Visual Model Architecture Benchmarking for Robust Crop Disease Classification
Overview
I designed and implemented a unified evaluation framework to systematically compare CNNs, contrastive Vision–Language Models, and generative Vision–Language Models for fine-grained crop disease classification under real-world domain shift.
The goal was not just accuracy under controlled conditions, but robustness across lab and field environments.
Problem
Many computer vision systems for agriculture perform well on curated lab images but degrade significantly when deployed in uncontrolled field conditions.
This project evaluated multiple architectural paradigms under explicit Lab vs Field separation to understand performance trade-offs and robustness characteristics.
Dataset
AgriPath-LF16
111,307 images
16 crops
41 diseases
65 crop–disease pairs
Explicit Lab vs Field source separation
Examples of images of a tomato crop with bacterial spot in AgriPath-LF16. Image (a) is a lab-based image, and (b) is a field-based image