SSM-Net for Plants Disease Identification in Low Data Regime

TL;DR

This paper introduces SSM-Net, a novel few-shot learning architecture using stacked siamese and matching networks, to improve plant disease detection accuracy in low-data scenarios, outperforming traditional transfer learning methods.

Contribution

The paper proposes a new metrics-based few-shot learning model, SSM-Net, specifically designed for plant disease detection with limited data, demonstrating superior accuracy over existing approaches.

Findings

Achieved 92.7% accuracy on mini-leaves dataset.

Achieved 94.3% accuracy on sugarcane dataset.

Improved decision boundaries compared to VGG16 transfer learning.

Abstract

Plant disease detection is an essential factor in increasing agricultural production. Due to the difficulty of disease detection, farmers spray various pesticides on their crops to protect them, causing great harm to crop growth and food standards. Deep learning can offer critical aid in detecting such diseases. However, it is highly inconvenient to collect a large volume of data on all forms of the diseases afflicting a specific plant species. In this paper, we propose a new metrics-based few-shot learning SSM net architecture, which consists of stacked siamese and matching network components to address the problem of disease detection in low data regimes. We demonstrated our experiments on two datasets: mini-leaves diseases and sugarcane diseases dataset. We have showcased that the SSM-Net approach can achieve better decision boundaries with an accuracy of 92.7% on the mini-leaves dataset and 94.3% on the sugarcane dataset. The accuracy increased by ~10% and ~5% respectively, compared to the widely used VGG16 transfer learning approach. Furthermore, we attained F1 score of 0.90 using SSM Net on the sugarcane dataset and 0.91 on the mini-leaves dataset. Our code implementation is available on Github: https://github.com/shruti-jadon/PlantsDiseaseDetection.