LALE: Lightweight-Transformer Architecture for Land-Cover Estimation

LALE (Lightweight-transformer Architecture for Land-cover Estimation) is a novel end-to-end semantic segmentation model designed for remote sensing imagery, integrating ConvMixer and transformer stages to optimize for both local and global features while minimizing computational overhead. The architecture achieves a strong efficiency-performance trade-off, with its smallest variant having only 1.6M parameters, yet performing within 2.6 F1 points of the best baseline (UPerNet) while using significantly fewer resources (4.5x fewer parameters, 7x less storage, and 17x fewer GMACs). This model is particularly relevant for practitioners seeking efficient segmentation solutions in resource-constrained environments.