{"id":9335,"date":"2026-02-10T11:01:38","date_gmt":"2026-02-10T09:01:38","guid":{"rendered":"https:\/\/teodorcostachioiu.com\/publicatii\/"},"modified":"2026-11-03T13:33:33","modified_gmt":"2026-11-03T11:33:33","slug":"publicatii","status":"publish","type":"page","link":"https:\/\/teodorcostachioiu.com\/ro\/publicatii\/","title":{"rendered":"Publica\u021bii"},"content":{"rendered":"
<\/a><\/form>

2026<\/span><\/h3>

Journal Article<\/span> MDPI RS<\/span> Q2<\/span> IF<\/abbr> 4.2<\/span><\/span><\/p><\/div>

SENSERO: A Multi-Scale, Multi-Label Sentinel-2 Benchmark for Romania with Cross-Scale Land-Cover Classification Baselines<\/a><\/p>

Cost\u0103chioiu, Teodor; Vasilescu, Vlad; Grivei, Alexandru-Cosmin; Griparis, Andreea; Faur, Daniela<\/p>

In: <\/span>Remote Sensing, <\/span>vol. 18, <\/span>no. 21, <\/span>2026<\/span>, ISSN: 2072-4292<\/span>.<\/p>

Abstract<\/a><\/span><\/span>Links<\/a><\/span><\/span>BibTeX<\/a><\/span><\/p>

@article{rs18213655,
\r\ntitle = {SENSERO: A Multi-Scale, Multi-Label Sentinel-2 Benchmark for Romania with Cross-Scale Land-Cover Classification Baselines},
\r\nauthor = {Teodor Cost\u0103chioiu and Vlad Vasilescu and Alexandru-Cosmin Grivei and Andreea Griparis and Daniela Faur},
\r\nurl = {https:\/\/www.mdpi.com\/2072-4292\/18\/21\/3655},
\r\ndoi = {10.3390\/rs18213655},
\r\nissn = {2072-4292},
\r\nyear  = {2026},
\r\ndate = {2026-01-01},
\r\nurldate = {2026-01-01},
\r\njournal = {Remote Sensing},
\r\nvolume = {18},
\r\nnumber = {21},
\r\nabstract = {Patch size is an important design parameter in patch-based Earth observation (EO) learning, yet it is rarely treated as an experimental variable because existing benchmarks use a single fixed scale. We introduce SENSERO, a multi-scale, multi-label Sentinel-2 benchmark for Romania consisting of 10,000 geographically fixed locations extracted at eight patch sizes, from 64\u00d764 to 336\u00d7336 pixels, at a constant 10 m ground sampling distance, so that all scales share identical patch centres and an identical, spatially disjoint 70\/15\/15 partition. Each patch provides twelve-band bottom-of-atmosphere reflectance, a rasterised CORINE Land Cover (CLC) 2018 reference map from which five deterministic multi-label annotation layers are derived (three levels of the CLC hierarchy, the BigEarthNet nomenclature, and a scheme adapted to the Romanian landscape). The benchmark task is patch-level multi-label classification; the rasterised reference maps serve to derive and trace the labels, not as a segmentation target. Using ResNet-50 baselines trained from scratch, we find that larger patches bring little benefit at CLC Level 1 but clear gains at Levels 2 and 3 (macro mAP at Level 3 rising from 0.42 at 64 px to 0.61 at 336 px), with per-location gains saturating around 224\u2013256 pixels; that transfer across scales is asymmetric, models trained on wide patches transferring to narrow ones far better than the reverse; and that the choice of nomenclature affects measured performance as much as patch size. Leave-scene-out cross-validation, an ablation of the labelling rule, pretrained ResNet-50 and ViT-S\/16 controls and bootstrap confidence intervals show that the scale effect is not attributable to spatial leakage, to the annotation rule, or to the baseline architecture. Comparison with an independent 10 m land-cover product shows that the gain from context is concentrated where CLC labels generalise over heterogeneous ground.},
\r\nkeywords = {},
\r\npubstate = {published},
\r\ntppubtype = {article}
\r\n}
\r\n<\/pre><\/div>
Close<\/a><\/p><\/div>
Patch size is an important design parameter in patch-based Earth observation (EO) learning, yet it is rarely treated as an experimental variable because existing benchmarks use a single fixed scale. We introduce SENSERO, a multi-scale, multi-label Sentinel-2 benchmark for Romania consisting of 10,000 geographically fixed locations extracted at eight patch sizes, from 64\u00d764 to 336\u00d7336 pixels, at a constant 10 m ground sampling distance, so that all scales share identical patch centres and an identical, spatially disjoint 70\/15\/15 partition. Each patch provides twelve-band bottom-of-atmosphere reflectance, a rasterised CORINE Land Cover (CLC) 2018 reference map from which five deterministic multi-label annotation layers are derived (three levels of the CLC hierarchy, the BigEarthNet nomenclature, and a scheme adapted to the Romanian landscape). The benchmark task is patch-level multi-label classification; the rasterised reference maps serve to derive and trace the labels, not as a segmentation target. Using ResNet-50 baselines trained from scratch, we find that larger patches bring little benefit at CLC Level 1 but clear gains at Levels 2 and 3 (macro mAP at Level 3 rising from 0.42 at 64 px to 0.61 at 336 px), with per-location gains saturating around 224\u2013256 pixels; that transfer across scales is asymmetric, models trained on wide patches transferring to narrow ones far better than the reverse; and that the choice of nomenclature affects measured performance as much as patch size. Leave-scene-out cross-validation, an ablation of the labelling rule, pretrained ResNet-50 and ViT-S\/16 controls and bootstrap confidence intervals show that the scale effect is not attributable to spatial leakage, to the annotation rule, or to the baseline architecture. Comparison with an independent 10 m land-cover product shows that the gain from context is concentrated where CLC labels generalise over heterogeneous ground.<\/div>
Close<\/a><\/p><\/div>