Revisiting decoder-side improvements in Landsat-8 cloud detection: a controlled multi-seed and multi-target cross-dataset evaluation

Accurate cloud masking is a prerequisite for most optical remote-sensing applications. Reported 1–3% improvements from decoder-side modifications (attention, auxiliary supervision, hybrid losses) in cloud detection are commonly based on single-run evaluations, making it difficult to determine whether they exceed training variation. We revisit these components on Landsat-8 imagery under a controlled protocol: 96 training runs (60 ablation, 12 configurations × five seeds; 21 architecture/encoder, seven models × three seeds; and 15 parameter-matched control, three blocks × five seeds), evaluated in-domain on 38-Cloud and cross-dataset on two independent same-sensor targets (78 SPARCS scenes and 26 829 tiles from 96 L8 Biome scenes). In-domain, no component produced a statistically significant gain: all 12 configurations fell within ±0.41% of the baseline; the only two significant paired differences were negative. Cross-dataset behaviour diverged by component type. Attention-free interventions transferred with negligible change (SPARCS −0.26% to −0.63%; L8 Biome +0.03% to −1.13%), whereas all eight configurations containing CBAM-family attention modules degraded systematically on both targets (complete 5/5 seed-level separation in 6/8 SPARCS and 8/8 L8 Biome configurations) (SPARCS −2.19% to −7.74%; L8 Biome −3.89% to −6.49%). Confusion-matrix diagnostics identify the degradation as over-prediction of cloud on high-reflectance out-of-distribution surfaces (precision falls, recall essentially unchanged), not under-detection: the attention modules over-generalize the high-reflectance cloud cue of the bright-cloud-dominated 38-Cloud source, so the finding is scoped to transfer from such a source rather than to attention in general. A parameter-matched non-attention control shows the effect is clearly attention-driven on SPARCS, where capacity alone reproduces almost none of it, whereas on L8 Biome non-gated capacity and spatial processing account for roughly two-thirds of the floor and attention for the remainder. Reported single-seed 1–3% decoder-side gains cannot be relied upon without multi-seed and cross-dataset validation; the present benchmark provides a reference for judging future improvements against training noise.