Comparison of U-Net and SAM-LoRA Deep-Learning Models for Paddy-Field Detection in High-Resolution Imagery of Poncosari Village, Bantul Regency, in 2023 Using ArcGIS

Abstract

The object of paddy field cover in Indonesia in high-resolution images is difficult to identify due to its unique characteristics and similarity to surrounding objects. This study aims to compare the performance of two Deep Learning architectures, UNet with ResNet34 backbone and SAMLoRA with ViT-B backbone, to perform segmentation using high-resolution images with accuracy down to the paddy field unit. Both models were trained using identical hyperparameters, with and without data augmentation, to evaluate accuracy and computational efficiency. The results show that SAMLoRA outperformed UNet across all evaluation metrics, achieving a higher mean Intersection over Union (mIoU) of 0.7909 compared to 0.7568. Although SAMLoRA required a longer training time, it produced clearer and more interpretable segmentations, capturing complex details more effectively. In conclusion, both Deep Learning models offer a much faster and more effective alternative to manual digitization for mapping paddy fields.