Deep Delta Vision Mamba: A Lightweight State Space Architecture with Deep Delta Learning for Efficient Remote Sensing
Abstract
Real-time land cover classification on autonomous satellites requires models that are accurate, lightweight, and computationally efficient within strict hardware constraints. Vision Transformers and convolutional neural networks reach state-of-the-art results on benchmarks, but their quadratic self-attention cost and large numbers of parameters make them unsuitable for edge computing. We propose Deep Delta Vision Mamba (DDV-Mamba), a hierarchical model consisting of two principled components. First, we extend the Deep Delta operator from one-dimensional to two-dimensional feature maps: each DDV block chooses to erase redundant spectral data along a learned projection direction and write discriminative data through an SSM-gated pathway. Second, an SSM-inspired gated aggregation module substitutes self-attention with depthwise convolution and channel-wise gating, restoring global context at linear rather than quadratic complexity. Assessed on EuroSAT, DDV-Mamba reaches 96.95% accuracy with 5.08 M parameters at 510 frames per second.
Key Methodologies & Contributions
- Deployment Efficiency Score (DES): Introduced a novel multi-criteria metric (DES = Accuracy × FPS / Parameters) to quantify hardware-centric edge deployment suitability, balancing accuracy, throughput, and memory footprint.
- 2D Deep Delta Operator: Successfully transitioned the Deep Delta framework to 2D spatial feature maps, enabling active, content-addressable routing that erases redundant spectral data and writes discriminative features.
- Unmatched Edge Performance: Achieved a DES of 9733.2 on the EuroSAT dataset, establishing a 13.1x relative deployment improvement over ResNet50 and an incredible 166.9x improvement over ViT-B/16.
- Architectural Necessity (Ablation): Conducted rigorous ablation studies proving the structural necessity of the Deep Delta block; removing it completely abolishes residual connections and causes a catastrophic 58.38 percentage-point accuracy collapse.