Development of a Prediction Method with Deep Learning and Unmanned Aerial Vehicle (UAV) for the Red Clover Coverage Ratio in Grass-Legume Mixed Swards

Abstract

A red clover coverage prediction method using drones and deep learning was developed. Six images of a grass-legume mixed swards were used as the training dataset and 180 images (60 test plots taken three times) as the estimation images. CNN models were created using three CNN architectures (InceptionV3, ResNet50, and VGG16), five optimization algorithms (MomentumSGD, Adagrad, RMSprop, Adadelta, and Adam), and five learning rate conditions (0.01, 0.001, 0.0001, 0.00001, and 0.000001), for a total of 75 conditions to create CNN models. From these, there was no significant difference among the CNN architectures, but VGG16 was suitable under the GPU-equipped PC environment, while InceptionV3 and ResNet50 were suitable under the general-purpose PC environment. In addition, the combined use of multiple AI models created in this study improved the accuracy of estimating the coverage of red clover compared to the previously reported method, achieving a maximum absolute error of less than 5％.