Efficient crop–weed semantic segmentation via knowledge distillation from DeepLabV3+ to Fast-SCNN

Precision agriculture plays a crucial role in sustainable food production and resource-efficient land management. A key component of precision agriculture is the semantic segmentation of crops and weeds that enables autonomous systems to minimize herbicide use and environmental impacts through targeted weeding. However, state-of-the-art segmentation models — such as DeepLabV3+ — demand high computational resources, making them impractical for low-power agricultural devices deployed in the field. To address this problem, this study introduces a green artificial intelligence approach using a novel knowledge distillation (KD) framework that transfers semantic knowledge from a DeepLabV3+ teacher to a lightweight Fast-SCNN student model. The proposed approach is designed specifically for energy-efficient crop–weed segmentation on the CWFID dataset. The framework employs a dynamic alpha scheduling approach to balance hard and soft label supervision and a patch-based training strategy to handle high-resolution field imagery efficiently. The distilled Fast-SCNN achieved a mean Intersection over Union (mIoU) of 0.8695 and a pixel accuracy of 0.9879, with a compact model size of only 4.54 MB and real-time inference capability of 4.84 FPS on an NVIDIA T4 GPU. Compared with several state-of-the-art lightweight deep learning architectures, our method achieved a competitive accuracy with significantly reduced computational and energy costs. These findings demonstrate that knowledge distillation supports sustainable AI practices by reducing the carbon footprint of deep learning models while maintaining high performance in precision agriculture applications.