Semantic Segmentation of Rice Field Bund on Unmanned Aerial Vehicle Image using UNet

This research contributes to the application of image processing in the optimization and automation of agricultural drones to modernize agricultural systems. This study aims to perform semantic segmentation of rice field bunds using UNet. The method proposed in this study starts from the dataset finalization stage, data preparation, and finally evaluates the segmentation model. The input from the UNet model will be an RGB image of a wet rice field with a resolution of 512 × 512 pixels which is the result of the frame extraction process and the output is a binary image of the predicted pixel of the rice field bund. Based on the results of the testing model that has been carried out on computers with Intel i7 processor, RTX 3070 Ti GPU, 32GB RAM, and 1 TB SSD, the UNet model underwent training with a dataset configuration of type split 1 (792 images of train set, 44 images of validation set, and 45 images of test set) and without going through the binarization stage has the highest performance with an average accuracy of 99% and an average segmentation time of 0.32 per second.