Attention-Enhanced U-Net with Residual Learning and ASPP for Efficient Crop-Weed Semantic Segmentation

Abstract

Efficient crop-weed discrimination is a crucial necessity for precision agriculture, which enables site-specific weed control, optimum herbicide use and sustained crop production. However, semantic segmentation of agricultural field photos is still problematic due to diverse backdrop, varied lighting conditions, overlapping vegetation, uneven plant structure and multi-scale item features. In such circumstances, traditional deep learning algorithms generally fail to keep the fine object boundaries and properly partition tiny weed instances. Thus, this work presents an Attention-Enhanced U-Net with Residual Learning and Atrous Spatial Pyramid Pooling (ASPP) to enhance crop-weed semantic segmentation using the PhenoBench dataset. Firstly, the transfer learning model ResNet50 was selected as the baseline image classification model and the traditional U-Net was selected as the baseline semantic segmentation model. The suggested design extends the basic U-Net architecture with residual blocks for efficient propagation of features, ASPP module for multi-scale contextual feature extraction and attention gates for adaptive feature refining. The model was trained using Adam optimizer using the hybrid loss function of Dice Loss and Sparse Categorical Cross-Entropy Loss. Experimental assessment was conducted using accuracy, precision, recall, F1-score, intersection over union (IoU) and dice coefficient, and qualitative evaluations comprising prediction maps, overlay visualizations and error maps. The suggested model performed better than the baseline ResNet50 and the traditional U-Net in segmentation, with greater border localization, better weed recognition and better retention of fine structural elements. The findings reveal that the suggested framework provides an efficient and robust solution for the automatic crop and weed semantic segmentation, and has excellent promise for real-world precision agricultural applications.