[2O4-GS-13] AI application: Agriculture, forestry and fisheries

[2O4-GS-13-02] Applications of Visual Explanations for Crop Yields Prediction

〇Masaki Shimomura¹, Kazuyuki Nakamura^1,2 (1. Meiji University, 2. JST, PRESTO)

Keywords:Convolutional Neural Network, Grad-CAM, Agriculture, field images

Predicting crop yields is important to stabilize yields and revenues. For this subject, high-accuracy prediction is expected by using Convolutional Neural Network (CNN). On the other hand, it is important to estimate the appropriate sample size because getting agricultural data and training CNN models are needed long times and high costs. In this study, we focus on cucumber grown in greenhouses and we use Grad-CAM which can visualize explanations for CNN prediction to check that it can catch essential information. As a result, we got a suggestion that CNN focus on leaves around growth points. The area has been pointed out the importance. However, we used only 894 images so the visual explanations are not stable due to over-fitting. We need to get much more images and select CNN models as future works.