Cloud-based Paddy Plant Pest and Disease Identification using Enhanced Deep Metric Learning and k-NN Classification with Augmented Latent Fusion

Abstract

Identifying pests and diseases is vital for rice farming, but the lack of experts often limits it. We propose a novel method that uses deep metric learning and k-NN classifier to classify pests and diseases in paddy plants. We experimented using different distance metrics, latent dimensions, base models, and optimizers to train the neural model that produces latent representation from images. Then, we used k-NN retrieval approach to find and label similar images from the latent space. During the inference phase, we manipulated the original image using various transformations and fused them with its latent representation to enhance the quality of latent space. The results show that our method surpasses the conventional deep learning classification (softmax classifier). Specifically, our method achieved maximum accuracy of 0.920 on ResNet-50 and 0.878 on ResNet-152. In comparison, the softmax classifier only achieved maximum accuracy of 0.752 on the same modeling scheme. Our method can produce more discriminative and robust data representations for classification tasks. Moreover, latent fusion from input augmentation during inference can also improve accuracy up to 19.2%. We also deployed the best model from our experiment on serverless cloud computing, allowing users to use the platform for prediction and monitoring through GIS.