Functional ANOVA Analysis of Design Choices in Remote Sensing Multi-Label Classification

A recent preprint on arXiv (2608.04702) presents a novel approach for assessing deep learning models in the context of multi-label classification (MLC) in remote sensing, extending beyond mere rankings. Utilizing functional analysis of variance (fANOVA), the research meticulously evaluates how various design choices—like network architecture, fine-tuning methods, learning strategies, and initialization—affect performance variability. The study involved two empirical analyses, examining 48 and 20 deep learning models across seven remote sensing image datasets. Through fANOVA, the authors developed dataset meta-representations that illustrate sensitivity to design choices. Hierarchical clustering of these representations indicated that datasets tend to cluster based on their responses to design decisions, closely tied to their inherent properties. This research seeks to enhance the understanding of factors influencing model performance, offering insights that transcend dataset-specific rankings. The preprint can be accessed on arXiv with the identifier 2608.04702.