Accurate and timely land cover and land use (LCLU) classification from medium-spatial-resolution optical time-series data is essential for large-scale environmental monitoring. lightweight deep neural networks (DNNs) offer reduced computational and memory requirements, enabling efficient deployment in resource-constrained scenarios. While popular in computer vision tasks, their ability to simultaneously model spatial, spectral, and temporal information for medium-resolution optical time series is understudied. This study addresses this gap by evaluating seven existing lightweight models spanning four architectural families: convolutional and recurrent hybrids, convolutional and transformer hybrids, 3D convolutional models, and video transformers against a traditional hybrid convolutional transformer (CNNTransformer) benchmark across the Conterminous United States (CONUS). Models are trained on 500,000 Landsat time-series samples with 25 repetitions and evaluated across five model sizes (3k, 5k, 10k, 25k, and 50k parameters) to assess both accuracy and stability. Results show that Simple Recurrent Unit (SRU)-based lightweight hybrids provide the best performance. Specifically, MobileNetSRU consistently outperformed the benchmark at small-to-moderate model sizes (3k–15k), achieving peak relative improvement gains of ~2.5–7.5% at 7.5k parameters. MobileNetSRU also demonstrated superior robustness in limited-data scenarios (50k training samples), particularly for spectrally stable classes like water and bare land. This study reveals that the inherent inductive bias of recurrent-based lightweight models aligns more effectively with the sequential phenology of satellite data than more flexible, data-hungry attention mechanisms at small parameter scales. These findings suggest that strategically matching architectural priorities to temporal data structures can significantly reduce the trade-off between model efficiency and classification accuracy in scalable Earth-observation workflows.
Wang et al. (Mon,) studied this question.