# Help with Final Year Major Project **URL:** \ **Category:** Discussion & Solutions\ **Tags:** help\ **Created:** [September 10, 2026, 1:26pm UTC](https://onehack.st/t/help-with-final-year-major-project/325450 "2026-09-10T13:26:26Z")\ **Posts on this page:** 1\ **Showing post:** 3
**Author:** ![Indianapolis](https://onehack.st/user_avatar/onehack.st/indianapolis/32/150586_2.png) [@Indianapolis](https://onehack.st/u/Indianapolis)\ **Post date:** [October 3, 2026, 10:40am UTC](https://onehack.st/t/help-with-final-year-major-project/325450/3 "2026-10-03T10:40:15Z")
You probably do not need someone else’s finished project. You need a manageable baseline, a clear research question, and one small thing that makes your work different. Your professor is probably tired of seeing: ```plaintext Download a popular dataset→ train CNN/YOLO/LSTM→ report accuracy→ build a Streamlit interface ``` That does not mean deep-learning projects are useless. It means the project needs a stronger question than “can I train a model?” ## **The easiest way to make a project original** Take an existing method and change one important variable: - Use a different dataset or region - Compare lightweight and large models - Test performance with limited data - Study bias or failure cases - Make the model work offline - Run it on low-end hardware - Test it under noise or missing data - Compare accuracy against speed and memory - Reproduce an existing paper and explain where it fails - Add human feedback or active learning - Build a useful system around the model instead of only showing accuracy A good final-year project can be: > “I evaluated whether an existing method works reliably under a new constraint.” It does not have to invent a brand-new neural-network architecture. ## **Project ideas that are realistic** ### **1. Wildfire detection with uncertainty** Use satellite fire detections and weather data to classify areas as low, medium, or high risk. Possible research question: > Does adding weather and vegetation information improve wildfire-risk prediction compared with satellite data alone? You can compare: - Random Forest - XGBoost - CNN or LSTM - A simple baseline model Add a useful angle: - Compare different regions - Test seasonal performance - Measure false alarms - Show uncertainty instead of only a yes/no prediction - Build a map dashboard NASA FIRMS provides near-real-time active-fire data, although the near-real-time product is not considered science-quality, so your project should clearly describe its limitations ### **2. Lightweight plant-disease detection** Instead of training a large image model, compare models designed for low-resource devices. Research question: > How much accuracy is lost when a plant-disease model is compressed for mobile or edge-device deployment? Measure: ```plaintext AccuracyModel sizeInference timeRAM usageEnergy or CPU usage ``` Compare: - MobileNet - EfficientNet-Lite - Quantized versions - A larger model as the baseline This is stronger than simply saying “I trained a plant-disease classifier.” ### **3. Misinformation or claim classification** Build a system that classifies claims as supported, contradicted, or unverifiable using a public fact-checking dataset. The important part is not just the classifier. Study: - Performance across topics - Performance on short versus long claims - Whether the model relies on keywords - Explainability - False confidence - Performance when the claim is reworded Do not present it as a system that decides absolute truth. Present it as a research prototype that assists fact-checking. ### **4. Traffic or air-quality forecasting** Use public weather, pollution, or traffic data to predict the next hour or day. Compare: - Historical-average baseline - Linear regression - Random Forest - XGBoost - LSTM or Temporal Fusion model Then test: - Different forecasting horizons - Missing sensor values - Different cities - Extreme weather days - Model performance versus computational cost A project becomes much more interesting when you answer: > “When does the deep-learning model actually outperform simpler methods?” Sometimes the result will be that it does not. That can still be a valid conclusion if the experiment is well designed. ### **5. Offline disaster-report classifier** Create a small text classifier that categorizes emergency messages into: - Medical - Shelter - Food/water - Infrastructure - Transportation - Safety The novelty could be: - Works offline - Supports multiple languages - Runs on a low-end laptop - Handles noisy social-media text - Compares a small model with a cloud API - Measures performance when training data is limited This gives you a complete project: ```plaintext Dataset→ preprocessing→ baseline model→ improved model→ error analysis→ offline demo→ limitations ``` ### **6. Sports analytics with explainability** Use a public sports dataset to predict match outcomes, player performance, or injury risk. Do not stop at “my model achieved 87% accuracy.” Study: - Whether the model is overfitting to teams or players - How performance changes over time - Which features influence predictions - Whether the model remains useful on a new season - Whether a simple model performs almost as well The interesting part is often the analysis, not the sport itself. ## **A simple project structure** Use this structure for almost any topic: ```plaintext 1. Define the problem 2. Find one public dataset 3. Implement a simple baseline 4. Implement one stronger method 5. Add one constraint or research angle 6. Compare the results 7. Perform error analysis 8. Build a small demonstration 9. Document limitations 10. Make everything reproducible ``` Your final report should answer: - What problem are you solving? - Why does it matter? - What has already been done? - What is different about your project? - What is your baseline? - What experiment did you perform? - What failed? - What are the limitations? - Can someone else reproduce it? ## **Where to find inspiration** Use these places for ideas and baselines, not for copying and submitting someone else’s project: - GitHub project-based-learning repositories - Build-your-own-X projects - Devpost and hackathon submissions - University project archives - Kaggle datasets and notebooks - Papers on arXiv and Semantic Scholar - NASA, NOAA, USGS, and other public-data portals - Open-source repositories related to your chosen method NOAA maintains a large collection of openly accessible datasets across weather, climate, ocean, and environmental domains, which gives you many options beyond the usual image-classification datasets. **[National Oceanic and Atmospheric Administration](https://www.noaa.gov/nodd/datasets)** ## **The fastest way to choose one** Give yourself 30 minutes: ```plaintext Choose one area→ choose one public dataset→ find two related papers→ read their limitations and future-work sections→ write one research question→ check whether you can build a baseline in one week ``` For example: > “Can a lightweight model detect wildfire activity from public satellite data while using less memory than a standard CNN?” That is already a better project direction than: > “Wildfire detection using deep learning.” ## **What I would avoid** - Downloading a complete GitHub project and changing the title - Using ChatGPT to write the entire report without understanding it - Choosing a dataset before defining the research question - Reporting only accuracy - Using a huge model that you cannot explain - Making medical or security claims your experiment cannot support - Depending on a paid API for the entire project - Building a complicated mobile app before validating the model - Claiming novelty when your only change is the user interface The safest route is to use an existing implementation as a **baseline** , cite it properly, understand it, and then perform your own experiment. A strong final-year project could simply be: > “I reproduced an existing method, tested it on a different dataset or under a realistic constraint, compared it with simpler baselines, analyzed its failures, and built a reproducible demonstration.” That is much more defensible than submitting a copied “AI project” with a flashy interface. --- _[View the full topic](https://onehack.st/t/help-with-final-year-major-project/325450)._