Technology
July 5, 202610 min readHow AI is changing crop disease diagnosis
From lab to field: how machine learning helps farmers identify problems faster.
Daniel Ochieng
Head of AI, Krawp
The old way of diagnosing crop problems
For most of agricultural history, crop diagnosis has depended on a slow, expensive chain of human expertise. A farmer notices something wrong in the field — discoloured leaves, unusual spots, stunted growth — and carries that observation (sometimes literally carrying a plant sample) to the nearest source of expertise: an extension officer, an agro-dealer, or a university researcher.
In well-resourced systems, the extension officer may consult reference books, contact a plant pathologist, or arrange for laboratory analysis — processes that can take days to weeks and cost more than a smallholder farmer earns in a season. In practice, most smallholder farmers don't have access to any of these resources. They rely on their own experience, advice from neighbours, or whatever product the nearest agro-dealer recommends — regardless of whether it's appropriate for the actual problem.
The consequences of this system are measured in avoidable losses. Studies across sub-Saharan Africa consistently estimate that 20–30% of crop yields are lost to pests and diseases that could be managed effectively if diagnosed correctly and treated promptly. For a farmer earning $1–2 per day, a misdiagnosed disease isn't an academic problem — it's a food security crisis.
What machine learning does differently
Krawp's AI models are trained on over 280,000 annotated crop images spanning 38 crop species, 25 disease types, 12 nutrient deficiency categories, and 8 pest damage patterns. Each image was collected in real field conditions — taken by farmers or field officers on standard smartphones in variable lighting, at different angles, and with the kind of image noise that characterises real-world photography rather than laboratory specimens.
The models learn to identify visual features that correlate with specific problems: lesion shape and margin characteristics, colour gradients and their spatial distribution, the pattern of affected tissue relative to healthy tissue, and contextual cues like affected leaf age, position on the plant, and growth stage. These features are often too subtle for unaided human observation — a slight difference in lesion margin, a specific shade of yellow in a chlorotic band — but they're consistent and learnable by neural networks.
Critically, image recognition is only part of the system. Krawp combines visual analysis with environmental data (temperature, humidity, rainfall from the past 7 days), regional disease pressure history (what other farmers in the area are reporting), crop growth stage, and soil conditions to produce a context-aware diagnosis. A lesion that might look like early blight in one climate context could be bacterial speck in another — Krawp accounts for these variables.
Accuracy that holds up in the field
There's a well-known gap between laboratory accuracy and field accuracy in machine learning. A model that performs at 96% accuracy in controlled conditions — standardised lighting, high-resolution images, clean backgrounds — often drops to 70–75% when facing dirty lenses, poor lighting, motion blur, and overlapping symptoms. This 'domain gap' is the central challenge in agricultural AI.
Krawp's models were specifically designed to bridge this gap. The training dataset consists entirely of field-grade images: photos taken by farmers on mid-range smartphones (typical devices cost $80–$150), in direct sunlight or shade, at angles that reflect how farmers actually hold their phones when photographing crops. The models were trained to perform well on this data — not on laboratory-grade images that farmers will never take.
In field trials across 14 sites in East and West Africa — covering 2,347 individual crop scans across 23 disease and deficiency categories — Krawp achieved an overall diagnostic accuracy of 87%. For the five most common diseases in the regions tested (late blight, early blight, bacterial wilt, leaf rust, and rice blast), accuracy exceeded 91%. These results are comparable to the diagnostic accuracy of experienced plant pathologists working with reference materials.
How the model learns and improves
Krawp's AI models are not static. Every diagnosis that a farmer confirms or corrects feeds back into the training pipeline. When a farmer receives a diagnosis of 'late blight' and confirms it, that image-and-label pair reinforces the model's understanding. When a farmer corrects a diagnosis — indicating that the actual problem was bacterial speck, not early blight — that correction is flagged for model review and retraining.
This continuous learning loop means the models improve with every scan. As the dataset grows — now exceeding 1.2 million confirmed field observations — the models become better at handling edge cases, regional variations, and unusual symptom presentations that weren't well-represented in the initial training data.
The retraining process follows strict validation protocols. Before any updated model is deployed, it must demonstrate statistically significant improvement over the previous version on a held-out validation set that includes challenging cases. Models that perform well on average but degrade on specific crop-disease combinations are rejected and retrained with adjusted weighting.
What AI doesn't replace
Krawp isn't designed to replace agricultural expertise. It's designed to extend it — to bring the diagnostic capability of a trained plant pathologist to every farmer's phone, everywhere, in real time. The app provides a data-driven starting point, but final decisions about treatment — accounting for local regulations, input availability, farm-level economics, and the farmer's own experience and judgement — remain with the farmer and their advisory network.
The real transformation isn't that AI can identify a disease. It's that a farmer in a remote field, with no internet, no extension officer nearby, and no laboratory within 100 kilometres, can get a reliable diagnosis in 3–5 seconds. That speed and accessibility is what changes outcomes — not the sophistication of the underlying algorithm.
As Krawp's AI capabilities continue to evolve, the focus remains on practical impact: faster diagnoses, better treatment decisions, reduced input waste, and ultimately, more food on more tables. The technology is a means to that end — nothing more, nothing less.
A farmer in a remote field, with no internet and no extension officer nearby, can get a reliable diagnosis in 3–5 seconds.
Key Takeaway
AI crop diagnosis isn't about replacing human expertise — it's about making reliable identification accessible to every farmer. Krawp's models are trained on 280,000+ field-grade images, achieve 87% accuracy across 23 diseases, and improve continuously through farmer feedback.
Frequently asked questions
Does Krawp's AI work for crops not in its training data?
Krawp currently supports 38 crop species. For crops not in the supported list, the app will indicate that it cannot provide a diagnosis and recommend consulting a local expert. The supported crop list expands with each model update as new training data is collected.
Can the AI make dangerous misdiagnoses?
Krawp provides confidence scores with every diagnosis. Low-confidence results (below 70%) are flagged with a recommendation to seek expert verification. The app also provides general safety guidance for all treatment recommendations, including warnings about withholding periods and application safety.
How does Krawp handle diseases that look very similar?
For diseases with significant visual overlap (e.g., early blight vs. bacterial speck), Krawp provides a differential diagnosis — listing the most likely candidates with confidence scores — rather than a single definitive answer. This helps the farmer and their advisor make a more informed decision.
