Plant and crop disease detection is one of the most popular computer vision thesis topics, and for good reason: early, accurate diagnosis helps farmers act before a disease spreads, and a smartphone camera is now available in almost every village. Deep learning models can classify leaf images with impressive accuracy in the lab. The hard part, and the part examiners and reviewers care about, is whether the model still works on a real field photo taken in harsh sunlight with soil, weeds and other leaves in the background. This guide explains how to plan a plant disease detection study that is honest, reproducible and publishable.
1. Define the task precisely
"Plant disease detection" can mean several different computer vision problems. Decide early which one your thesis solves:
- Image classification: one label per image (for example "tomato early blight" or "healthy"). This is the most common starting point.
- Object detection: locating diseased leaves or lesions with bounding boxes, useful when one photo contains many leaves or plants.
- Segmentation: outlining the exact lesion area, which allows severity estimation (what fraction of the leaf is affected).
- Severity grading or early detection: predicting disease stage, which is more useful agronomically but needs careful expert labels.
Also fix the scope: one crop with several diseases (for example wheat rust, cotton leaf curl or rice blast, depending on your region) is often more valuable and realistic than a generic multi-crop classifier.
2. Choose datasets and understand domain shift
The best-known public dataset is PlantVillage, with more than 50,000 leaf images across 38 crop and disease classes. It is excellent for learning and benchmarking, but most images show a single detached leaf on a plain background under controlled lighting. Models trained only on such images often lose much of their accuracy on field photos. This gap between lab and field conditions is called domain shift, and it is the central issue of the topic.
Field-oriented datasets such as PlantDoc (web-collected images with real backgrounds) and crop-specific competition datasets such as the Cassava Leaf Disease collection are closer to real use. If you can, collect your own field images from local farms with help from plant pathologists, and record the device, date, location and growth stage. A strong design trains on one source and tests on another, so you can report both in-domain and cross-domain results.
3. Baselines: CNNs and Vision Transformers
Use transfer learning from ImageNet-pretrained models and compare several families under identical splits and training budgets:
- Classic CNNs: ResNet-50 and DenseNet as standard references.
- Efficient CNNs: MobileNetV3 and EfficientNet, important if you target phones or edge devices.
- Vision Transformers: ViT, DeiT or Swin Transformer, which often need more data or stronger augmentation but can generalise well.
- Your proposed method: for example attention modules, domain adaptation, self-supervised pre-training on unlabelled field images, or a lightweight hybrid model.
Python with PyTorch and OpenCV covers this whole pipeline; see our Python guide, PyTorch guide and OpenCV guide.
4. Augmentation and avoiding leakage
Augmentation should imitate real field variation: random crops and rotations, brightness and colour jitter, blur, partial occlusion and background replacement. Two common mistakes inflate results. First, near-duplicate images of the same leaf or plant appearing in both training and test sets; split by plant, field or capture session rather than by individual image. Second, models learning the background instead of the disease. Grad-CAM or similar explanation maps help you check that the model looks at the lesions, not the table or soil.
5. Validate in the field
A thesis that tests only on a random split of PlantVillage will be questioned. Add at least one of these: an independent field test set, a cross-dataset test (train on one dataset, test on another), or a small pilot with images captured by farmers or extension workers. Report where the model fails, for example early-stage symptoms, nutrient deficiency that looks like disease, or images with multiple diseases.
6. Report the right metrics
- Accuracy alone is misleading when classes are imbalanced. Report per-class precision, recall and F1, macro-averaged F1 and a confusion matrix.
- For detection, report mAP at the IoU thresholds you used; for segmentation, Dice or IoU.
- Repeat training with several random seeds and report mean and standard deviation.
- For deployment, report model size, parameter count and inference time on the actual target device.
7. Edge deployment
If your contribution is a practical tool, convert the trained model to TensorFlow Lite, ONNX or a similar format, apply quantisation or pruning, and measure the accuracy drop. Offline inference on a mid-range Android phone is a realistic target for rural areas with weak internet.
8. Concrete research questions
- "How much does accuracy drop when CNN and ViT models trained on PlantVillage are tested on field images, and can background augmentation or domain adaptation reduce this gap?"
- "Can a quantised lightweight model detect wheat rust on field images on a smartphone with accuracy close to a full-size model?"
Novelty does not come from training one more network. Build a literature comparison table of recent studies on the same crop and setting, identify a clear gap, and prove through baselines, cross-domain tests and ablation studies that your method actually closes it.
Quick checklist
- Task type, crop and disease classes clearly defined.
- Lab and field data described, with splits by plant or session.
- CNN, efficient CNN and ViT baselines under the same settings.
- Cross-domain or field validation, not only a random split.
- Per-class metrics, multiple seeds and device-level speed if deploying.
Planning a plant disease detection or smart agriculture thesis?
If you have your own research idea in computer vision for agriculture, the FHJCIS team works with you on literature review, dataset planning, experiment design, Python implementation, data analysis and writing, and helps you prepare the work for journal submission. We guide you through every step, and every contributor is named as an author. We support researchers across Pakistan, including Multan, and abroad, including Indonesia and Malaysia. WhatsApp: +92 331 2663342, +92 312 6243342 (main) or +92 332 1791155 (broadcast).
Research SupportData AnalysisWhatsApp usContact