DeltaGate-CNN: A Lightweight Convolutional Architecture for Efficient Remote Sensing Image Classification
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 the DeltaGate-CNN (DG-CNN), a hierarchical architecture combining two mechanisms for edge deployment. First, a Gated Depthwise Aggregation module efficiently captures local spatial context. Second, we extend the Deep Delta operator to 2D feature maps, allowing each block to selectively erase redundant spectral data and write discriminative features using a channelwise sigmoid gate. 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, DG-CNN 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
- 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.
Publication Status
Under Review at IEEE INDICON, 2026
Authors: L. Chhetri, A. Kumar, P. Ghosal