Abstract <p>Dust has an adverse effect on the environmental sensors of automated agricultural machinery. This issue highlights the need for further investigation. Refining existing deep learning methods for dust removal is critical for improving the performance and reliability of these machines in dusty environments. We propose an end-to-end trainable learning network (DedustNet) to address the challenge of agricultural dust removal in real-world scenarios. To our knowledge, DedustNet represents the first application of Swin Transformer-based units in wavelet networks for agricultural image dusting. DedustNet leverages frequency-dominated Swin Transformer-based blocks, namely, DWTFormer and IDWTFormer, in tandem with wavelet transforms. Acting as fundamental encoding and decoding units, these blocks enable the retrieval of intricate details, such as image structural and textural components. This approach effectively overcomes the limitations of the global receptive field in Swin Transformer under complex, dusty backgrounds. In addition, DedustNet incorporates a cross-level information fusion module (CIFM). This module adeptly integrates features from different levels, thereby facilitating the capture of global and long-range feature relationships. Moreover, DedustNet is enhanced by a dilated convolution module (DCM). This module leverages the guidance of wavelet transforms to extract contextual information at multiple scales. Compared with those of existing state-of-the-art methods, DedustNet achieves superior performance and more reliable results in agricultural image dedusting, validating the powerful application of applied intelligence in agriculture. Additionally, the impressive performance on real-world hazy datasets and application tests highlights DedustNet’s superior generalizability and computer vision-related application performance.</p> Graphical abstract <p></p>