TRANSFER LEARNING ARCHITECTURE SELECTION FOR REMOTE SENSING SCENE CLASSIFICATION
Jurnal Riset Informatika · 16.06.2026 · 8 (3) · 433-447
DOI: 10.34288/jri.v8i3.515 ↗Kurz gefasst
Fachartikel in der Zeitschrift Jurnal Riset Informatika, online erschienen am 16.06.2026. Verfasst von Akhiyar Waladi und 1 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Selecting a deep learning architecture for classifying remote sensing scenes usually involves comparing published accuracy across papers that each use different training protocols, making it unclear whether accuracy gaps reflect architecture or training differences. We isolate the architecture variable by evaluating eight models from three design families, five classical CNNs (ResNet-50, ResNet-101, DenseNet-121, EfficientNet-B0, EfficientNet-B3), two vision transformers (ViT-B/16, Swin Transformer), and one modernized CNN (ConvNeXt-Tiny), under identical training conditions on EuroSAT (10 classes, 27,000 Sentinel-2 patches) and UC Merced (21 classes, 2,100 aerial photographs). Every model shares the same ImageNet-1K initialization, AdamW optimizer, augmentation pipeline, and early stopping rule. ConvNeXt-Tiny reached the highest accuracy on EuroSAT (99.11%) and Swin-T on UC Merced (99.76%), but the accuracy range on EuroSAT was only 0.41 percentage points (1.66 on UC Merced). McNemar's test confirmed that most pairwise differences were not significant. EfficientNet-B0, the smallest model at 4.0M parameters, reached 98.76% and 99.52% while using 21x fewer parameters than ViT-B/16. On these two well-studied benchmarks, a single uniform training configuration was sufficient to bring all architectures to near-identical performance. This convergence, observed under one fixed protocol and a single data partition, suggests that on saturated classification tasks the choice of architecture may be secondary to the choice of training procedure. Whether this convergence holds on harder benchmarks, under architecture-specific optimal configurations, or with domain-specific pretraining remains to be tested
Zitieren
APA
Waladi, A., & Iftitah, H. (2026). TRANSFER LEARNING ARCHITECTURE SELECTION FOR REMOTE SENSING SCENE CLASSIFICATION. Jurnal Riset Informatika, 8(3), 433-447. https://doi.org/10.34288/jri.v8i3.515
Vancouver
Waladi A, Iftitah H. TRANSFER LEARNING ARCHITECTURE SELECTION FOR REMOTE SENSING SCENE CLASSIFICATION. Jurnal Riset Informatika. 2026;8(3):433-447. doi:10.34288/jri.v8i3.515
BibTeX
@article{Waladi2026,
author = {Waladi, Akhiyar and Iftitah, Hasanatul},
title = {TRANSFER LEARNING ARCHITECTURE SELECTION FOR REMOTE SENSING SCENE CLASSIFICATION},
journal = {Jurnal Riset Informatika},
year = {2026},
volume = {8},
number = {3},
pages = {433--447},
doi = {10.34288/jri.v8i3.515}
}
EUVIMED