With the rapid development of remote sensing technology, high-resolution satellite imagery has been increasingly applied to land resource monitoring, urban planning, and environmental assessment. Automatically assigning semantic labels to remote sensing image patches remains a fundamental challenge due to pronounced intra-class variation and high inter-class visual similarity. To address the trade-off between model capacity and limited labeled data, this paper proposes a remote sensing image scene classification framework based on an improved EfficientNetV2-S architecture. The proposed model integrates a Squeeze-and-Excitation (SE) channel attention module between the final 1 × 1 expansion convolution and the Global Average Pooling layer, where it functions as a late-stage channel gating mechanism that adaptively recalibrates channel-wise responses, though its accuracy benefit is seed-sensitive rather than consistently reproducible at the current dataset scale. A two-stage optimization strategy was evaluated, comprising a fully unfrozen supervised baseline followed by a pseudo-label semi-supervised fine-tuning stage utilizing a strict confidence threshold (τ=0.90). Evaluated on a 10-class subset of the public NWPU-RESISC45 benchmark, the purely supervised SE-EfficientNetV2-S delivers 98.71% independent test accuracy, matching or exceeding the much larger ResNet50 (98.50%, 24.1 M parameters) despite using only 20.4 M parameters. Multi-seed variance analysis further reveals that semi-supervised fine-tuning yields a small test-set improvement for the No-SE configuration that is consistent in sign across all three seeds (+0.46 pp mean) but not statistically significant at this sample size, and an even smaller, likewise non-significant gain for the SE-augmented model (+0.08 pp), suggesting that channel gating moderates pseudo-label effectiveness in small-data regimes.
Liao et al. (2026) studied this question.