Monitoring vegetation patterns and their drivers to infer resilience: Automated detection of vegetation and megaherbivores from drone imagery using deep learning

James, Rebecca K. | Daniels, Freek | Chauhan, Aneesh | Wicaksono, Pramaditya | Hafizt, Muhammad | Harahap, Setiawan Djody | Christianen, Marjolijn J.A.

Ecological Informatics, Open Access, Volume 81, July 2024, Article number 102580

Abstract

Ecological pattern theory has highlighted spatial vegetation patterns that can be used as indicators of ecosystem resilience. Combining this spatial pattern theory with aerial imagery from drones and automated image processing with deep learning methods, we show how monitoring of natural ecosystems can be enhanced through quantifying vegetation spatial patterns. We demonstrate this approach in a tropical seagrass ecosystem with a high abundance of turtles that generate vegetation patches when grazing. Past field observations suggest that patch size and density reflect the seagrass meadow resilience, but understanding the natural variation in vegetation patchiness is crucial. Employing the deep learning methods of semantic segmentation and object detection, we quantify vegetation patchiness metrics and turtle distribution across 12 ha of seagrass meadow in the years 2012 and 2022. The resulting output facilitates spatial and temporal comparisons, revealing areas of low resilience. In 2012, turtle grazing across the entire site yielded vegetation patch sizes averaging 2 ± 0.2 m2 (95% confidence interval). Reduced patch sizes of 0.24 ± 0.05 m2 and 0.67 ± 0.6 m2 at the reef edge and beach slope respectively, in conjunction with a reduced patch density, indicated lower resilience at the seagrass meadow edges. Analysis of the 2022 dataset indicates a general decrease in patch size over time, suggesting declining resilience. A retraining experiment of the semantic segmentation model was conducted where the initial model was retrained on the 2022 dataset and demonstrated the adaptability of the deep learning methods. Despite using different equipment, the model achieved high accuracy with only 5–10 additional training images. By providing the tools to conduct these analyses, we aim to stimulate the uptake of deep learning for enhancing the data obtained from aerial imagery to improve the monitoring and conservation of natural ecosystems. © 2024 The Authors read more