Andhra University Students: Planning a Computer Vision Mini Project Around Local Agriculture Images

Yes, a crop-leaf classifier built on photos from your own region is a realistic computer vision mini project for a student, and it can be done in about two weeks if you keep the scope to three or four classes on one crop. The plan below covers consent and ethics, a labeling guide, baseline targets and a scorecard you can use to judge your own work before you present it.

Why local agriculture images make a stronger mini project

Most students download a public leaf dataset, fine-tune a pretrained network, report 98% accuracy and stop. The result looks good but says little, because public datasets are usually photographed against plain backgrounds under controlled light. A model trained on them often struggles with a leaf photographed in a field, with soil, shadows, water droplets and a farmer’s thumb in the frame.

Coastal Andhra Pradesh is a good setting for a regional project because the nearby farming landscape includes paddy, groundnut, chilli, banana, coconut, cashew and vegetable plots. Pick one crop that you can actually access within a short travel distance. A project built on photos you gathered yourself shows that you handled the full pipeline: collection, cleaning, labeling, training, error analysis and honest reporting.

Keep the claim modest. A student classifier is a learning exercise and a decision-support sketch, not a diagnostic tool. It should never be described as a replacement for an agriculture extension officer or a plant pathologist.

The 14-day plan for a crop-leaf classification project

Fourteen days is short, so the schedule front-loads decisions and leaves time for error analysis. This is a suggested plan, not a prescribed one.

Data collection ethics for farm and field photographs

Photographing crops looks harmless, but a field belongs to someone, and images can reveal more than leaves. Treat the ethics step as part of the project, not an afterthought.

• Ask the owner first. Get verbal permission at minimum, and note the date and the person’s name in your log if they agree. Explain in simple Telugu or the local language what the photos are for and that the work is a student exercise.
• Do not photograph people. Frame shots to exclude faces, children and identifiable workers. If a person appears, delete the image.
• Avoid exact location exposure. Turn off GPS tagging or strip EXIF data before sharing anything. Record only a coarse area, such as mandal or village cluster, in your metadata.
• Do not enter restricted spaces. Stay out of fenced research plots, nurseries and government farms unless you have written permission from whoever manages them.
• Respect the crop. Do not pluck large numbers of leaves. Photograph leaves on the plant where possible.
• Never present a label as a diagnosis. Do not tell a farmer their crop has a disease based on your model.
• Decide on sharing in advance. If you plan to publish the dataset, tell the owners. If they are unsure, keep it private and share only code and a few non-identifying examples.

Metadata worth recording for every session

• Date, time of day and weather (sunny, overcast, after rain)
• Crop, variety if known, and growth stage
• Camera or phone model and distance from leaf
• Coarse area (not coordinates)
• Whether the owner or an agriculture officer commented on the condition

Session metadata matters because it lets you split data honestly later and explain failures such as “all disease images were taken in the evening.”

A labeling guide for healthy, nutrient-stressed and diseased leaves

Label quality will limit your model more than architecture choice. Write the guide before you label, and keep it short enough that a friend can follow it.

Choose classes you can actually tell apart

Students are rarely qualified to separate similar-looking conditions. Nutrient deficiency, pest damage and fungal infection can look alike. A safe starting set is visual categories rather than confirmed diagnoses:

1. Healthy: uniform colour, no spots, holes or curling beyond normal ageing.
2. Spotted or lesioned: distinct spots or patches covering part of the leaf surface.
3. Discoloured or yellowing: broad colour change without clear lesions.
4. Damaged or eaten: holes, torn edges or visible chewing.
5. Unclear (excluded from training): blurry, partial, or ambiguous images.

If you can get an agriculture officer, a lecturer in a plant science or agriculture department, or an experienced farmer to review a sample, note that in your report and use disease names only for the images they actually reviewed. Otherwise say “visually labeled by the project team.”

Labeling rules to write down

• One dominant leaf per image, in focus, filling roughly a third to two-thirds of the frame.
• If two classes are visible, use the more severe one, and record the rule.
• Put uncertain images in “unclear” rather than guessing.
• Label in a spreadsheet with image ID, label, labeler initials and a notes column.
• Have a second person label at least 10% of the images independently and compare the results. Disagreements show where the guide needs rewriting.

Baseline targets for a crop-leaf classifier

A target is only meaningful next to a baseline. These are suggested planning ranges, not benchmarks from any study. Your numbers will depend on your crop, class balance and image quality.

Macro-F1 matters because field data is usually imbalanced. Healthy leaves tend to be far more numerous than diseased ones, so plain accuracy can look good while the model ignores minority classes.

Splitting data so the test is honest

Random splitting leaks information when several photos come from the same plant or the same minute. Split by session, plot or plant instead. Hold out one complete session as the final test, and touch it only once.

Worked illustrative example: groundnut leaf classifier with three classes

This example is illustrative and uses invented numbers to show the reasoning. No code or tests were run for it.

A team of two students chooses groundnut and defines three classes: healthy, spotted and yellowing. Over two sessions they collect 600 usable photos. After cleaning, the counts are 300 healthy, 180 spotted and 120 yellowing. They hold out session two (about 240 images) as the test set.

The majority-class baseline predicts “healthy” every time. In an imagined test set with 50% healthy leaves, that gives 50% accuracy but zero recall for the other two classes. A fine-tuned small pretrained network might reach, say, 80% accuracy on validation yet only 68% on the held-out session. The team investigates and finds that most “yellowing” images in session one were taken at dusk, so the model partly learned lighting rather than leaf condition.

