UNET-SA: A Spatial Attention Enhanced UNET for Crop–Weed Segmentation in UAV-Based Pigeon Pea Fields

Abstract

As the global population continues to expand and the effects of climate change become increasingly evident, the demand for sustainable agricultural practices has grown more urgent. A persistent challenge in crop cultivation lies in the intense competition between crops and weeds for essential resources such as water, nutrients, and sunlight—often leading to substantial yield losses. Conventional approaches that rely heavily on herbicides and pesticides, while effective in the short term, can degrade soil health and harm the surrounding ecosystem. Hence, developing environmentally friendly and efficient weed management strategies has become a priority in precision agriculture. In this study, we introduce UNET-SA, an improved semantic segmentation framework that integrates a spatial attention mechanism into the traditional UNet architecture. The addition of spatial attention enables the model to better identify small or scattered weeds by concentrating computational focus on key regions within the image—areas that standard segmentation networks often overlook. The proposed model was trained and evaluated using a dataset of 1,727 annotated images collected from pigeon pea fields in the Vidarbha region of India. To correct manual annotation inconsistencies, HSV color space transformation was applied during preprocessing. Experimental findings demonstrate that UNET-SA delivers notable performance gains over the baseline UNet, achieving a mean Intersection over Union (IoU) of 94.44% and an overall accuracy of 98.64%, reflecting improvements of +1.74% and +1.04%, respectively. Additional testing on the larger CropAndWeed dataset further validated the model’s generalization capability, where UNET-SA achieved 98.81% accuracy and a 55.79% mean IoU, outperforming the baseline UNet (98.49% accuracy, 51.81% mean IoU). The disparity between high accuracy and moderate IoU highlights the impact of class imbalance—large background regions can inflate accuracy without reflecting true segmentation precision. Consequently, mean IoU serves as a more reliable indicator of model effectiveness. Overall, UNET-SA surpasses leading architectures such as DeepLabv3+, SegFormer, PSPNet, and LinkNet, demonstrating strong potential for practical, long-term deployment in crop–weed segmentation tasks under real agricultural conditions.