Wheat Crop Monitoring and District Yield Estimates from Sentinel-2 for Punjab

The problem: wheat decisions are made before the harvest is known

Wheat is Pakistan's staple food, and Punjab grows most of it. Every spring, governments decide on procurement targets, support prices and imports before the harvest is fully known. Flour mills, traders and farmers' groups make their own decisions on the same uncertainty. When estimates are late or wrong, the result shows up in flour prices and in farmers selling below cost.

Official crop estimates rely heavily on field surveys and reporting staff. Satellites cannot replace that work, but they can give a consistent, early, field-by-field view of the whole province every few days, for free.

What already exists and the gap

• Provincial crop reporting services and statistics departments publish district area and production figures, which you will use as ground truth.
• The GEOGLAM Crop Monitor gives monthly, expert-assessed crop conditions for major producing countries, at a coarse national or regional level.
• Commercial agritech firms sell field monitoring dashboards to large farms and lenders, usually without published accuracy for smallholder landscapes.

The gap is an open, documented pipeline for Punjab smallholder wheat that maps the crop, tracks the season against past years and gives district yield estimates with honest error bars.

Inspiration

Lobell and colleagues' paper "A scalable satellite-based crop yield mapper" (Remote Sensing of Environment, 2015) introduced SCYM, which uses crop simulation models to learn the link between satellite observations and yield, so yields can be estimated without ground data in every field. It was developed for maize and soybean in the US Corn Belt. Applying a simpler, statistical version of the idea to small wheat fields in Punjab, and measuring where it breaks, is a real research contribution.

What you build

• The mapping stage builds a cloud-masked Sentinel-2 time series from November to April and classifies wheat from its distinctive green-up and senescence curve.
• The monitoring stage computes season metrics per field and summarises them per district.
• The yield stage trains a model on several past seasons, linking district-level satellite and weather features to published yields.
• The dashboard shows each district's current season against previous ones and the yield estimate with its error range.

The technically interesting part

• Small fields and mixed pixels. Many Punjab fields are narrow, so 10 metre pixels mix wheat with field edges, trees and other crops. Measure accuracy by field size and report it.
• Cloud and haze. Winter fog and haze hide the ground for weeks in some seasons. Use cloud masks carefully and interpolate the time series, then test how missing data affects the estimates.
• Labels. You need labelled wheat and non-wheat points. Combine a field visit with geotagged photos and careful labelling on high resolution imagery. Document how many labels you have and how they were checked.
• Few training years. With only a handful of seasons since Sentinel-2 began, overfitting is a real risk. Use leave-one-season-out validation, keep features few and physically meaningful (peak greenness, season length, heat days), and compare against a naive baseline such as the previous five-year average.
• Weather and heat. Late heat during grain filling cuts wheat yields. Add temperature features from reanalysis data and test whether they improve estimates.

How to evaluate it

• Mapping: overall accuracy, F1 for wheat and wheat area per district compared with published area figures.
• Yield: mean absolute percentage error per district across held-out seasons, compared with the five-year average baseline.
• Timeliness: how the error changes when the estimate is made in February, March or April.
• Robustness: performance in districts with the smallest fields and the most cloud.

Suggested plan

• Weeks 1 to 4: learn Earth Engine and Sentinel-2, build cloud-free time series for one district.
• Weeks 5 to 9: labels, wheat classifier, accuracy assessment.
• Weeks 10 to 14: extend to more districts, season metrics, collect published yields.
• Second semester: yield model and validation, weather features, dashboard, write-up.

Path to a product

Banks and microfinance lenders exposed to agriculture, crop insurers, flour mills and provincial agriculture departments need early, consistent crop information. A data service with a public district dashboard and a paid field-level feed is a known model. Agricultural remote sensing is a strong research and employment niche, and the project is a natural fit for a research publication.

Scope warning

Start with three to five districts, not the whole province. Do not train deep models on a handful of seasons; simple, validated models are more convincing. Never present satellite estimates as official numbers, and keep the baseline comparison in every chart. If a stage takes longer than planned, cut districts before you cut validation.

Learn what you need on Sefism

• Machine learning roadmap: validation, overfitting and feature choice.
• Data science roadmap: the statistics behind honest error bars.
• scikit-learn User Guide: classifiers for the wheat mask and regressors for yield.
• Python Data Science Handbook: NumPy and array thinking for time series stacks.
• Pakistan Bureau of Statistics: agricultural statistics and context.

Get help with this idea

This is an advanced, research-shaped project, and the common failure is a beautiful map with no validation. Tauseef can help you set up the labelling plan, the leave-one-season-out evaluation and a realistic two-semester scope. 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