Paired evaluation of U-Net with ResNet backbones for oil palm segmentation from Sentinel-2 RGB imagery

This study compared U-Net models with ResNet-18, ResNet-34, ResNet-50, and ResNet-101 encoders for oil-palm segmentation using 10 m Sentinel-2 RGB true-colour imagery in Riau, Indonesia. A pooled multi-year dataset comprised 1,283 image–mask patches: 1,026 for training and 257 for patch-level testing. Sixty matched configurations produced 240 model fits. Mean intersection over union (IoU) ranged from 0.789 to 0.812, with no significant overall backbone effect (χ2(3) = 1.858, p = 0.602). ResNet-18 exhibited the lowest IoU coefficient of variation (18.3%) while requiring approximately 72% fewer parameters and 77% fewer floating-point operations than ResNet-101. Increasing encoder depth did not consistently improve performance, whereas ResNet-18 offered a favourable observed accuracy–efficiency trade-off. These findings do not establish statistical equivalence or generalisation beyond the evaluated RGB input, training configurations, reference masks, and patch-level test set.