The targeted improvement on Day 12 is a short extra collection in midday light plus brightness augmentation. They report the before and after figures, state clearly that the labels were visual rather than expert-confirmed, and list the lighting bias as a limitation. That honest story is worth more in a viva than a single high accuracy figure.

Self-review scorecard for your computer vision mini project

Score each row from 0 to 2 (0 = missing, 1 = partly done, 2 = done and documented). A total of 16 or more out of 20 suggests a presentable project.

Andhra University students: fitting a crop-leaf project into your internship and project requirements

Andhra University’s home page carries a notice-board entry dated 23/03/2026 announcing an Internship Training Program for II Degree and III B.Tech students of affiliated and constituent colleges. That wording suggests internship notices can apply across different college types, but the content of the notice has not been verified here, so details such as eligibility, duration, evaluation and documents are unknown. Treat the notice as a pointer, and read the full version, which may have changed since.

The same home page lists campus facilities including an incubation hub called Aa-Hub and a NASSCOM skill centre. Whether either supports student AI or agriculture projects, offers mentoring or accepts outside work is not stated, so ask before assuming. If you want to present your classifier more widely, these may be natural places to enquire.

You can also use a short, self-paced project like this alongside the free two-week online internship at CareerSwami, which provides an offer letter and a completion certificate after successful completion. That is separate from any university requirement, and it should not be assumed to satisfy your department’s internship or project rules.

Because the university’s rules on whether an online or self-directed project can count toward an internship requirement are unverified, get them in writing before you rely on it.

• Open the 23/03/2026 notice on the university home page and note eligibility, dates, documents and any nodal officer named.
• Ask your department coordinator or training and placement officer whether online or self-arranged internships are accepted, and by which regulation.
• Email your project guide with a one-page scope and ask whether a locally collected image dataset is acceptable for a mini project.
• Ask whether data collected from private farms needs a formal permission letter or ethics clearance at your college.
• Check with the Aa-Hub or skill centre contact whether students outside a particular programme can use their facilities or GPU access.
• Confirm the report format, evaluation scheme and submission deadline for your semester in the official project guidelines.

Tooling choices that keep a two-week crop-leaf project manageable

• Annotation: a spreadsheet is enough for whole-image classification. Use a labeling tool only if you decide to add bounding boxes.
• Model: a small pretrained convolutional network fine-tuned in a free notebook environment is sufficient for a few hundred images.
• Augmentation: flips, small rotations and brightness changes are sensible; avoid heavy colour shifts because colour carries the signal for leaf condition.
• Tracking: keep a simple experiment table with run name, data version, settings and test result.
• Deployment: a demo that takes an uploaded photo and returns a label with a confidence score is enough. Add a visible disclaimer that it is a student prototype.

What to put in your report and presentation

1. The regional problem in two sentences, without exaggeration.
2. Data collection method, permissions and the ethics measures you took.
3. Class definitions and the labeling guide, with two example images per class that contain no people.
4. Baseline and model results in one table, including the held-out session.
5. Three failure examples and what you learned from them.
6. Limits and next steps, such as more seasons, more varieties or expert-verified labels.

Frequently Asked Questions

How many leaf images do I need for a crop-leaf classification project?

For three or four classes and a fine-tuned pretrained model, a few hundred cleaned images per class is a workable target, though fewer can still teach the pipeline. What matters more is variety across days, lighting and plants. A smaller, varied set is better than a large set of near-duplicates from one plant.

Can I use a public plant disease dataset alongside my own local images?

Yes, as long as you state the source and licence and keep the two separate in your evaluation. A useful design is to train on public data, then test on your local images to show how much performance drops. Check the licence terms of any dataset before reuse.

Do I need a farmer’s written consent to photograph leaves?

Verbal permission with a logged date and name is a reasonable minimum for a student exercise, but your college may require more. Ask your project guide or department whether a permission letter is expected, especially if you plan to publish the dataset.

Which metric should I report besides accuracy?

Report macro-F1 and a confusion matrix. Field datasets are usually imbalanced, so accuracy alone can hide a model that never recognises the rarer, more interesting classes.

Why should I split train and test data by session instead of randomly?

Photos from the same plant or visit look very alike, so a random split lets near-copies appear in both sets and inflates your score. Holding out a whole session or plot shows how the model behaves on genuinely new conditions.

Can I name specific diseases in my labels?

Only for images reviewed by someone qualified, such as an agriculture officer or plant science faculty. Otherwise use visual categories like spotted or yellowing and say the labels are visual, not diagnostic.

Will a 14-day crop-leaf project be enough to show on a resume?

It can be, if it includes your own collected data, baselines, error analysis and a clear limits section. Describe it precisely as a student prototype and avoid claims about real-world accuracy or farm deployment.

Can I run this as part of a free online internship?

The two-week format of the CareerSwami internship suits a scoped project like this, but whether your university will accept it toward an internship requirement must be confirmed with your department in writing.

Related Artificial Intelligence guides

• Manav Rachna University Students: A Viva Question Bank for Your AI Internship Presentation
• Gokhale Institute Students: Applying AI to Economics Data in a Short Internship
• Nrupathunga University Students: How to Pick Between Two AI Internship Offers Using a Weighted Scorecard
• All Artificial Intelligence internship guides