Large scale mapping of winter cereals was often constrained by the
scarcity of ground truth labels and the limited transferability of
existing global products. However, an annually updated winter cereal
map for Australia was not available, and glob...
Large scale mapping of winter cereals was often constrained by the
scarcity of ground truth labels and the limited transferability of
existing global products. However, an annually updated winter cereal
map for Australia was not available, and global layers such as ESA
WorldCereal and the Winter Triticeae Crops Index showed area
mismatches and regional bias over Australian croplands. We
developed a two-stage PL based pipeline that mapped winter cereals
across Western Australia without local ground truth, using
Queensland as the reference region. First, using the 2021 wintercereal
labels in Queensland and Sentinel-2 composites, we identified
a two-dimensional vegetation-index feature space in which winter
cereals were robustly separated from other land-cover types.
Among many candidates, the CIRE–LSWI pair provided the most
stable separability across major cropland regions in Queensland. We
trained a topology-guided U-Net on seed pixels sampled from highdensity
regions of this feature space and applied it to Western
Australia Sentinel-2 composites to generate high-confidence
winter-cereal PLs for 2021. We defined high-confidence PLs using
the intra-QLD criterion that achieved ≥90% precision. Second, we
processed Sentinel-2 time-series data from sowing to pre-harvest,
trained a Random Forest classifier, and generated annual 30 m
winter-cereal probability maps for Western Australia (2020–2025).
We calibrated a separate probability threshold for each year to match
Western Australia government area statistics. Using each year’s
calibrated threshold, the 2020 map achieved R² = 0.94 and RMSE =
634.8 km² against SA2-level government area statistics (n = 25), iv
and the 2023 map achieved OA = 87.0%, UA = 100%, PA = 80.5%,
and F1 = 0.892 against Google Street View–derived polygons. These
results demonstrated the feasibility of winter-cereal mapping via
spatiotemporal transfer in large regions where ground truth
collection was constrained.