ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Test-Time Adaptive Latent Feature Restoration for UAV Crop–Weed Segmentation under Altitude-Induced Resolution Shifts

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601493.(doi:10.13031/aim.202601493)
Authors:   Yu Tian, Xintong Jiang, Nazia Hossain, Asahi Suzuki, Jan Franklin Adamowski, Shangpeng Sun
Keywords:   integrated weed management; crop–weed segmentation; test-time adaptation; domain shift; precision agriculture

Abstract. Weeds are among the leading causes of agricultural production losses and remain one of the greatest threats to global food security. Accurate crop–weed segmentation is essential for understanding weed spatial distribution and supporting field-scale precision management. However, for large-scale field applications, increasing the flight altitude is necessary to reduce flight time. Nevertheless, the resulting loss of spatial detail can substantially degrade crop–weed segmentation performance when using existing models trained on low-altitude, high-resolution imagery. Although existing image-space restoration methods reconstruct visually plausible detail by enforcing pixel-level statistics, such appearance-driven recovery does not guarantee preservation of the semantic representations on which crop–weed boundary prediction depends. In this study, we address this limitation by performing generative diffusion in the latent feature space, restoring segmentation-relevant representations through a pre-trained latent diffusion U-Net with lightweight LoRA adaptations, interfaced with features produced by a pre-trained DINOv3 encoder. A degradation-aware test-time adaptation stage further refines the restored representation for each test image by enforcing consistency across multiple degradation variants of the input, without modifying model parameters. The approach is evaluated on UAV imagery acquired over the same field at 10 m, 30 m, and 70 m flight altitudes, corresponding to progressively coarser ground sampling distances. Under direct low-to-high altitude transfer, mIoU falls from 78.71% at 10 m to 47.47% at 70 m, with crop IoU declining more severely than weed IoU. Latent feature restoration recovers 14.8 percentage points of mIoU at 70 m over direct transfer, and the full pipeline with degradation-aware test-time adaptation yields the highest average performance across all three altitude subsets, with gains concentrated on the most degraded imagery. The proposed method can empower farmers and breeders with innovative insights for smart weed management, such as precision spraying, to reduce chemical inputs.

(Download PDF)    (Export to EndNotes)