Development of advanced tools for the early warning of cyanobacterial blooms from a multi-scale perspective

• Autores: Claudia Fournier Ubeda
• Directores de la Tesis: Antonio Quesada del Corral (dir. tes.), Samuel Cirés (dir. tes.)
• Lectura: En la Universidad Autónoma de Madrid ( España ) en 2025
• Idioma: inglés
• Número de páginas: 270
• Títulos paralelos:
  • Desarrollo de herramientas avanzadas para la alerta temprana de blooms de cianobacterias desde una perspectiva multi-escala
• Tribunal Calificador de la Tesis: José Luis Risco Martín (presid.), Elvira Perona Urízar (secret.), Lars Wörmer (voc.)
• Programa de doctorado: Programa de Doctorado en Microbiología por la Universidad Autónoma de Madrid
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Bancos de datos
    • Estadística
      • Análisis de datos
  • Ciencias de la vida
    • Microbiología
    • Biología vegetal (botánica)
      • Fisiología vegetal
• Enlaces
  • Tesis en acceso abierto en: Biblos-e Archivo
• Resumen

  • This thesis addresses the problem of cyanobacterial blooms through the development of monitoring tools that can support early warning systems. Cyanobacteria are photosynthetic microorganisms present in aquatic ecosystems that can proliferate massively and produce toxins, deteriorating water quality and limiting its use. Traditional monitoring methods are effective for characterizing cyanobacterial communities and their toxicity but present limitations in providing early warnings. For this reason, the present Doctoral Thesis aims to develop advanced monitoring tools that offer early alerts and support a more preventive and less reactive management approach. The introduction presents the theoretical framework necessary to understand the development of the chapters, including the problem and the state of the art regarding existing and novel monitoring tools. Chapter 1 studied the connectivity between a reservoir and an estuary using genetic analyses (16S rRNA gene metabarcoding and PCRs for toxic genes) during six sampling campaigns in 2022 and 2023. The results revealed that the reservoir serves as a source of potentially toxic cyanobacteria for the estuary. More than 50% of the cyanobacterial species in the estuary possibly originated from the reservoir, and over 80% of the transmitted ASVs belonged to potentially toxic genera. Genes for microcystins and anatoxins were detected, although toxin concentrations were low. ASVs from the most relevant genera showed different transfer success rates, but in most cases, a dominant genotype was identified, leading the transmission of the genus and its presence in the estuary. This finding presents an opportunity for targeted monitoring and management. Chapter 2 applied machine learning and deep learning models to predict blooms using data from multiparametric probes in the Cuerda del Pozo reservoir (2016–2021). Variables such as temperature, chlorophyll a, and phycocyanin were used. The data were preprocessed to create biologically optimized time series, and linear models, Random Forest, and LSTM (Long Short-Term Memory) neural networks (both autoregressive and multivariate) were trained and evaluated with a hybrid metric system. The multivariate LSTM models (using phycocyanin and temperature) achieved over 90% accuracy in predicting alerts at 10 µg PC/L and delivered positive results up to 16 days in advance. Chapters 3 and 4 explored hyperspectral technology to identify potentially toxic cyanobacterial genera. In Chapter 3, a library of 140 hyperspectral images was created for five genera (Microcystis, Planktothrix, Aphanizomenon, Chrysosporum, and Dolichospermum) grown under different light and nutrient conditions to maximize spectral variability. Random Forest models trained with preprocessed spectral data achieved 90% classification accuracy, with Microcystis and Chrysosporum exceeding 95%. In Chapter 4, a library of 123 hyperspectral images was developed for Microcystis, Chrysosporum, and Dolichospermum, in both pure cultures and binary mixtures. Neural networks were trained, achieving 95% accuracy in genus detection and 90% accuracy in determining their proportions, identifying genera in proportions as low as 6%. Both chapters identified key wavelengths in the visible and near-infrared ranges. Finally, the discussion integrates the findings, highlighting opportunities and challenges in using AI and hyperspectral technology for early warnings, and analyzes the main barriers to their adoption in water management. The conclusions summarize the main contributions to the monitoring and management of cyanobacterial blooms