Artificial Intelligence + Agricultural Technology
Recognise plant diseases from leaf photos
Train a model on public leaf images to tell healthy leaves from diseased ones and test it on your own photos.
- DifficultyAdvanced
- Estimated time1–3 months
- Budget₹0
- ClassCollege, years 2–4
- Team2–4 people
- You needComputer only
Prerequisites: Python, a first neural-network project, basic statistics
Original · by TechnifyedAI/ML ProjectFinal-Year ProjectFree to doComputer onlyPortfolio focusedWell documentedTwo subjects
Overview
Public datasets hold tens of thousands of labelled leaf images. With transfer learning, a pretrained network can be adapted to classify diseases in a few hours. The hard and honest part is testing on real field photos, where accuracy usually drops.
Why build this?
It is a capstone with a real user, and the gap between lab and field accuracy is itself a strong finding.
What you will learn
- How transfer learning reuses a pretrained network
- How augmentation reduces overfitting
- Why test data must come from a different source
- How to report results per class
Skills you will use
Materials and tools
- Computer with Python (a free cloud notebook with a GPU helps)
- A public plant-disease image dataset
- A phone to photograph real leaves
Tools and software
Step-by-step roadmap
Start
You will finish with: a trained model, per-class metrics, a field test and a discussion of what would be needed for real use.
- 1
Prerequisites
Know this first: python, a first neural-network project, basic statistics. Then collect the 3 items in the materials list.
- 2
Learn
- How transfer learning reuses a pretrained network
- How augmentation reduces overfitting
- Why test data must come from a different source
- How to report results per class
- 3
Plan
List the features for version one, sketch the screens or data flow and pick the tools. Keep the first version small.
- 4
Build
- Download the dataset and pick three crops with four or five classes each.
- Split by image into training, validation and test sets.
- Fine-tune a small pretrained network with augmentation.
- Report accuracy per class and a confusion matrix.
- Photograph 50 real leaves in a garden or farm and test on them.
- Compare lab and field accuracy and explain the gap.
- 5
Test
Field accuracy is the real result; report it even if it is much lower.
- 6
Document
Write a README: what it does, how to run it, screenshots and what you learned. Report structure
- 7
Present
Show a live demo of one complete task, then the design and the hardest problem you solved. Presentation structure · Viva questions
- 8
Publish
Put the code on GitHub with the README and, if you can, deploy a live demo.
Expected outcome
A trained model, per-class metrics, a field test and a discussion of what would be needed for real use.
Other versions of this project
Beginner version
Classify only healthy versus diseased for one crop.
Advanced version
Add a heat map that shows which part of the leaf the model looked at.
Research version
Study how background, lighting and leaf age affect accuracy with a controlled photo set.
Social-impact version
Test it with a local farmer or nursery and record what they would need to trust it.
Portfolio version
Build a phone-friendly page where a farmer uploads a photo and sees the top two guesses with confidence.
Make this project better
- BasicClassify only healthy versus diseased for one crop.
- This projectRecognise plant diseases from leaf photos
- AdvancedAdd a heat map that shows which part of the leaf the model looked at.
- ResearchStudy how background, lighting and leaf age affect accuracy with a controlled photo set.
- PortfolioBuild a phone-friendly page where a farmer uploads a photo and sees the top two guesses with confidence.
Ways to do this project
Text marked "You write this" is a prompt for your own work; everything else is specific to this project.
- Research question
- You write thisOne question you can answer with data, narrow enough to finish in the time you have.
- Hypothesis
- You write thisWhat you expect to find, and why.
- Variables
- You write thisWhat you change or compare, what you measure, and what you hold constant.
- Methodology
- Study how background, lighting and leaf age affect accuracy with a controlled photo set.
- Data collection
- You write thisSay what you will record, how many times, and how you will keep the records safe.
- Analysis
- You write thisChoose the table, chart or test that answers the question; report averages with their spread.
- Limitations
- You write thisList what could have affected the result: small sample, instrument limits, things you could not control.
Literature review
Find five to eight sources (your textbook, review articles, reports) and note what each says about your question. Group them by idea, not one after another, and end with what is still not known.
References
List every source you used in one style throughout (author, year, title, where it was published, link and the date you opened it).
Use this as the outline of your README or case study.
- Problem
- Public datasets hold tens of thousands of labelled leaf images.
- Why it matters
- It is a capstone with a real user, and the gap between lab and field accuracy is itself a strong finding.
- What was built
- A trained model, per-class metrics, a field test and a discussion of what would be needed for real use.
- Technology
- OpenCV, Python, TensorFlow
- Architecture
- Photo → resize and normalise → pretrained network → new final layer → class probabilities.
- Implementation
- You write thisThe 6 build steps in your own words, with one code or design decision you are proud of.
- Challenges
- You write thisThe hardest bug or decision, and how you got past it.
- Results
- Field accuracy is the real result; report it even if it is much lower.
- Demo and code
- You write thisA link to a live demo or a short video, and to the GitHub repository.
- Lessons learned
- You write thisWhat you would do differently next time.
- Future improvements
- Add a heat map that shows which part of the leaf the model looked at.
