Analyzing the computational cost of EffBaGAN in data-scarce remote sensing scenarios

Generative Adversarial Networks (GANs) applied to remote sensing classification have shown strong performance in scenarios with limited labeled data. However, their deployment on resource-constrained edge computing devices, such as autonomous aerial vehicles or satellites, is often limited by their high computational cost. This work presents a preliminary study on the impact of computational optimization techniques applied to EffBaGAN, aiming to facilitate the projection over edge computing devices. EffBaGAN is a neural architecture that integrates EfficientNet and a Balancing GAN (BAGAN) to tackle the classification of multispectral images under conditions of data scarcity and class imbalance. We evaluate the effect of several optimization techniques, including Automatic Mixed Precision (AMP), Post-Training Quantization (PTQ), Quantization-Aware Training (QAT), and structured magnitude-based pruning, on the computational efficiency of the EffBaGAN neural architecture when executing on a multicore CPU + GPU system. Experimental results on high-resolution multispectral datasets for biodiversity assessment show that the network can be significantly optimized without degrading classification accuracy. In particular, PTQ and QAT achieved a network size reduction of over 72%, while structured pruning successfully removed 30% of the parameters and floating-point operations. Moreover, we show that suitable combinations of optimization strategies, such as QAT combined with pruning, can substantially reduce computational overhead, thereby improving the feasibility of deploying EffBaGAN in resource-constrained environments.

keywords: High Performance Computing, edge computing, multispectral, classification, deep learning, data scarcity