Title: LALE: A Lightweight-Transformer Architecture for Land-Cover Estimation

Abstract

Semantic segmentation in remote sensing demands models that balance global context with local detail, all while operating within strict computational constraints. Previous research has generally focused on optimizing a single aspect: utilizing attention mechanisms for global awareness, convolutions for local precision, or compact designs for efficiency. Although hybrid methods attempt to address both needs, they often necessitate architectural modifications and encoder backbones that introduce significant computational overhead, thereby compromising both speed and performance.

To address these limitations, we introduce LALE (Lightweight-transformer Architecture for Land-cover Estimation), an end-to-end segmentation framework for remote sensing imagery. LALE splits its encoder by resolution: it employs lightweight ConvMixer stages to process high-resolution local features, while transformer stages manage low-resolution global context. This design restricts the quadratic complexity of self-attention to deep, downsampled feature maps. Additionally, the architecture incorporates an all-MLP multi-scale decoder, along with RMSNorm and StarReLU layers throughout the network, to further minimize both parameter counts and computational demands.

Evaluated on the large-scale ARAS400k remote-sensing segmentation benchmark, LALE demonstrates a superior efficiency-performance balance compared to CNN, transformer, and hybrid baseline models. The model’s smallest variant, containing only 1.6 million parameters, achieves an F1 score within 2.6 points of the top-performing baseline, UPerNet. Notably, this variant requires 4.5 times fewer parameters, occupies 7 times less storage, utilizes 17 times fewer GMACs, and delivers 1.8 times higher throughput.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC