SSMR-Net and Across Feature Mapping Attention are jointly applied to the UAV imagery semantic segmentation task of weeds in early-stage wheat fields

Abstract

Wheat is a critical global food crop, and its yield is significantly affected by various factors, including weeds, which can pose a major threat. Accurate identification and localization of weeds is essential for precision weeding in modern smart agriculture, with early prevention playing a key role. However, during the early growth stages, the challenge intensifies due to the significant variation in weed size, the abundance of small weeds, and the complexity of the field environment, all of which make segmentation more difficult. To address this challenge, this study combines Across Feature Mapped Attention (AFMA) with a proposed SSMR-Net model based on an improved U-Net architecture to improve weed identification. AFMA leveraged multilevel features from the original image to quantify the intrinsic relationships between large and small objects within the same category, compensating for the loss of high-level features in small target extraction and enhancing segmentation performance. SSMR-Net incorporated a multiscale feature structure by connecting the encoder and decoder with an Atrous Spatial Pyramid Pooling (ASPP) module with a small expansion rate, preserving the small target features during the information transfer and facilitating the multiscale feature extraction of weeds. The semantic differences between feature layers at the same depth were optimized through the upsampling and connection modules, whereas the encoder and decoder layers integrated an improved residual module. The skip mechanism further enabled SSMR-Net to capture features at various levels. This makes SSMR-Net maintain high segmentation performance in different complex scenarios. The combination of SSMR-Net and AFMA is more suitable for the UAV imagery semantic segmentation task of weeds in early-stage wheat fields. The experimental results demonstrated that the proposed SSMR-Net combined with AFMA achieved superior segmentation accuracy for weed and wheat identification on a custom-built wheat and weed dataset, outperforming existing models with a weed accuracy of 0.774, an IoU score of 0.696, and an mIoU of 0.865. This study presents a promising approach to precise weed identification and control in agriculture.