Automated detection of wild plant indicators in agricultural fields: Integrating UAV technology and deep learning for result-based payments

The decline of wild plant species across European agricultural landscapes threatens biodiversity and vital ecosystem functions. While result-based payments to farmers show promise for species conservation, implementing such programs has been hindered by the high costs of traditional biodiversity monitoring. Our study explored a novel solution using uncrewed aerial vehicles (UAVs) equipped with an RGB camera and deep learning technology to efficiently detect and monitor these important plant species.

We conducted our research across four winter barley fields in Germany under different management intensities. Using the YOLO deep learning model, we analysed UAV imagery to detect segetal flora species across multiple flight altitudes. To validate and enhance our detection methodology, we collected detailed field measurements of plant traits and species coverage, and investigated whether spatial co-occurrence patterns and canopy height variations could help predict the presence of species that are challenging to detect from aerial imagery.

Our findings revealed that UAV-based monitoring could successfully detect 50% of the observed species on-site, with optimal results achieved for developing manual annotations at a ground sampling distance of 1.22mm. Plant height emerged as a crucial factor in detection success, with detection probability increasing with plant height. Based on the trait analysis, we projected similar detection success rates for key indicator species not present in our study area. The YOLO models showed accuracy rates vary between 49% to 100% depending on the management type, and performed effectively at a flight height of 40m enabling rapid field surveys that required only eight minutes per hectare. Notably, we found that both the spatial co-occurrence with easily detectable species and variations in canopy height structure showed potential as predictors for the presence of harder-to-detect species. While these findings are promising, additional research is needed to validate these relationships across broader landscape scales.

This study demonstrates the feasibility of implementing large-scale, cost-effective monitoring of wild plant indicators in agricultural settings. Our results provide a foundation for developing sophisticated 'smart indicators' for future biodiversity monitoring practices. This technological approach could make result-based conservation payments more practical and widespread, ultimately supporting the preservation of vital plant species in agricultural ecosystems.