Computer Engineer (or so I try)
Huelva, Spain
Joined June 2023
- Tweets67
- Following93
- Followers57
- Likes118
Pinned Tweet
🆕 New paper out in Eng. applications of AI!
We propose iterative deep learning for detecting cetacean whistles in one of the noisiest marine regions: the Strait of Gibraltar 🐋🔊
While baseline models collapsed under noise, we reached 0.88 F1
🔗 sciencedirect.com/science/ar…
Alba Márquez-Rodríguez retweeted
Detectan silbidos de cetáceos en el Estrecho de Gibraltar usando Inteligencia Artificial lavozdigital.es/provincia/ca…
Alba Márquez-Rodríguez retweeted
Investigadoras de la Universidad de Cádiz desarrollan un sistema para detectar silbidos de cetáceos en el Estrecho de Gibraltar
🌐 csur.red/43FN50YFxaU
Alba Márquez-Rodríguez retweeted
Un equipo del Instituto de Investigación Marina (INMAR) de la Universidad de Cádiz ha desarrollado un sistema de inteligencia artificial para detectar silbidos de cetáceos en el Estrecho de Gibraltar, uno de los entornos marinos más ruidosos y complejos del mundo.
Alba Márquez-Rodríguez retweeted
🌊 Investigadoras de la Universidad de Cádiz han desarrollado un sistema de inteligencia artificial capaz de detectar silbidos de cetáceos en el Estrecho de Gibraltar, uno de los entornos marinos más complejos por su intenso tráfico y ruido submarino.
Alba Márquez-Rodríguez retweeted
Not only birds!
Can AI detect cetacean movements whistles in one of the noisiest seas on Earth? Yes! 🐬 In our new paper, we show it’s possible using deep learning in the Strait of Gibraltar.
🔗 shorturl.at/kNI55
All code, data pipelines, and tools are open source to support reproducible and scalable PAM workflows.
Check it on GitHub!
Choosing the “optimal” ROC threshold is not always optimal for ecology.
We show how confidence threshold calibration can balance automation vs expert review depending on monitoring goals.
Baseline models collapsed under noise.
After iterative adaptation + threshold tuning, our best model reached:
➡️ F1 = 0.88
➡️ Recall = 0.91
With minimal extra annotation effort thanks to the semi-supervised approach!
Why the Strait of Gibraltar?
Because it is an important migratory bottleneck with:
🚢 Heavy maritime traffic
🌊 Complex geophony
🐬 Multiple cetacean species
A perfect stress test for real-world bioacoustic AI.
We combine:
• Transfer learning (BirdNET & Perch)
• Iterative, semi-supervised fine-tuning
• Expert-in-the-loop validation
• Confidence-threshold calibration
To adapt models to local soundscapes.
Passive Acoustic Monitoring is powerful, but real-world data is messy.
Most models can score >0.95 on clean datasets but also can completely fail under real noise data (F1 ≤ 0.10).
Accuracy alone is misleading.
Alba Márquez-Rodríguez retweeted
I'm thrilled to share with you our lattest publication with the brilliant @GrunCrow as the first author. We propose here a two-stage pipeline for analyzing audios from a acoustic monitoring programme in Doñana; 1) using a detector of vocalizations, 2) a fine-tuned classificator.
First week of June I went to the SIBECOL & AEET meeting in Pontevedra
I presented advances from #SEANIMALMOVE on using ecoacoustics for bird & marine mammal monitoring
Some results are saved for upcoming papers — but the talk still won Best Multidisciplinary Predoc Oral Comm!
All data and methods are shared and publicly available:
📰Paper: arxiv.org/abs/2503.15576
📁Dataset: huggingface.co/datasets/Grun…
🤖Simplified Repository: github.com/GrunCrow/Bird-Son…
🤖Main Repository: github.com/GrunCrow/BIRDeep_…
General models like BirdNET are powerful, but local adaptation is key, but sometimes even fine-tuning or custom classifiers are not enough.
By separating detection and classification, we boost precision and reduce noise in PAM workflows.
We also checked and compared BirdNET as a detector with our Bird Song Detector, our detector detected more birds with fewer false positives.
📷 FNs dropped 67%
📷 FPs stayed under 5%
The main innovation is that we trained a YOLOv8-based Bird Song Detector to filter only segments with bird vocalizations before running species classifiers.
The results are that for all classifiers, metrics improve when the Bird Song Detector is first applied:
🎙️461 minutes manually annotated
📊3,749 annotated vocalizations (Frequency + Time of vocalization)
🐦34 classes (29 species-specific)
It took 224 hours to annotate this data. For our Bird Song Detector we only used temporal annotations.
📁huggingface.co/datasets/Grun…
