Benchmarking U-Net and Transformer-Based Architectures for Paddy Field Semantic Segmentation Using Very High-Resolution Satellite Imagery in Indonesia

Accurate mapping of paddy fields at the phenological stage level is critical for food security monitoring in Indonesia, yet annotated very high resolution (VHR) benchmark datasets for tropical agricultural environments remain scarce. This letter presents a multisite benchmark evaluation of four deep learning semantic segmentation architectures—U-Net, Vision Transformer (ViT), TransUNet, and UNetFormer—on three newly compiled VHR paddy datasets from Pleiades/Pleiades Neo imagery over Subang, Klaten, and Yogyakarta, Indonesia, covering binary, three-class, and four-class scenarios. Binary-task results, reported as mean ± standard deviation (SD) over three seeds, show U-Net, TransUNet, and UNetFormer perform comparably at all sites (overall accuracy (OA) within 0.02 everywhere), revising an earlier single-seed impression of a clear per-site winner; a controlled ablation confirms ImageNet pretraining gives a consistent but modest benefit (1.5–4.5-pp OA), smallest at the site with the highest class count and most severe imbalance. Extending evaluation beyond the binary task, U-Net and TransUNet are competitive with, and on Yogyakarta’s three-class task exceed, UNetFormer, whose advantage instead concentrates in the most severe four-class setting, where U-Net and TransUNet both collapse (OA = 0.33) while UNetFormer remains functional (OA = 0.74). The from-scratch ViT configuration collapses to trivial predictions on all datasets, indicating that this specific Transformer design is not viable under these small-sample conditions. Spatial resolution further matters. 30-cm/pixel imagery consistently outperforms 50-cm/pixel imagery for phenological stage discrimination. The presented datasets and results provide a reference foundation for future deep learning research on Indonesian agricultural remote sensing.