A reproducible pipeline for generating analysis-ready paddy-field segmentation datasets from cadastral vectors and Pléiades very-high-resolution imagery.

This method converts official Lahan Baku Sawah cadastral vectors and Pléiades very-high-resolution imagery into analysis-ready paddy-field segmentation tiles. It aligns coordinate reference systems, rasterizes polygons onto the reference image grid, tiles paired image and mask arrays using 1024 × 1024 windows with a 512-pixel stride, applies per-band 2nd-98th percentile normalization while excluding no-data pixels, removes tiles below 1% paddy coverage, and computes connected-component landscape descriptors. Applied to five Indonesian regions, the workflow produced 3108 image-mask pairs spanning a wide range of Largest Patch Index values (9.40-34.43%). An exploratory validation used a separate 48-tile metadata subset: a custom U-Net was trained on two tiles per region and evaluated on the remaining 38 tiles, while SAM 1 was evaluated with a ground-truth-derived bounding-box prompt. The results confirm that the exported PNG pairs can be consumed by standard segmentation pipelines, but the small, overlapping validation subset is not intended as a benchmark. The source TIFFs are publicly available on Figshare, and the processed dataset is available on Mendeley Data, Version 3. • Produces co-registered 1024 × 1024 RGB image and binary-mask tiles from cadastral vectors and very-high-resolution imagery. Handles coordinate-system mismatch, no-data borders, overlapping windows, coverage filtering, and normalization.