A multiscale representation learning algorithm for internal pest and disease damage imaging in trees

Forestry is a key pillar for the sustainable development of the economy and society, and assessing the health of trees is of significant importance for forestry protection. To address the challenges encountered in imaging tree internal pests and diseases, where the target size is small and its dielectric constant is similar to that of the surrounding medium—leading to blurred boundaries and potential target disappearance—this paper proposes the AtteNIS-subspace optimization method (SOM) imaging algorithm. The algorithm uses the output of the SOM algorithm as an estimate to establish a multiscale input–output learning framework. The frequency decoupling module, through an attention-based fusion strategy, combines with channel separation strips unit and frequency strips unit to achieve distributed outputs within the multiscale representation architecture, thereby enabling the inversion of the dielectric constant distribution in tree cross-sections. This study conducts comparative experiments with different percentages of moisture content, employing various quantities and sizes of tree cavities and insects based on models of cavities, insects, and hybrid models. Experimental results show that the proposed AtteNIS-SOM algorithm performs excellently across various models, demonstrating its robustness and adaptability. This provides new optimization insights for the application of electromagnetic wave-based nondestructive testing technologies in tree health monitoring and intelligent forestry.