Water Quality Hotspot Prediction

ISEF Category: Environmental Engineering

Subcategory: Water Resources Management  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A stream can look fine and still carry pollution. Nitrate often hides in water that seems clear. That makes it a perfect problem for a student project, because you can measure the clue, map the pattern, and test whether land use helps explain it.

What Is It?

This project turns everyday water checks into a prediction problem. You collect water quality data with a turbidity tube, colorimetric strips, and a phone GPS app. Turbidity tells you how cloudy the water is. Nitrate strips give you a rough chemical readout. GPS tags each sample so you can map where the data came from.

Then you ask a bigger question, can you predict where nitrate hot spots will show up before you sample them? A machine-learning classifier is just a model that learns patterns from examples. In this case, it can look at features like nearby land use, slope, elevation, and distance to roads or fields. Think of it like training a spam filter, but for water pollution risk.

Why This Is a Good Topic

This is a strong science fair topic because you can measure real water quality, connect it to a real environmental problem, and build a model that makes predictions. You do not need a university lab to start. You can collect your own field data, use public map data, and test whether simple landscape features can explain nitrate patterns. That gives you both hands-on science and data analysis.

Research Questions

• How does nearby land use affect nitrate levels in small streams?
• What is the effect of elevation on predicted nitrate hot spots?
• Does turbidity help predict nitrate concentration when land use is held constant?
• To what extent does distance from agricultural land improve classifier accuracy?
• Which landscape features, such as slope, land use, or road density, best predict high nitrate sites?
• How does a simple rule-based map compare with a machine-learning classifier for hotspot prediction?

Basic Materials

• Turbidity tube or Secchi-style tube.
• Nitrate colorimetric test strips with a known range.
• Phosphate test strips, if you want a comparison variable.
• Smartphone with GPS and camera.
• Field notebook or data sheet.
• Portable cooler or sample bottles.
• Permanent marker for sample labels.
• Ruler or measuring tape for local site notes.
• Laptop or desktop computer.
• Spreadsheet software such as Google Sheets or Excel.

Advanced Materials

• Handheld spectrophotometer or colorimeter.
• Portable GPS logger with higher location precision.
• Water sample filtration setup.
• Conductivity meter.
• pH meter.
• Dissolved oxygen meter.
• GIS software such as QGIS.
• Python environment for model building.
• Access to land-cover, elevation, and watershed datasets.
• Calibration standards for nitrate or turbidity validation.

Software & Tools

• Google Sheets: Organizes field data, cleans records, and makes simple charts.
• QGIS: Maps sample points and overlays land-use, slope, and elevation layers.
• Python: Builds and tests the classifier with your water quality data.
• scikit-learn: Trains machine-learning models and compares prediction scores.
• ImageJ: Helps if you photograph strips and want to measure color intensity.

Experiment Steps

1. Define the water bodies you will sample and the exact question your model should answer.
2. Decide which field measurements you will collect at every site, then keep those variables consistent.
3. Build a site table that links each sample to land-use, elevation, and distance features from map data.
4. Choose a target label for the model, such as high nitrate versus not high nitrate, and define it before analyzing results.
5. Split your data into training and test sets so you can judge prediction honestly.
6. Compare a simple baseline model with a more complex classifier, then check which features drive the result.

Common Pitfalls

• Using nitrate strips in bright sunlight at different angles, which makes color readings hard to compare.
• Sampling only one part of each stream, which can hide real variation across the watershed.
• Mixing up site labels and GPS points, which breaks the link between water data and map features.
• Trying to predict nitrate hot spots with too few samples, which makes the classifier overfit noise.
• Skipping a baseline model, which makes it hard to know whether machine learning really helped.

What Makes This Competitive

A competitive version of this project uses clean field methods and careful data labeling. You can raise the level by comparing several models, testing feature importance, and checking whether your results hold in a separate watershed or season. Strong projects also handle uncertainty well, for example by reporting confidence intervals, class balance, and false positive rates. That kind of analysis shows that you understand both the water science and the prediction problem.

Project Variations

• Use urban streams instead of agricultural waterways to see whether impervious surfaces predict nitrate hot spots differently.
• Swap nitrate strips for phosphate strips and test whether the same land-use features still predict hot spots.
• Add rainfall history as a feature and compare dry-weather samples with storm-runoff samples.

Learn More

• USGS Water Science School: Search the USGS site for clear explanations of turbidity, nitrate, and watershed basics.
• NOAA National Water Center: Search NOAA for watershed and runoff resources that help connect rainfall to water quality.
• EPA Water Research: Search the EPA site for background on nutrient pollution, stormwater, and freshwater monitoring.
• PubMed: Search for review articles on citizen science water monitoring and nitrate detection methods.
• QGIS Documentation: Use the official QGIS manual for mapping sample points and working with elevation and land-cover layers.
• MIT OpenCourseWare: Search for introductory classes on machine learning, data analysis, and environmental engineering.