Resources

PrONe-RS: A Graph-Isomorphic Pruning with Pareto-Driven Optimization Framework for Network Sizing in Remote Sensing LULC Applications

Deep Learning (DL) models have demonstrated remarkable performance in Remote Sensing (RS) land use land cover (LULC) classification. Yet, their high computational complexity demands can limit their deployment in resource-constrained edge computing environments. Existing structured pruning approaches are often weight-dependent and lack a principled mechanism to jointly coordinate the pruning process and select the optimal network sparsity ratio. To address these limitations, this work introduces PrONe-RS, a graph-isomorphic pruning coupled with Pareto-driven optimization framework for efficient neural network sizing in RS LULC applications. Through the incorporation of Isomorphic Structured Pruning (ISP), the network's structure is modeled as a graph to identify and create isomorphic groups of structurally similar computational blocks, which are consequently pruned according to a user-defined ratio. With this principally coordinated pruning process across layers, ISP preserves the network's structural coherence making it attractive for DL models utilizing diverse computing blocks. PrONe-RS is enhanced with a Pareto-driven multi-objective optimization scheme that first identifies the optimal pruning ratios achieving the best classification performance and complexity trade-off, and then automatically selects the most efficient one through a dedicated metric, thereby guiding the pruning ratio selection process. Extensive ablation study on the effective pruning values and the selection of the optimal one balancing the performance-complexity trade-off using five widely-used DL models including EfficientNet, ResNet, ResNeXt, VGG and Vision Transformer variations across two datasets, EuroSAT-Multispectral and AID, demonstrate significant reduction in number of trainable parameters, corresponding to 64−98% for the EuroSAT-Multispectral case and 44−70% for the AID case, all while preserving >91% classification performance across all metrics. Comparisons with existing pruning approaches on AID dataset and EuroSAT-RGB, demonstrate improved classification accuracy along with a substantial reduction in parameters, achieving a 91% reduction for EuroSAT-RGB, 36% for ResNet, and 75% for VGG on AID, respectively.