Design, implementation and evaluation of a functional plant-phenology monitoring system using low-cost sensors

Thesis Abstract

Plant phenology is a sensitive indicator of climate variability, yet traditional monitoring often relies on labor-intensive observations and expensive equipment, limiting scalable deployment across diverse ecosystems. This study addresses the gap in accessible, continuous phenological data by designing, implementing, and evaluating a functional monitoring system that leverages low-cost sensors to capture key phenophases (bud burst, leaf expansion, flowering, fruiting) under real-world field and greenhouse conditions. The aim is to develop a repeatable, sensor-driven workflow capable of delivering high-resolution phenology data with quantified uncertainty, suitable for integration with climate and ecosystem models. Specific objectives are to (1) identify a minimal set of affordable sensors (ambient light, temperature, humidity, soil moisture, and a low-cost time-lapse camera) and to calibrate phenological indicators against ground-truth observations; (2) implement an edge-computing data pipeline for real-time anomaly detection, data quality control, and storage in a relational database; (3) validate the system against conventional manual scoring across multiple taxa, including Arabidopsis thaliana, Prunus avium, and Helianthus annuus, in three contrasting environments (temperate greenhouse, outdoor field plot, and controlled growth chamber); (4) perform statistical and machine learning analyses to derive phenophase onset thresholds and phenology curves, and (5) evaluate system robustness, cost-benefit trade-offs, and scalability for broader deployment. The methodology adopts a mixed-methods design anchored in the theoretical framework of phenology monitoring and sensor fusion theory, with data triangulated across quantitative sensor outputs, time-lapse imagery, and expert manual ratings following the BBCH scale. The population comprises plant individuals from three functional groups (herbaceous, deciduous woody, and fruit crops) with a total of 180 specimens distributed across the three sites. A stratified sampling approach yields 60 plants per site, with equal representation of species to enable cross-taxa comparisons. Data collection instruments include low-cost sensors (DS1923 soil temperature sensors, DHT22/AM2302 for ambient conditions, TSL2591 light sensors, soil volumetric moisture sensors, and a 12 MP Raspberry Pi high-definition camera). A parallel human observer panel of five phenology experts provides reference annotations at biweekly intervals over a 12-month cycle. Instrument validity is established through calibration against a laboratory environmental chamber and USDA phenology references, while reliability is evaluated via inter-instrument replication and test-retest assessments. Data analysis proceeds in four stages. First, sensor data are preprocessed with outlier detection, missing-value imputation, and sensor fusion using Kalman filtering to generate continuous phenology signals. Second, image streams are analyzed with convolutional neural networks trained to classify phenophase states, with ground-truth labels from expert observers used to compute precision, recall, and F1-scores. Third, phenophase onset is estimated using hazard-function modeling and threshold-based methods, complemented by regression analysis to relate environmental drivers (temperature sums, photoperiod, soil moisture) to phenophase timing. Fourth, a comparative evaluation employs repeated-measures ANOVA to test differences between sensor-derived and manually observed onset dates across taxa and environments, and a cost-effectiveness analysis compares the low-cost system against conventional monitoring. The study also investigates uncertainty propagation through Bayesian inference to quantify confidence in phenophase estimates. Expected findings include high concordance between sensor-based onset dates and manual ratings (mean absolute error < 3 days for most taxa), robust phenology curves that track seasonal transitions, and clear relationships between thermal time accumulation and phenophase initiation. The low-cost system is anticipated to demonstrate favorable cost-per-sample metrics, with edge-computing yielding near-real-time alerts and acceptable data completeness (>92% across monitoring periods). The research is expected to identify species- and environment-specific sensor calibrations, revealing which measurements most strongly predict particular phenophases and where camera-based imagery adds incremental value. The study contributes to knowledge by validating a scalable, affordable phenology monitoring approach that integrates sensor fusion, image analytics, and classical phenology theory, enabling broader deployment in climate research, agriculture, and phenology-informed ecosystem management. It offers a replicable methodological blueprint for architects of agricultural and ecological sensing networks, including guidance on sensor selection, data governance, and analytical workflows. Practical recommendations include standardized calibration protocols, open-source software pipelines, and deployment guidelines for multi-taxa phenology networks, along with policy-relevant implications for crop phenology forecasting and climate adaptation strategies. The main conclusion posits that low-cost, integrated phenology monitoring is feasible and accurate enough for operational use, with recommendations emphasizing continual calibration, expanded species coverage, and investment in automated ground-truthing to sustain data reliability over time.