FreqDyn-YOLO: A High-Performance Multi-Scale Feature Fusion Algorithm for Detecting Plastic Film Residues in Farmland

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 16; p. 4888
• Main Authors: Zhang, Mingyang, Zhang, Jianjie, Peng, Yihang, Wang, Yi
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 08.08.2025; MDPI
• Subjects: Agricultural land; Algorithms; Annotations; Computer vision; Datasets; DCNv4; Deep learning; Efficiency; Environmental protection; Environmental stewardship; FreqFTA; HPMSFPN; Loam soils; Localization; Machine learning; Morphology; Plastics; Remote sensing; residual film; Unmanned aerial vehicles; Waste management; YOLO11; China; DCNv4; FreqFTA; YOLO11; residual film; HPMSFPN
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25164888

Plastic mulch technology plays an important role in increasing agricultural productivity and economic returns. However, residual mulch remaining in agricultural fields poses significant challenges to both crop production and environmental sustainability. Effective recovery and recycling of residual plastic mulch requires accurate detection and identification of mulch fragments, which presents a substantial technical challenge. The detection of residual plastic film is complicated by several factors: the visual similarity between residual film fragments and soil in terms of color and texture, as well as the irregular shapes and variable sizes of the target objects. To address these challenges, this study develops FreqDyn-YOLO, a detection model for residual film identification in agricultural environments based on the YOLO11 architecture. The proposed methodology introduces three main technical contributions. First, a Frequency-C3k2 (FreqC3) feature extraction module is implemented, which employs a Frequency Feature Transposed Attention (FreqFTA) mechanism to improve discrimination between residual film and soil backgrounds. Second, a High-Performance Multi-Scale Feature Pyramid Network (HPMSFPN) is developed to enable effective cross-layer feature fusion, enhancing detection performance across different target scales. Third, a Dynamic Detection Head With DCNv4 (DWD4) is introduced to improve the model’s ability to adapt to varying film morphologies while maintaining computational efficiency. Experimental findings on a self-developed agricultural field residual film dataset confirm that FreqDyn-YOLO outperforms the baseline approach, achieving improvements of 5.37%, 1.97%, and 2.96% in mAP50, precision, and recall, respectively. The model also demonstrates superior performance compared to other recent detection methods. This work provides a technical foundation for precise residual film identification in agricultural applications and shows promise for integration into automated recovery systems.