Training and inference Time Efficiency Assessment Framework for machine learning algorithms: A case study for hyperspectral image classification

The increasing complexity and scale of remote sensing datasets, coupled with the challenges of accurately estimating algorithmic time efficiency, often lead to significant resource waste or even failure when using machine learning algorithms in urgent or resource-constrained scenarios. Accurate time efficiency estimation is critical for deploying effective algorithms, yet it remains challenging due to the many factors influencing computational performance. Traditional methods of evaluating time efficiency often neglect the effects of core model parameters and complex data scales in spectral and temporal dimensions. In addition, inference time, an essential factor in real-world applications, is often overlooked. To address these limitations, we propose the Time Efficiency Assessment Framework (TEAF), a novel method for evaluating the time efficiency of machine learning algorithms. Through mathematical reasoning, TEAF models the training and inference time as functions ($\psi$) of complex data scales and core model parameters. The strong linear correlation between $\psi$ and the actual runtime allows TEAF to accurately predict the time and cost of machine learning tasks with a low computational overhead before algorithm execution. To validate this framework, we derived TEAF formulations for five classical machine learning algorithms and tested them on state-of-the-art hyperspectral image datasets and Sentinel-2 multispectral datasets. The results demonstrated that TEAF could accurately predict both training and inference time for various algorithms, with a strong linear correlation between $\psi$ and actual runtime ($R^2>0.942$). This study offers a practical solution to the challenges posed by the increasing volume and complexity of data in remote sensing image processing. The code is available at https://github.com/SCUT-CCNL/TEAF.