking-water intakes, irrigation lifts and industry along its course, yet its condition is usually discussed through raw parameter tables that no planner can interpret at a glance.

The study acquires about ten years of station-level records from CPCB's National Water Quality Monitoring Programme (NWMP) data portal and the India-WRIS system, harmonises them into a tidy dataset, and computes the WQI by the weighted-arithmetic method on BIS IS 10500:2012 limits, cross-checked against the NSF-WQI with class-agreement reported as Cohen's kappa. Monthly and seasonal trends are tested with the Mann–Kendall test and Sen's slope estimator, with lag-1 autocorrelation handled explicitly. A second stage models the index itself: multiple linear regression, random forest and gradient boosting predict WQI from raw parameters under leave-one-year-out cross-validation, so no future year can leak into training. Parameter importance identifies which constituents drive the index along the stretch.

Everything runs in Python with pandas, SciPy, scikit-learn and pymannkendall; QGIS, an industry-standard extra, maps the stations. Written for the VTU M.Tech 2022 scheme Project Work Phase-2 (22SXX41) and shaped by the Research Methodology & IPR course (22RMI16), this bundle documents the full protocol with empty results tables — every number must come from your own runs.

CPCB's monthly river-monitoring records for the Tungabhadra stretch are published as raw parameter tables: one number per parameter per station per month. District officials, basin boards and researchers cannot easily answer the questions that matter — is the stretch improving or deteriorating over ten years, which months and seasons carry the worst load, which parameters drive the degradation, and how do stations rank against drinking-water criteria? Parameter tables hide answers that an index, a trend test and a simple model reveal immediately.

Existing studies of the river either cover short periods, use undisclosed cleaning rules, or fit models with random cross-validation that leaks future information. There is a need for a reproducible, protocol-first study that converts public NWMP records into a defensible Water Quality Index, quantifies statistically significant trends, models the index from raw parameters without temporal leakage, and documents every assumption so the results survive examination.

Abstract

River-water quality in India is monitored monthly at thousands of stations, but the records reach administrators as long parameter tables. This dissertation builds a reproducible analysis pipeline for one heavily used Karnataka river: the Harihar–Hospet–Siruguppa stretch of the Tungabhadra. About ten years of monthly records for three to five NWMP stations are cleaned, converted into a weighted-arithmetic Water Quality Index on BIS IS 10500:2012 limits, tested for trend, and modelled with machine-learning regressors under leakage-safe cross-validation.

Introduction

The Tungabhadra, a major Krishna tributary, carries industrial effluent from towns such as Harihar, irrigation return flows from the Tungabhadra command area, and domestic sewage from growing towns. CPCB's NWMP publishes monthly values for pH, dissolved oxygen, BOD, conductivity, nitrate, total coliform and more at fixed stations. CPCB itself uses these data to notify polluted river stretches, so a station-level index and trend analysis connects directly to regulatory practice.

Literature Gap

• Most Indian WQI studies compute an index for one or two years and stop; trend statistics and index modelling rarely appear together.
• Papers often skip the boring decisions that change answers: treatment of below-detection values, missing months, and station relocations.
• Gap: a single, fully documented pipeline — declared cleaning rules, two WQI formulations compared by class agreement, non-parametric trend tests with autocorrelation checks, and regression models evaluated by leave-one-year-out cross-validation — for a Tungabhadra stretch, reproducible by any student from public data.

Proposed Work

• A tidy station-month-parameter dataset with a written audit trail of every cleaning decision.
• WQI by weighted arithmetic (BIS limits) cross-checked with the NSF-WQI; agreement measured by Cohen's kappa.
• Monthly, seasonal (pre/post-monsoon) and full-record Mann–Kendall tests with Sen's slope and lag-1 autocorrelation handling.
• MLR, random forest and gradient-boosting models of WQI from raw parameters, with RMSE, MAE and R² plus feature importance.
• Decision-ready artefacts: station map, WQI time series, exceedance tables and a station-ranking narrative.

Feasibility

All data are public and free; every tool is open-source and runs on a laptop. The main cost is care: the dissertation earns its marks from the declared protocol, the audit trail and honest treatment of data gaps, not from software. A single student can complete the analysis in about twelve working weeks inside the 24-week Phase-2 slot.

In scope

• Secondary data only: CPCB NWMP monthly records and India-WRIS downloads for the selected Tungabhadra stations, roughly 2014–2024.
• Three to five fixed stations on the Harihar–Hospet–Siruguppa stretch, subject to data completeness.
• Conventional parameters published by NWMP: pH, temperature, dissolved oxygen, BOD, COD (where reported), conductivity, TSS or turbidity, nitrate, total and faecal coliform.
• Two index formulations (weighted arithmetic on BIS limits; NSF-WQI), non-parametric trend analysis, and three regression model families.

