Fish Finder AT3b

Presenting to the client

This website is the client presentation layer. It walks through the problem, shows how the data was prepared, and presents evidence from visualisations, the spreadsheet, and intelligent-system outcomes — ending with what we learnt and what was delivered.

The live Streamlit app runs locally for hands-on demo; this site explains the project results for client review without hosting models or raw data files.

Scroll through sections 2–11 for detailed evidence, then section 12 for a summary of learnings and client outcomes.

Weather and moon enrichment

scripts/enrich_fish_data.py enriches every row with moon phase, wind direction, and rainfall from the Open-Meteo Historical Weather API (batched by ~10 nm grid cells and cached in enrichment_cache.csv). Output: GameFish_Releases_Master_enriched.csv — still 203,300 rows, now with extra weather columns.

Change from previous stage: Δ 0 rows (columns added, no rows removed). Sea surface temperature was also fetched but returned 100% null for every row, so the Sea_Surface_Temp_C column was unusable and later excluded.

Weather-filtered subset

Only rows with usable Wind_Direction_10m and Rain_mm values were kept (~24.5% of enriched rows had Open-Meteo API coverage for their date/location). This produced Weather_GameFish_Releases_enriched.csv with 49,893 rows.

Change from previous stage: Δ −153,407 rows (203,300 → 49,893).

Row count drops here because many older or remote records fall outside the weather archive API's coverage — a documented data limitation, not random deletion.

Final primary dataset

Rows missing a recorded Length or Weight were removed, yielding Cleaned_Weather_GameFish_Releases_enriched.csv — 41,440 rows. The Sea_Surface_Temp_C column was also filtered out at this stage because it was 100% null across the enriched dataset.

Change from previous stage: Δ −8,453 rows (49,893 → 41,440).

This is the final primary dataset used by both the Streamlit app (BasicProduction/fish_app.py) and the neural network trainer (BasicProduction/train_species_nn.py), ensuring the app and model always use the same features.

What we learnt: a single cleaned primary dataset keeps the app, model, and spreadsheet aligned; provenance CSVs at each stage are kept as evidence.

Stage 2: Streamlit + Folium dashboard

Replaced static exports with a live, in-browser dashboard (BasicProduction/fish_app.py). Folium renders an interactive Leaflet map; Streamlit adds species/year/month filtering, manual and AI-assisted modes, trip estimation, and route export.

This prototype later became the base for the intelligent systems (neural network, what-if analysis, certainty-factor engine) — added in later project stages, not in this website section.

What we learnt: the dashboard became the integration point — map filters, offline LLM summaries, and user ratings later fed the expert system.

Stage 3: Power BI reference dashboard

A Power BI dashboard (Dashboard-FishFinder.pbix copy) was built as a business-intelligence-style visualisation alongside the Streamlit app — useful for comparing chart types, slicers, and BI layout patterns during design.

The .pbix file stays in the repository unchanged. This presentation site references it through screenshots only — the file is never copied into the website.

What we learnt: Streamlit suits interactive intelligent systems; Power BI suits polished reporting — both visualise the same data story for different audiences.

Model evaluation and alignment

The model was evaluated using an 80% / 20% train–test split (stratified by species class): 33,152 rows for training and 8,288 rows held out for testing.

Metrics on the test set: 63.4% accuracy, 56.3% macro F1, 62.1% weighted F1 — across 13 classes (12 named species + an "Other" bucket).

A feature-mismatch bug was fixed: the app now reads model_info['feature_columns'] from the saved model and builds every prediction input dynamically, so app inputs always match what the network was trained on.

Moderate accuracy is expected — species distribution depends on many factors not captured here (bait, vessel, angler skill, time of day).

What we learnt: location, season, and weather explain part of species patterns (~63% accuracy) — useful for guidance, not a guarantee of catch.

Baseline vs adjusted comparison

When the user clicks "Run what-if comparison", the app generates two predictions using the same neural network and feature pipeline: a baseline (current context) and an adjusted scenario (with the user's changed values).

Results appear as top-species metrics, a bar chart comparing baseline vs adjusted probabilities, and a table showing the percentage-point change per species — making the effect of each variable change visible and explainable in a walkthrough video.

Like the neural network, what-if analysis runs locally in Streamlit — this website explains and shows screenshot evidence only.

What we learnt: changing month, wind, or location visibly shifts species probabilities — clients can see why a recommendation changes.

Thresholds and output

Confidence bands: score ≥ 75 → Strong recommendation; 45–74 → Moderate; 1–44 → Weak (advise changing species, season, or location); 0 → Insufficient data (species not in dataset).

Each evaluation returns: a recommended species, certainty score, confidence band, a table of rules fired (with points and plain-English explanations), and a summary sentence. Example: "Moderate recommendation (50/110) for ALBACORE in February."

The expert system and neural network are toggleable/separate panels in the AI Assisted page — together they demonstrate both predictive modelling and expert-system reasoning.

What we learnt: rule-based scoring gives clients a readable explanation (rules fired + points) alongside the neural network's probability output.

Rule engine, spreadsheet & consistency checks

Certainty-factor rules — each rule (species, season, catch density, location, user preference) was tested individually with known inputs to confirm the correct points are added and the running total never exceeds 110.

Spreadsheet generation — build_nesa_spreadsheet.py was re-run end-to-end to confirm it produces 18,225 rows with a working Weight_Classification formula and no missing-column errors.

Dataset path / feature consistency — verified that the app, trainer, and spreadsheet script all read from the same cleaned CSV, and that the app's prediction inputs match model_info['feature_columns'] exactly (the fix for the earlier feature-mismatch bug).

The pass/fail checklist in the left visual panel (sections 9a and 9b) summarises each quality check — so you can see at a glance which parts of the system were verified during development.

Walkthrough steps

The live demo walks through the system in this order:

1. Launch the Streamlit app — load the interactive map and dataset.
2. Filter the map — choose species, year, and month to explore catch patterns.
3. Neural Network Explorer — run a species prediction from location, time, and weather.
4. What-If Analysis — change month, wind, rain, moon phase, or location and compare probabilities.
5. Certainty-Factor Expert System — evaluate a species with rule-based scoring and plain-English explanations.
6. Spreadsheet export — review the generated Excel workbook with cleansed data and formulae.
7. Power BI reference — view dashboard screenshots for reporting-style visualisations.

A recorded walkthrough video can be linked from the visual panel once available.

Outcomes delivered to the client

The enterprise system delivers evidence through visualisations, a spreadsheet, and intelligent-system outcomes — all grounded in the same cleaned dataset.

• Streamlit + Folium dashboard — interactive map, filters, AI-assisted mode, intelligent-system panels (live app).
• Excel spreadsheet — 18,225 cleansed rows with Weight_Classification IF formula.
• Power BI dashboard — BI visualisation reference (screenshot evidence on this site).
• Neural network + certainty-factor expert system — species prediction, what-if analysis, and rule-based recommendations.

Predictions and certainty scores support fishing decisions — they do not guarantee a catch. Known limitations are documented in section 10.

Client takeaway: tagging data can be turned into an explorable, offline-capable enterprise tool with both data-driven and rule-based guidance.