Smartphone-based Cyanobacteria Bloom Early-Warning System via Image Analysis and Crowdsourcing

Thesis Abstract

The proliferation of cyanobacterial blooms in freshwater bodies poses significant threats to public health, ecosystem services, and local economies, yet real-time monitoring remains fragmented due to resource constraints and spatial gaps in conventional sensor networks. This study proposes a smartphone-based early-warning system that leverages image analysis and crowdsourcing to detect and forecast cyanobacterial blooms, addressing the need for accessible, scalable, and timely information to inform risk management and public communication. The aim is to develop and validate a mobile platform that (i) processes user-captured water images to identify bloom indicators using computer vision and machine learning, (ii) aggregates crowdsourced observations with environmental metadata to enhance spatiotemporal sensitivity, and (iii) integrates with a probabilistic warning model to issue threshold-based alerts to stakeholders. Specific objectives include (a) constructing a labeled image dataset of lake and reservoir phenotypes (n ? 8,000 images) annotated for cyanobacterial presence, pigment concentration proxies, and turbidity; (b) developing a lightweight on-device convolutional neural network (CNN) for real-time bloom likelihood estimation with an accuracy target of F1 score ?0.85; (c) designing a crowdsourcing workflow that validates user reports through cross-verification, geo-tagging, and temporal sequencing; (d) calibrating a Bayesian hierarchical model that fuses image-derived features, meteorological data, and water quality indicators to generate probabilistic bloom warnings; and (e) evaluating system performance against traditional monitoring datasets from three public reservoirs over a 12-month period. The study adopts a pragmatic research design combining software engineering, computer vision, and statistical modelling within a transdisciplinary framework influenced by the Theory of Planned Behavior and the Technology Acceptance Model to understand user adoption and data quality. Data collection comprises (i) a stratified sample of public reservoirs with historical bloom events, (ii) 8,000 curated images collected via citizen scientists using the mobile app, and (iii) in-situ water quality readings (phycocyanin, microcystin, chlorophyll-a, dissolved oxygen, temperature) from partner laboratories (n ? 1,200 tuples). Instruments include a validated image annotation schema, a mobile data capture form, and laboratory-verified water quality meters. Data analysis employs a multi-method approach computer vision techniques (transfer learning with ResNet-50, data augmentation, edge-enhanced segmentation) to derive bloom features; statistical validation using confusion matrices, ROC-AUC, and kappa statistics to assess annotation reliability; Bayesian data fusion to integrate heterogeneous data streams; and time-series analysis (ARIMA/Prophet) to examine lead times between predicted and observed bloom events. The anticipated findings indicate that image-derived pigment proxies, when coupled with crowdsourced corroboration and meteorological context, can achieve bloom warnings with lead times ranging from 24 to 72 hours, depending on reservoir characteristics. The Bayesian model is expected to reduce false alarm rates below 15% while maintaining a true-positive rate exceeding 80%, and the on-device CNN will maintain inference times under 1 second per image. The study contributes to knowledge by (i) demonstrating the viability of participatory sensing and mobile vision for environmental risk monitoring, (ii) advancing the methodological integration of computer vision and probabilistic data fusion in aquatic ecology, and (iii) providing a scalable blueprint for low-cost, citizen-enabled cyanobacteria bloom surveillance applicable to water-management agencies and at-risk communities. Practical implications include improved situational awareness for water providers, enhanced public risk communication, and support for timely decision-making regarding water withdrawals, recreational advisories, and nutrient management strategies. The main conclusion anticipates that a smartphone-based, crowdsourced, vision-enabled early-warning system can complement conventional monitoring by offering rapid, geographically dense, and user-driven data streams, with recommendations emphasizing ongoing model recalibration, data quality controls, user training, and integration with existing water-quality dashboards to sustain long-term adoption and impact.