STA-UNet: Lightweight Spectral Segmentation for Mowing Robots

Abstract

Achieving robust semantic segmentation for autonomous mowing robots requires discriminating between materials with subtle visual differences under challenging outdoor conditions. To address this, we present a novel framework that co-designs feature-specific spectral imaging and a lightweight deep learning network. On the sensing front, we develop a low-cost multispectral camera using a custom multi-band-pass (MBP) filter, engineered to capture distinctive spectral signatures of key classes like grass and shrubs, enabling parallax-free and compact data acquisition. For efficient inference, we propose STA-UNet, a lightweight encoder-decoder network that innovatively incorporates a Spectral Attention (SA) module to adaptively fuse the acquired spectral bands, and a Cross-Scale Texture Perception (CSTP) module to enhance spatial feature representation. Trained on Grass8K—a new 8,191-image multispectral dataset for lawn environments—STA-UNet achieves a remarkable mIoU of 88.63%, outperforming a standard U-Net by 7.11% while utilizing only one-quarter of the parameters. The system achieves real-time segmentation on an edge device with 5 TOPS of computing power, demonstrating a practical, high-performance perception solution for resource-constrained robotic applications. The dataset and code are made available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://drive.google.com/drive/folders/</uri> 1g1dzyc-Rn6iUWGGvIbnRr wRKUn2bic4?usp=drive link