UnMixMamba splits four crops out of coarse MODIS pixels, lifting barley recall to 0.55 in Alberta at roughly a thousand times the compute
Synopsis
A MODIS pixel often covers one or more Prairie fields, so the coarse signal is a mixture of crop types and direct coarse-to-fine classifiers struggle. UnMixMamba uses the abundance field from unmixing as a shared intermediate, so reconstruction, classification and sparse refinement all read the same fractions, while a learnable per-crop variant bank absorbs intra-class spectral spread. On 183 spatially blocked Alberta 2015 test patches it reaches OA 0.8397, mIoU 0.6640, AA 0.7837 and Kappa 0.7715, ahead of ST-DRes and other baselines, with the best barley recall in the comparison at 0.5496; the task is closed-set four-crop mapping inside a known agricultural mask.
Interpretation
Crop mapping no longer maps coarse MODIS pixels directly to high-resolution classes; it first estimates each pixel's crop fractions and lets those fractions drive the high-resolution decision. Classical subpixel mapping estimated abundances and then allocated hard labels with hand-designed spatial or temporal priors, while deep models often mapped the coarse cube straight to high-resolution classes and rarely required reconstructed spectra and abundances to agree across scales. On the 183 spatially blocked Alberta 2015 test patches, UnMixMamba reaches OA 0.8397, mIoU 0.6640, AA 0.7837 and Kappa 0.7715, above the dedicated MODIS crop subpixel mapping baseline ST-DRes at OA 0.7970 and mIoU 0.5662, and above FineCrop, AgriFM and five classical allocators.
The network models the temporal, spectral and spatial axes separately and fuses them into per-pixel crop fractions, instead of flattening the whole MODIS series into one convolution. The MODIS cube consists of 16-day composites with six bands per date; flattening discards phenological order, and mixed pixels concentrate along field boundaries. Temporal, spectral and depthwise spatial convolutions run in parallel and feed an abundance head that outputs normalized fractions; replacing this decoupled module with a shallow low-resolution encoder drops OA to 0.6181 and barley recall to 0.0000.
High-resolution refinement runs at only a small subset of locations, chosen by predicted crop fractions rather than spread uniformly over the high-resolution grid. Uniform refinement treats every subpixel alike, so most compute lands in easy field interiors while the error-heavy minority class gets no extra capacity. A fixed token budget is shared by abundance mass and area, with a floor for barley and canola; at the default budget refinement covers about 11% of the high-resolution grid, and tightening the budget further lowers OA by 8.6 percentage points and barley recall to 0.0000.
Perspective
Evaluation is limited to closed-set four-crop mapping inside a known agricultural mask for Alberta 2015, with non-crop and minor-crop pixels excluded from training and evaluation and metrics assuming that mask is available at inference; the cross-province check applies Alberta-trained weights directly to Saskatchewan 2015.
Barley recall is still only 0.55, so almost half of barley pixels are misclassified, mostly as spring wheat or canola; across multi-seed retraining overall accuracy is stable, but barley recall ranges from 0.493 to 0.606 and trades off mainly against canola. All experiments use the single 2015 crop year, with training in Alberta and testing in Alberta and Saskatchewan, and other years have not been tested.