Residual U-Net and SAM 3 fusion maps farmland with 92% Dice accuracy

A new study presents a robust workflow for mapping farmland extent and visible boundaries using 1-meter NAIP RGB imagery, combining a residual U-Net (ResUNet) with a frozen, text-prompted SAM 3 model. The researchers annotated 37 scenes covering diverse agricultural landscapes—open cropland, peri-urban interfaces, semi-arid irrigation, and fragmented mosaics—and extracted 5,698 non-overlapping 256x256 patches. The ResUNet, trained with a Dice-dominant hybrid loss (L = 2.5(1 - Dice) + BCE), achieved strong performance: test accuracy 0.8808, IoU 0.8605, Dice 0.9234, precision 0.8766, and recall 0.9794.

The key innovation is fusing the ResUNet output with a frozen SAM 3 branch prompted with the text "agricultural farmland field" via a logical OR operation. While the stand-alone ResUNet performed well, SAM 3 refinement dramatically improved results on challenging subregions—orchard rows saw Dice climb from 0.858 to 0.955, and fragmented parcels improved from 0.804 to 0.903. Sliding-window stitching produced coherent regional masks with tile Dice scores up to 0.919. The resulting semantic layer focuses on farmland extent rather than cadastral parcel boundaries, making it ideal for agricultural monitoring where proprietary or updated field maps are missing.