文献速递|无人机辅助深度学习支持农业环境计划生物多样性监测

结果支付计划依赖对生物多样性指标物种的可靠监测，但传统人工核查成本高、效率低，难以支撑大规模方案推广，也制约了农业环境政策的效果评估精度。德国Eco-Scheme 5以草原植物物种或属内物种组作为生物多样性指示指标，各地区计数方式不统一，影响跨区域可比性和方案实施的公平性。本研究聚焦于利用无人机搭载高分辨率相机采集草原影像，结合深度学习目标检测实现指示物种的自动化识别与计数，为解决人工核查瓶颈提供了新技术路径。

Fig. 1. (a). Number of indicators used to assess grasslands in various regions of Germany (status in 2020). The regions are grouped as follows: NO - Mecklenburg-Vorpommern and Brandenburg, NW - Schleswig-Holstein and Lower Saxony, MW - Hesse, Rhineland-Palatinate and Saarland, MO - Saxony-Anhalt and Thuringia, SN - Saxony, BW - Baden-Wuerttemberg and BY - Bavaria. Fig. 1 (b). Distribution of indicators based on their commonality across the regions. For example, 31 plant species and/or genera (left-end) are region-specific indicators, recognized only in any one of the seven regions, whereas five species and/or genera (right-end) are common indicators recognized in all seven regions.

方法论包含数据采集、模型训练和跨域泛化验证三个核心环节。数据采集在德国ATB研究所实验草地进行，同时整合了既往研究和开源数据库的地基图像扩充训练样本，有效提升了模型跨域泛化能力。

Fig. 2. Sample images of the selected indicators from the UAV and ground-based image (GBI) datasets, illustrating variations in perspective, resolution, and development stages. GBIs I refers to ground-based images from our previous work, whereas GBIs II includes images sourced from the Global Biodiversity Information Facility database (GBIF).

目标检测采用EfficientDet单阶段模型，训练采用随机梯度下降优化器并利用数据增强策略提升跨域泛化能力，标注采用开源LabelImg工具完成。

Fig. 3. Schematic overview of the proposed data enhancement approach for cross-domain knowledge transfer between UAV and ground-based image domains. The high-resolution UAV training dataset was enriched with ground-based imagery to enhance model generalization and improve species detection in UAV images. The data enhancement approach was based on the assumption that deep learning models can be applied to learn universal representation across these two image domains.

实验结果表明：增强模型在无人机和地基影像上均表现稳健，验证了跨模态泛化的有效性；微调策略在标注数据有限时显著优于从零训练，证明了迁移学习的重要价值；大多数指示物种在营养期和花期均被成功检测，宽叶物种检测分数最高，Daucus carota检测分数最低。结果表明只要无人机影像中保留部分植物可见部位即可实现有效检测，为草地监测提供了技术可行性验证，为农业环境方案的规模化推广奠定了数据基础。

本研究成功展示了无人机与深度学习整合用于草原生物多样性监测的巨大潜力，有效解决了结果支付计划中环境指标可靠验证的核心障碍，为农业环境方案的规模化推广提供了可验证的技术路径。方法已验证可支持RBP方案的核查流程，具备向其他农业环境计划迁移的可行性。后续工程应用应重点关注：积累更多草原类型和气候区的训练数据以提升跨域泛化能力；探索多无人机协同采集以扩大单次任务覆盖范围；完善模型在花期以外的季节性检测鲁棒性；推动技术标准的建立以支撑结果支付计划的规范化实施。