Deep learning and drone imagery for the automatic detection and quantification of common vole burrows in agricultural landscapes

[EN] Accurate detection of vole burrows is critical for monitoring rodent outbreaks and mitigating agricultural damage. This study compared deep learning (DL) and traditional supervised classification approaches -Random Forest (RF) and Support Vector Machines (SVM)- using high-resolution UAS imagery...

Descripción completa

Detalles Bibliográficos
Autores: Santiago-Aliste, Alberto, Sánchez Martín, Nilda, Charfolé de Juan, José Francisco, Pérez Sánchez, Rodrigo, Plaza Martín, Javier
Tipo de recurso: artículo
Estado:Versión publicada
Fecha de publicación:2025
País:España
Institución:Universidad de Salamanca (USAL)
Repositorio:GREDOS. Repositorio Institucional de la Universidad de Salamanca
OAI Identifier:oai:gredos.usal.es:10366/167801
Acceso en línea:http://hdl.handle.net/10366/167801
Access Level:acceso abierto
Palabra clave:Deep learning
Common vole
UAS imagery
Burrows
Object-oriented classifications
Aprendizaje profundo
Topillo común
Imágenes aéreas no tripuladas
Madrigueras
Clasificaciones orientadas a objetos
1203.04 Inteligencia Artificial
5102.01 Agricultura
2506.16 Teledetección (Geología)
3308.08 Tecnología del Control de Roedores
Descripción
Sumario:[EN] Accurate detection of vole burrows is critical for monitoring rodent outbreaks and mitigating agricultural damage. This study compared deep learning (DL) and traditional supervised classification approaches -Random Forest (RF) and Support Vector Machines (SVM)- using high-resolution UAS imagery at two flight heights (15 m and 25 m). DL, implemented via convolutional neural networks, consistently outperformed classification methods, mainly due to a drastic reduction in misclassifications. The best DL configuration achieved a precision of 0.92, recall of 0.81, and an F1 score of 0.83, reflecting strong detection capability with limited false alarms. A sensitivity analysis further indicated that reducing the object-oriented classification to two classes (burrow vs. non-burrow) improved efficiency without major performance losses compared to a four-class scheme. Higher spatial resolution (15 m flights) enhanced detection across all methods, especially benefiting DL by improving recall without sacrificing precision. The findings highlight that DL is particularly suited for detecting relatively discrete objects like burrows, offering a reliable tool for early intervention in pest control strategies. In contrast, traditional classifications processed the imagery >95 % faster than the DL workflow but were hindered by excessive misclassifications that compromised overall accuracy. This work underscores the potential of UAS-based DL detection pipelines to improve pest monitoring in agricultural ecosystems, providing a scalable and accurate solution for early warning and management programs. Future research should focus on linking burrow density to vole population estimates to further assist in plague management.