Abstract
Effective management of protected areas is critical to mitigating the global biodiversity crisis. In water-limited trophic savannas, altered herbivory regimes lead to ecosystem degradation. Consequently, there is a need for monitoring tools that can track the impact of herbivores on vegetation in space and time to inform adaptive management and restoration efforts. Here, we present an innovative tool that couples an object detection model (YOLO v10) applied to waterpoint-based camera trap images to detect herbivores with a deep-learning model to estimate vegetation category fractions from high-resolution satellite imagery (Sentinel-2). This tool allows for monitoring of herbivory across finer spatial and temporal scales than previously possible, enabling assessments at resolutions down to a few km² depending on the waterpoint density, and intervals as frequent as weekly. Our framework allows adaptive herbivory management by tying data collection on herbivore dynamics to surface water, a primary determinant of large herbivore distribution in semi-arid environments, and provides an approach to disentangle the ecological drivers influencing plant-herbivore interactions. This refined monitoring capability can enhance conservation strategies and promote the restoration of savannas, by providing detailed, real-time data on herbivore densities and vegetation changes, allowing for more targeted and adaptive management interventions.
Content
herbivory-monitoring-tool_v1.0.exe is the executable to run the herbivory monitoring tool locally.
camera_image_classifier_v1.0.exe is the executable for inference of the YOLOv10 model on locally stored camera trap images.
yoloweightsv10_lib.pt are the trained YOLOv10 model weights to detect 10 large herbivore species on camera trap images in semi-arid savannas.
DLmodelweights.h5 are the trained model weights to estimate vegetation category cover from Sentinel-2 1C imagery.
Test data.zip contains all the required data to test STEHM.
training data DL model.zip contains the training dataset used to train the vegetation category model.
YOLO training dataset.zip contains the training dataset used to train the YOLO detection model.
All scripts necessary to develop and implement STEHM and results presented in the paper are available at: https://github.com/ManuelABWeber/STEHM.git
