Rice (Oryza sativa) is one of the staple foods worldwide. Paddy, the raw grain before removal of husk, is cultivated in tropical climates, mainly in Asian countries. Paddy cultivation requires consistent supervision because several diseases and pests might affect the paddy crops, leading to up to 70% yield loss. Expert supervision is usually necessary to mitigate these diseases and prevent crop loss. With the limited availability of crop protection experts, manual disease diagnosis is tedious and expensive. Thus, it is increasingly important to automate the disease identification process by leveraging computer vision-based techniques that achieved promising results in various domains.
The main objective of this competition was to develop a machine or deep learning-based model that would classify the given paddy leaf images accurately. Competition organizers provided a training dataset of 10,407 (75%) labeled images across ten classes (nine disease categories and normal leaf). Moreover, additional metadata for each image, such as the paddy variety and age was also provided. The task was to classify each paddy image in the given test dataset of 3,469 (25%) images into one of the nine disease categories or a normal leaf.
This was a simple multi-class classification problem, getting a public score more than 0.95 was pretty easy, the actual challenge was to get a score above 0.98.
Hence, I tried several already available finetuned models like EffificientNet(B0-B7), resnet50, resnet50v2, resnet101, resnet101v2, resnet150, resnet150v2 and InceptionResNetV2 along with data augmentation. Out of all the experiments InceptionResnetV2 gave the best results.
| Model | InceptionResNetV2 |
|---|---|
| #Params | 54,732,778 |
| Train Accuracy | 99.04% |
| Validation Accuracy | 99.42% |
| Public LB score | 0.9777 |
| Private LB score | 0.97926 |
- Competition page: Paddy Doctor: Paddy Disease Classification
- Pretrained models: Keras applications