Testing how segmentation affects wheat-disease classification
A controlled comparison of wheat stripe-rust classification workflows under different image-segmentation conditions, supported by published research.
Case record
The same seven questions, every time.
- Context
- Wheat stripe rust appears within complex field imagery where leaves, lighting and background conditions vary between images.
- My role
- Developer and first author — built and compared the classification workflows, and wrote the resulting paper.
- Constraint
- A classifier can be affected not only by its architecture but by how the plant region is prepared beforehand, so the preparation choice had to be isolated from everything else.
- Decision
- Hold the evaluation setting explicit and constant while varying only the segmentation condition, so any difference could be interpreted within the study.
- Artefact
- Comparable classification workflows across original and segmented inputs, with conditions and evidence documented for review, published as a first-author IEEE Access paper.
- Result
- A documented comparison showing how one preparation choice affected classification within a defined dataset and protocol. The work contributed to AgriVision, recognised with the NUST Rector's Gold Medal.
- Boundary
- The study does not establish universal performance across farms, cameras or seasons.
Evidence ledger
What the work can show.
- Comparison
- Different segmentation conditionsClassification workflows were compared within the same study evaluation.
- Verification
- Published researchThe method and findings are available in the linked first-author paper.
- Boundary
- Study dataset and protocolThe result is not presented as universal field performance.
The challenge
Wheat stripe rust appears within complex field imagery where leaves, lighting and background conditions vary. A classifier can therefore be affected not only by its architecture, but by how the relevant plant region is prepared before classification.
The useful question was practical and testable: does separating the plant and disease region from the surrounding image change classification performance under the study conditions?
Testing the classification workflow
I developed and compared classification workflows using different image-segmentation conditions. The comparison kept the evaluation setting explicit so that any observed difference could be interpreted within the study rather than treated as a general claim about field performance.
Testing was used to examine a design choice, not simply to report a final score. Original and segmented inputs moved through comparable classification workflows, with the conditions and resulting evidence documented for review.
- Define the segmentation conditions before model comparison.
- Keep preprocessing and evaluation steps repeatable.
- Compare workflows within the same study protocol.
- Report the result alongside its dataset and operating limits.
Evidence and limits
The work produced a first-author IEEE Access paper and contributed to the AgriVision system recognised with the NUST Rector’s Gold Medal. These provide external evidence that the comparison was completed and communicated.
The study does not establish universal performance across farms, cameras or seasons. Its value is a documented comparison showing how a specific preparation choice affected classification within a defined dataset and protocol.
Connected evidence
Where this sits in the rest of the work.
- Published workThe first-author IEEE Access paper this comparison produced.
- Testing, in the working methodThe Testing stage this study is the worked example for.