Scholarlink

<p>Accurate and timely detection of deforestation is a prerequisite for effective forest conservation policy. While hybrid architectures combining Transformer Encoders and Graph Neural Networks (GNN) have demonstrated competitive performance in remote sensing classification tasks, the choice of fusion strategy for integrating multi-branch representations remains largely underexplored. This study presents a systematic comparative evaluation of six feature fusion strategies — Concatenation, Sum, Average, Gated, Attention, and Gated-Attention — within the TE-GNN RFFA (Random Forest Feature Adaptation) framework applied to binary forest/non-forest classification on the Amazon Forest Dataset. The TE-GNN RFFA framework first employs a Random Forest to empirically validate the informativeness of eleven engineered spectral features before feeding them into parallel Transformer Encoder and GNN branches. Six fusion variants are evaluated under identical experimental conditions using 5-fold stratified cross-validation with multiple random seeds. Results reveal that all six fusion strategies achieve comparable classification accuracy within a narrow performance band (Δmax = 0.49%, range: 0.9285–0.9334), with a non-parametric Friedman test confirming statistically significant rank ordering among strategies (χ² = 21.35, <i>p</i> = 0.0007) but no individually significant pairwise differences after Bonferroni correction. This “flat performance landscape” — where Gated fusion achieves the highest mean rank (2.42) with medium effect sizes yet without practically meaningful separation — indicates that in low-dimensional, expert-engineered spectral feature regimes, input representation quality rather than fusion mechanism complexity is the primary performance determinant. These findings provide empirically grounded guidance for practitioners: fusion architecture selection in this regime should prioritize parameter efficiency and interpretability over architectural complexity.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative analysis of feature fusion strategies in a transformer encoder–GNN framework for Amazon deforestation detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Wibowo Harry Sugiharto,
  • Agung Budi Prasetijo,
  • Heru Susanto

摘要

Accurate and timely detection of deforestation is a prerequisite for effective forest conservation policy. While hybrid architectures combining Transformer Encoders and Graph Neural Networks (GNN) have demonstrated competitive performance in remote sensing classification tasks, the choice of fusion strategy for integrating multi-branch representations remains largely underexplored. This study presents a systematic comparative evaluation of six feature fusion strategies — Concatenation, Sum, Average, Gated, Attention, and Gated-Attention — within the TE-GNN RFFA (Random Forest Feature Adaptation) framework applied to binary forest/non-forest classification on the Amazon Forest Dataset. The TE-GNN RFFA framework first employs a Random Forest to empirically validate the informativeness of eleven engineered spectral features before feeding them into parallel Transformer Encoder and GNN branches. Six fusion variants are evaluated under identical experimental conditions using 5-fold stratified cross-validation with multiple random seeds. Results reveal that all six fusion strategies achieve comparable classification accuracy within a narrow performance band (Δmax = 0.49%, range: 0.9285–0.9334), with a non-parametric Friedman test confirming statistically significant rank ordering among strategies (χ² = 21.35, p = 0.0007) but no individually significant pairwise differences after Bonferroni correction. This “flat performance landscape” — where Gated fusion achieves the highest mean rank (2.42) with medium effect sizes yet without practically meaningful separation — indicates that in low-dimensional, expert-engineered spectral feature regimes, input representation quality rather than fusion mechanism complexity is the primary performance determinant. These findings provide empirically grounded guidance for practitioners: fusion architecture selection in this regime should prioritize parameter efficiency and interpretability over architectural complexity.