1School of Geography and Planning, Sun Yat-sen University, Guangzhou, 510700, China
2Department of Environmental Health Sciences, University of California Los Angeles, Los Angeles, CA, 90095, USA
3Department of Ocean Sciences, University of California Santa Cruz, Santa Cruz, CA, 95061, USA
4These authors contributed equally to this work.
| Received 30 Jun 2024 |
Accepted 12 Dec 2024 |
Published 24 Feb 2025 |
The threat of crop diseases can lead to reduced yields and significantly hinder progress towards achieving the sustainable development goal of ”zero hunger.” In the process of detecting crop diseases, variations in data collection conditions can lead to significant differences in the spatial distribution features of training and testing data. Models trained on specific datasets often perform poorly when applied to detect crop diseases in new datasets, significantly affecting the performance of cross-domain object detection tasks. To address the challenges of cross-domain crop disease object detection, this paper proposes a Multi-Granularity Alignment (MGA) domain adaptation framework that is compatible with other object detectors and is generalizable. This approach integrates the multi-granularity alignment and omni-scale gated fusion domain adaptation components into an enhanced object detector, aiming to align features between the source and target domains and reduce their disparities. MGA conducts scale-aware convolutional aggregation on the feature maps of the object detector, and it utilizes three different levels of discriminators—category, instance, and pixel—to identify the domain source, aligning features across different domains from a granularity-dependent perspective, thereby achieving cross-domain object detection. The experimental results demonstrate that MGA achieves the highest mAP in various datasets collected from different regions, environments, and styles, with scores of 47.9 % (dataset: PVi → CDi), 48.3 % (dataset: PDc → PVi), and 49.2 % (dataset: Data with style transfer → CDi). This performance significantly surpasses other object detection technologies. When integrated into Faster R-CNN, MGA also achieves a remarkable mAP of 44.7 % on the CDi → Data w/style transfer dataset, demonstrating robust generalization capabilities.