- Problem
- Public datasets hold tens of thousands of labelled leaf images.
- Target user
- A small farmer or gardener with a phone and no nearby expert.
- Solution
- Train a model on public leaf images to tell healthy leaves from diseased ones and test it on your own photos.
- First version (MVP)
- Classify only healthy versus diseased for one crop.
- Tech stack
- OpenCV, Python, TensorFlow
- Architecture
- Photo → resize and normalise → pretrained network → new final layer → class probabilities.
- Demo
- Field accuracy is the real result; report it even if it is much lower.
- Stretch goals
- Add a heat map that shows which part of the leaf the model looked at.
Team roles
Plan by length
24-hour
- Hours 0–2: agree the problem, the user and the one thing the demo must show
- Hours 2–16: build only the first version
- Hours 16–20: test the demo path end to end and fix what breaks
- Hours 20–24: slides, a 2-minute demo script and a backup recording
48-hour
- Evening 1: problem, user, sketch and task split
- Day 1: first version working end to end
- Day 2 morning: one stretch goal and testing
- Day 2 afternoon: polish, slides and demo practice
1-week
- Day 1: research the problem and talk to two possible users
- Days 2–4: first version
- Day 5: stretch goals
- Day 6: testing and write-up
- Day 7: demo video and presentation
Working as a team
2–4 people. Suggested roles:
Report and presentation
Report structure
- Title and summary
- Problem and target user
- Features
- Tools and technology
- Design: architecture, data model or screens
- Implementation
- Testing
- Results and screenshots
- Challenges
- Future improvements
- References
Presentation structure
- The problem
- Who it is for
- Live demo
- How it is built (one diagram)
- The hardest part
- Testing and results
- What you learned
- What comes next
Viva questions
Try answering before you open each one.
What is transfer learning?
Starting from a network trained on a large dataset and retraining only part of it for a new task.
Why did accuracy fall on your own photos?
The training images have plain backgrounds and even lighting; real photos do not.
What is data augmentation?
Making altered copies (flipped, rotated, brightened) so the model generalises better.
Should a farmer rely on this?
Only as a first hint; it is not a replacement for an expert.
Resources
Image Processing Using OpenCVCampusX · Free video course
OpenCV with Python for Image and Video AnalysisSentdex · Free video course
OpenCV Python TutorialsTech With Tim · Free video courseOpenCV documentationdocs.opencv.orgPython documentationdocs.python.orgTensorFlow tutorialstensorflow.orgOpen Government Data Platform Indiadata.gov.inSource and attribution
Original
Written by Technifyed. Free to use for your own school or college project; write the report in your own words. Added 1 Oct 2026.
Similar projects
More Artificial Intelligence projects
Artificial Intelligence · Computer Vision
Recognise Indian sign language letters
Collect images of static signs, train a classifier and test it with people who did not provide training data.
Artificial Intelligence · Computer Vision
Build an image search engine
Find visually similar images by comparing feature vectors from a pretrained network.
Artificial Intelligence · Computer Vision
Count vehicles in a video with computer vision
Use a pretrained detector to count vehicles crossing a line in traffic footage.
Artificial Intelligence · Computer Vision
Recognise hand gestures to control slides
Detect hand landmarks from a webcam and map gestures to next and previous slide.
Same tools
Software Development + Applied Mathematics
Create a school timetable generator
Generate a clash-free timetable from teachers, subjects, rooms and periods.
Cybersecurity · Web Security
Find and fix vulnerabilities in a deliberately weak web app
Run a training app on your own computer, find the classic flaws, then write the fixed code.
Development and Applied Economics · Econometrics
Test whether education and earnings are related
Use a public survey dataset to estimate how much an extra year of schooling relates to earnings.
Artificial Intelligence · Recommendation Systems
Build a movie recommender
Recommend films from the public MovieLens ratings using item similarity.
Two-subject alternatives
Computer Engineering + Systems and Cloud
Design a simple processor in a logic simulator
Build an 8-bit processor with an ALU, registers and a control unit, and run a small program on it.
Agricultural Technology + Robotics, IoT and Hardware
Build a low-cost weather and soil station for a farm
Log temperature, humidity, soil moisture and rainfall in a field and turn the data into simple advice.
Electrical Engineering + Robotics, IoT and Hardware
Build a PID temperature controller
Hold a small heater at a set temperature and tune the controller from its step response.
Programming + Systems and Cloud
Build a tiny key-value store in Rust
Write a persistent key-value database with get, set and delete and an append-only log.
Same difficulty and budget
Software Development · Mobile Apps
Develop an attendance app with QR codes
Teachers show a changing QR code; students scan it to mark attendance, with safeguards against proxies.
Data · Data Engineering
Build a data pipeline that runs every day
Fetch a public data feed daily, clean it, store it in a database and chart the history.
Artificial Intelligence + Media and Journalism
Detect fake news headlines with machine learning
Train a classifier on a public dataset and show why it learns the source's style, not the truth.
Games, Graphics and New Tech + Pure Mathematics
Render a 3D scene with a ray tracer
Write a program that draws spheres with shadows and reflections by tracing rays.