Dynamic Bit-Plane Augmentation Framework for Enhancing Data Quality and Robustness in Remote Sensing

In AI-driven remote sensing, data quality issues critically affect the generalization and reliability of deep learning models. To address this problem, we propose the Dynamic Attention Bit-Plane Augmentation (DAPW) framework, designed for task-relevant and structure-adaptive augmentation in multi-scene remote sensing tasks. At the core of DAPW is an adaptive selection mechanism that utilizes a dynamic perception window and entropy-gradient-contrast metrics to identify informative bit-plane regions. This approach effectively reduces redundancy and enhances the structural expressiveness of augmented data. To further improve robustness and intra-class diversity, we incorporate a lightweight tunable augmentation strategy that introduces controlled appearance-level perturbations while maintaining semantic integrity. The integration of these two components enables a unified augmentation pipeline, which is validated across three representative tasks to demonstrate cross-domain generalizability. To our knowledge, this is the first study to apply bit-plane augmentation to tree counting. Extensive experiments show consistent improvements across diverse scenarios: on the London Tree Counting dataset ( $\mathbf {E}_\mathbf {MAE}$ = 19.9, $\mathbf {E}_\mathbf {F1}$ = 72.4%), DOTA object detection (mAP = 83.6%), and CIFAR-100 classification (Top-1 = 84.17%). Compared to prior approaches such as BIRD and conventional bit-plane recombination, DAPW offers systemic advances in adaptive region selection, redundancy suppression, and alignment between augmentation strategy and downstream tasks, providing a more efficient and interpretable solution for improving data quality and model performance in AI-based remote sensing.