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, controlled by a channel-wise sigmoid gate. 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. To assess deployment feasibility beyond pure accuracy, we define the Deployment Efficiency Score (DES = Accuracy×FPS / Parameters(M)), a multi-criteria score that balances model performance, speed, and efficiency. Assessed on EuroSAT, DDV-Mamba reaches 96.95% accuracy with 5.08 M parameters at 510 frames per second, achieving DES = 9733.2, a 13.1× relative improvement over ResNet50 (DES = 743.2) and 166.9× over ViT-B/16 (DES = 58.3). An ablation experiment validates the architectural necessity of the Deep Delta block, whose removal causes a 58.38 percentage-point accuracy collapse.

Key Methodologies & Contributions

  • 2D Deep Delta Operator: Extended the Deep Delta operator from one-dimensional to two-dimensional spatial feature maps. This operator enables each block to erase redundant spectral data along a learned projection direction and write discriminative information.
  • SSM-Gated Aggregation: Replaced quadratic self-attention with a State Space Model (SSM)-inspired gated aggregation module that utilizes depthwise convolution and channel-wise gating to restore global context at linear computational complexity.
  • Deployment Efficiency Score (DES): Introduced a novel, hardware-centric compound metric that balances model classification accuracy, inference speed (FPS), and parameter count into a unified scalar value.
  • High-Throughput Edge Performance: Reached 96.95% accuracy with 5.08 M parameters at 510 frames per second on the EuroSAT dataset. This resulted in a DES of 9733.2, achieving a 13.1x relative improvement over ResNet50 and 166.9x over ViT-B/16.
  • Architectural Ablation: Conducted a controlled ablation experiment confirming that removing the Deep Delta block abolishes all residual connections and triggers a 58.38 percentage-point accuracy drop, validating its necessity as the architectural backbone.


Code & Resources

Status: Accepted at IEEE CONECCT, 2026. Authors: L. Chhetri, A. Kumar