How the AI works
Volunteer observations can vary. A common fix is to let AI “correct” them, but that throws away what only a person at the stream can know, and it hides the AI's own mistakes. StreamSense uses AI to prompt, explain and cross check. Every decision stays with a person.
Step by step
Steps in pencil blue involve the AI.
The volunteer takes photos
Photos are made smaller on the phone to save mobile data.
The photo assistant looks (uses AI)
The AI suggests answers for the 9 questions a photo can show, each with how sure it is and what it saw. It can say "can't tell", and it never guesses smell or chemistry.
Safety checks
Server code checks every suggestion against the survey, drops anything it doesn't recognise, and keeps one suggestion per question.
The volunteer decides
Suggestions sit next to the question but are never filled in. Every accept or reject is saved for reviewers.
Fixed scoring
The health score and alerts come from the published rules below. The same answers always give the same result.
Data checks
Impossible values, answers that clash, and places where the volunteer and the AI disagree are flagged with friendly prompts to double check.
AI second opinion (uses AI)
An optional plain English explanation of the results and what they mean for people, animals and nature. It can read the score but can't change it.
Human reviewer
Flagged reports wait in the review queue, where an expert verifies them or asks for a follow up visit.
The health score formula
Each answer has a stress level from 0 (no pressure) to 3 (severe). The score is 100 × (1 − Σ(weight × stress) / (3 × Σ weight)), counting answered questions only. It needs at least 3 answers.
- Healthy
75 and above - Moderate
50 to 74 - Stressed
25 to 49 - Critical
below 25
| What is measured | Weight | Can the AI suggest it? |
|---|---|---|
| Turbidity | 1.2 | Yes, from the photo |
| Apparent colour | 1 | Yes, from the photo |
| Odour | 1.3 | No, volunteer only |
| Surface film and foam | 1 | Yes, from the photo |
| Algae and eutrophication | 1.2 | Yes, from the photo |
| Anthropogenic litter | 0.8 | Yes, from the photo |
| Flow regime | 0.6 | Yes, from the photo |
| Riparian vegetation | 1 | Yes, from the photo |
| Bank erosion | 0.8 | Yes, from the photo |
| Outfalls | 1.1 | Yes, from the photo |
Known limits
- Photos can't show smell, chemistry or anything outside the frame. That's why those questions are for the volunteer only.
- Glare, shade and camera colour settings change how the water looks. The assistant is told to lower its confidence when this happens.
- The score is a screening tool for volunteers. It doesn't replace lab testing.
- Without an API key the app runs in demo mode with sample AI answers, and these are clearly marked.