Estimation of Rice Lodging Using Semantic Segmentation Based on Deep Learning Model

Abstract

Rice lodging is an annual occurrence that causes enormous damage to rice production by typhoons and rainy seasons. Therefore, it is necessary to find an effective method to prevent the damage to rice yield and pre-harvest sprouting through early detection. This paper proposes an estimation method for rice lodging based on RGB images captured by unmanned aviation vehicles. The proposed method constructs the DeepLabV3+ semantic segmentation model based on ResNetV2 101 backbone network and estimates the area of lodging, non-lodging, and background. To train and evaluate the proposed model, we captured 816 images related to rice lodging using an unmanned aerial vehicle in Gyeongsang-do, Jeolla-do, and Chungcheong-do. The collected dataset was divided into 748 training data, 40 validation data, and 28 evaluation data, which were then used in various methods such as transfer learning and focal loss function for the improved estimation performance. The evaluation of performance using 28 evaluation data shows that the DeepLab V3+ semantic segmentation model, to which the focal loss function was applied, yields the best results with 93.16% pixel accuracy and 87.75% mIoU. Furthermore, the estimation result can be used to find the distribution of lodging and non-lodging and analyze the trend of spreading lodging, damage, and shape.