FieldVision: Image-based crop disease detection with visual explanations

FieldVision uses computer vision models to identify crop diseases from a photo. It returns ranked predictions with confidence scores and an AI focus area overlay that shows exactly which part of the image influenced the diagnosis — so you can see what the AI sees.

What is FieldVision?

FieldVision is FildraAI's image-based crop disease detection system. A farmer or extension officer uploads a photo of a crop showing visible symptoms — yellowing, lesions, spots, wilting, or discolouration — and FieldVision analyses the image using a trained computer vision model.

Rather than returning a single label, FieldVision provides a ranked list of the most likely diagnoses, each with a confidence score. It also produces an AI focus area map showing which region of the leaf or plant the model focused on — making the reasoning transparent and verifiable in the field.

Design principle

FieldVision is a diagnostic aid, not a final verdict. Results are intended to support decision-making alongside field observation, local knowledge, and guidance from extension officers — not replace them.

How it works

FieldVision processes each uploaded image through a four-stage pipeline:

Image capture

Upload a photo taken on a mobile camera or select from your gallery. FieldVision works best with close, in-focus shots of affected leaves or stems in good light.

CNN classification

A convolutional neural network trained on labelled field images processes the photo and assigns probability scores across all supported disease classes for the selected crop.

AI focus area overlay

Gradient-weighted Class Activation Mapping highlights the image regions that most influenced the top prediction — producing a heatmap you can compare against what you see in the field.

Ranked results

The top predictions are returned in order of confidence. You can then hand off the top result to FieldGuide for follow-up guidance, management discussion, and contextual interpretation.

Supported crops and diseases

In Phase 1, FieldVision supports two core crops. These were chosen to match the primary disease burden facing smallholder farmers in Zambia, Malawi, and DRC in the current season.

Maize

  • Fall Armyworm (FAW) damage
  • Gray Leaf Spot
  • Maize Streak Virus
  • Northern Leaf Blight
  • Common Rust
  • Healthy (no disease detected)

Rice

  • Rice Blast (leaf and neck)
  • Brown Spot
  • Bacterial Leaf Blight
  • Sheath Blight
  • Tungro
  • Healthy (no disease detected)

Scope note

If you submit a photo of a crop not in the supported list, FieldVision will still process the image, but results should not be treated as crop-specific expertise. The system will indicate when it is operating outside its training distribution.

Reading the results

FieldVision returns up to three ranked predictions for each image. Each prediction includes:

  • Disease name — the identified condition or "Healthy" if no disease pattern is detected
  • Confidence score — how strongly the model associates this image with that class (shown as a percentage)
  • AI focus area map — a colour overlay on the original image highlighting the region the model focused on

A high confidence score means the visual pattern in the image closely matches what the model has learned for that disease. A low or split confidence (for example, 45% for one disease and 40% for another) is a signal that the image is ambiguous and field confirmation is especially important.

How to interpret confidence

Confidence above 80% indicates a strong visual match. Between 50–80%, treat the result as a likely candidate requiring field verification. Below 50%, the image is likely ambiguous — retake the photo or consult an extension officer.

AI focus area explanations

FieldVision automatically highlights which part of the image most influenced the diagnosis result. This visual focus map helps you verify whether the AI is responding to an actual crop symptom.

The focus map uses a colour scale: warm colours (red, orange) mark the regions with the highest influence; cool colours (blue) mark regions with low influence. Use this to confirm whether the model is focusing on the actual lesion or symptom — rather than an unrelated part of the image such as soil or background.

Good sign

The heatmap highlights the discoloured area, lesion, or spot on the leaf. This means the model is responding to the actual symptom.

Use caution

The heatmap highlights background, soil, or an unaffected leaf area. The prediction may be unreliable — consider retaking the photo with the affected area more clearly framed.

Photo requirements

Image quality directly affects prediction accuracy. The model was trained on clear, close-up photos taken in natural light. The following guidelines apply:

Recommended

  • Take the photo outdoors in natural daylight
  • Fill the frame with the affected leaf or stem
  • Keep the camera steady — blur reduces accuracy
  • Use the closest leaf to the camera that shows the symptom clearly
  • If there are multiple symptoms, submit one photo per type

Avoid

  • Dark, shadowy, or flash-lit photos
  • Photos taken from far away where the leaf is small in the frame
  • Blurry or out-of-focus images
  • Heavily crumpled or folded leaves that obscure the symptom
  • Composite photos or screenshots of other images

For a detailed walkthrough with examples, see our photo guide: Taking good diagnostic photos →

Limitations

Understanding what FieldVision can and cannot do helps you use it appropriately.

FieldVision can

  • Identify visual patterns matching supported disease classes
  • Return confidence-ranked predictions from a single image
  • Show which image region drove the prediction via AI focus areas
  • Pass results to FieldGuide for follow-up guidance
  • Process photos taken on standard Android or iOS cameras

FieldVision cannot

  • Diagnose from a description alone — an image is required
  • Reliably identify diseases outside its training classes
  • Guarantee accuracy on severely degraded or mislabelled images
  • Distinguish between diseases with visually similar symptoms without field confirmation
  • Replace soil tests, laboratory analysis, or agronomist assessment

Important

Never apply pesticides or make irreversible field decisions based solely on a FieldVision result. Always confirm against observed symptoms and, when in doubt, consult a local extension officer before treatment.

What to do after diagnosis

A FieldVision result is the starting point for a decision, not the end of it. There are two natural paths from here:

Follow up with FieldGuide

Send the FieldVision result directly to FieldGuide to ask follow-up questions: what are the management options, what look-alikes should you check, and what should you do next given your location and crop stage.

Verify in the field first

If confidence is low or the AI focus area looks off (highlighting background rather than the crop symptom), confirm the symptom pattern manually — check neighbouring plants, look for secondary signs, and take a new photo before proceeding.

Open FieldVision About FieldGuide Photo guide