Estimate of Rice Lodging Based on Image Processing (Part 2)

—Effect Verification of Convolutional Neural Network and Learning to Rank—

Abstract

In this study, an image processing system is developed for application on rice plants to determine their status with regard to lodging, which is a contributing factor to declining harvest efficiency and rice grain quality. Test images were obtained from photographs of a rice field captured by CCD cameras placed inside cabins, immediately prior to harvesting. A total of 2,601 images were collected for training and evaluation. A convolutional neural network and an order prediction model using the learning to rank algorithm, RankNet, were used. The normalized discounted cumulative gain (nDCG) was used to assess the quality of the model. The image processing results yielded nDCG@101 was 1.0 and showed that the system was usable as a lodging diagnostic tool.