Out of scope

• New field sampling or laboratory analysis — no primary data are collected.
• Real-time continuous water-quality monitoring and sensor instrumentation.
• Hydrodynamic or pollutant-transport modelling of the river.
• Source apportionment or attribution of loads to specific industries beyond descriptive upstream–downstream comparison.
• Groundwater quality and reservoir (Tungabhadra dam) limnology.
• Predicting future policy scenarios — the models describe and explain, they do not forecast regulation.

The study uses a quantitative, longitudinal, secondary-data research design — a single observational cohort of station-month records followed across ten years. The protocol is declared before analysis begins, in the spirit of the Research Methodology & IPR course (22RMI16): every rule below is written down first and never changed after seeing results.

Declared protocol. (1) Data acquisition: download NWMP monthly records from the CPCB data portal and cross-verify station metadata on India-WRIS; record station coordinates and any relocation events. (2) Cleaning: parameter-months with more than 30% missing values across years are excluded; values reported as below detection limit are replaced by half the detection limit and flagged; unit harmonisation table is part of the dissertation appendix. (3) Index: weighted-arithmetic WQI with quality rating qi = 100 × (vi / si) for harmful parameters and the DO-based form for dissolved oxygen, unit weights wi = K / si with K = 1 / Σ(1/si); classes on the standard excellent-to-unsuitable bands; NSF-WQI computed in parallel as a robustness check. (4) Trends: monthly Mann–Kendall, seasonal Mann–Kendall (pre-monsoon and post-monsoon strata) and Sen's slope; lag-1 autocorrelation tested and trend-free pre-whitening applied when significant. (5) Modelling: features are raw parameters, target is the index; leave-one-year-out CV; metrics RMSE, MAE, R²; permutation feature importance on the best model.

Threats to validity: station relocation, changes in laboratory methods, and detection-limit revisions are logged and discussed; results are reported per station so a single odd station cannot drive conclusions.

The study is organised as a one-way analysis pipeline with explicit artefacts between stages, so each stage can be checked — and defended — independently. Raw downloads land in a data directory; the curation stage emits a frozen tidy dataset plus an audit trail; the index, trend and modelling stages read only that frozen dataset, never the raw files; the reporting stage consumes only their outputs.

flowchart TD
  A["CPCB NWMP monthly records"] --> D
  B["India-WRIS station metadata"] --> D
  C["SPCB reports for cross-checks"] --> D
  D["Data curation and audit trail"] --> E[("Frozen tidy dataset<br/>station-month-parameter")]
  E --> F["WQI engine: weighted arithmetic<br/>plus NSF-WQI cross-check"]
  E --> G["Trend module: Mann-Kendall,<br/>Sen slope, pre-whitening"]
  E --> H["Modelling module: MLR, RF, GBM<br/>leave-one-year-out CV"]
  F --> I["WQI series, class bands,<br/>exceedance tables"]
  G --> J["Trend table with slopes<br/>and significance"]
  H --> K["Metrics and feature<br/>importance"]
  I --> L["Synthesis: QGIS map,<br/>figures, station ranking"]
  J --> L
  K --> L

The data model behind the frozen dataset is deliberately simple — three tables — so that the schema fits on one dissertation page.

erDiagram
  STATION ||--o{ SAMPLE : "is measured at"
  SAMPLE ||--o{ READING : contains
  PARAMETER ||--o{ READING : "is recorded as"
  STATION {
    string station_id PK
    string name
    float lat
    float lon
    date relocation_date "nullable"
  }
  SAMPLE {
    string sample_id PK
    string station_id FK
    date sample_date
    string season
  }
  READING {
    string reading_id PK
    string sample_id FK
    string parameter_id FK
    float value
    boolean below_detection
  }
  PARAMETER {
    string parameter_id PK
    string name
    string unit
    float bis_limit
  }

This modularity matters in the viva: when an examiner asks "what happens if you treat below-detection values differently?", the answer is a controlled re-run of one module, not a rebuild.

• Forecasting: extend the best regression model to a seasonal ARIMA or Prophet-style forecast of the index, with uncertainty bands, so the board can anticipate deteriorating months.
• Continuous data fusion: join the monthly NWMP records with real-time continuous water-quality monitoring station feeds where available, and quantify what the higher-frequency data add to trend detection.
• Satellite support: correlate in-situ turbidity and chlorophyll-related parameters with Sentinel-2 reflectance products to interpolate conditions between stations.
• Basin scale: replicate the pipeline for every NWMP station in the Krishna basin and publish an open dashboard with the WQI series and trend map.
• Bayesian trend models: replace or supplement Mann–Kendall with Bayesian change-point detection to date regime shifts, for example around monsoon anomalies or industrial expansion.