Pest Density Classification using MobileNet Architecture

Managing the pests is the biggest challenge in the field of agriculture across the world. Timely identification of pests affecting the crops helps to prevent the further damage and thus by saving the yield of the crop. Pest-trappers are one of the organic approaches to control them. Pest-trappers containing Pheromone loaded Nanogels are placed across the fields, and the pests gets attracted to this and gets trapped. The quantity of pests being trapped will give an idea about the density of pests in that area and helps the farmer to take further remedial actions. Hence, it is an important task to classify the pest-density automatically. In this paper, we present a novel approach to classify the density of pests being trapped on pest-trappers. As the algorithm built must be deployed into a handy devices like mobile phones, a lightweight MobileNet architecture is being used to build the model. The proposed algorithm reconfigures the pre-trained model MobileNetV2 to suit the need of the problem in question. As the pest-density image dataset created by GR Agri – Tek lab is small in size with 534 images, the image augmentation technique is used. The regularization techniques like early-stopping and drop-out layers are used to minimize overfitting problem. The proposed model is trained on a GPU based system for faster convergence. The classification model is designed to classify five different density classes, and it has the accuracy of 92.59%. The weights learnt by the best possible model is then saved in the HDF5 format for the deployment purposes. In the practical environment, the pest-trapper image is sent from the fields to the server and the density is classified using the model built earlier. The result is displayed through the android mobile app to the farmer, to facilitate him to take further course of action.