Drone Plant Stand Counting and Gap Mapping for Smallholder Maize and Sugarcane Fields

The problem: knowing how many plants came up

Poor germination is a common, expensive problem for smallholders. Bad seed, uneven sowing, waterlogging or pests can leave patches of a maize or sugarcane field empty. If the farmer knows early, he can resow or fill gaps; a few weeks later it is too late. Walking every row of a field to count is slow, so most farmers judge by eye, and seed companies and contract farming firms rarely get real numbers.

Small consumer drones now cost less than many farm machines and are used by agri service providers in Pakistan and India. Plant counting from drone images is a well-studied computer vision task but is rarely tested on small, irregular smallholder plots.

What already exists and the gap

• Pix4Dfields and similar commercial tools offer stand counting from drone images, aimed at large commercial farms.
• Drone spraying services are expanding in South Asia, so drones and pilots already exist in many farming areas.
• Research models count plants in clean, regularly planted research plots.

The gap is an open, tested pipeline for small, irregular fields with mixed weeds, shadows and broadcast sowing, that produces something a farmer can act on: a gap map.

Inspiration

TasselNet by Hao Lu and colleagues (Plant Methods, 2017) counted maize tassels from field images using a local counting regression approach rather than detecting each object, which copes better with overlap. Your project applies the same counting ideas earlier in the season to seedlings and compares counting by density with counting by detection.

What you build

• Farmers and agri service providers get a field report: estimated plants per acre and a map of gaps.
• Seed and contract farming companies can compare germination across fields.
• Researchers get an open dataset of annotated smallholder field images, if the farmers agree.

The technically interesting part

• Data collection is half the project. Plan flight height, overlap, time of day and sample plots before you fly. Bad images cannot be fixed later.
• Detection versus density. Young plants are small and close together. Compare a detector such as YOLO with a density map model and a classic colour-threshold baseline.
• From pixels to acres. Use georeferenced images so counts per area are correct, and report uncertainty.
• Generalising across fields. Train on one field and test on another to see how much the model depends on soil colour and light.

How to evaluate it

• Mean absolute error and percentage error of counts in hand-counted sample plots.
• Cross-field test: error when the model sees a field it was not trained on.
• Gap map accuracy: overlap between predicted gaps and gaps marked on the ground.
• Processing time for one field on a laptop.

Suggested plan

• Weeks 1 to 4: permissions, flight protocol, first flights and hand counts.
• Weeks 5 to 10: stitching or tiling, annotation and baseline counting.
• Weeks 11 to 18: deep learning models and the gap map.
• Weeks 19 to 24: second field, cross-field tests and the report generator.
• Weeks 25 to 28: evaluation, write-up and demo.

Path to a product

Drone service providers can add stand counting to spraying visits, and seed companies and contract farming firms want germination data. An open pipeline that works on smallholder fields is a credible starting point for a service business, and the project is a strong computer vision portfolio piece.

Scope warning

Drone flights are regulated. Check your country's current rules with the civil aviation authority, get written permission from the field owner and your university, and fly only with a trained pilot. If you cannot fly, use public drone datasets and say so. Keep to one or two crops.

Learn what you need on Sefism

• Machine learning roadmap: model training and evaluation.
• Data science roadmap: sampling and estimating counts with uncertainty.
• Ultralytics YOLO documentation: training a detector.
• OpenCV Python tutorials: image processing baselines.
• Streamlit documentation: the field report viewer.

Get help with this idea

The biggest risk is flying late in the season and running out of time for ground truth. Tauseef can help you plan data collection and choose between detection and density models. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session