Chianese, Flavia Vittoria (2026) Analysis and classification of satellite images to measure melliferous and pollen potential for predicting ecosystem services. [Tesi di dottorato]

Documento PDF TESI_PHD_CHIANESE_FLAVIA_VITTORIA_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (4MB) | Richiedi una copia

Abstract

The decline of managed honey bee colonies (Apis mellifera L.) and the degradation of trophic resources for pollinators represent growing threats to food security and ecosystem services on a global scale. Understanding and predicting the environmental conditions that determine foraging activity requires integrative approaches capable of combining heterogeneous data sources across multiple spatial and temporal scales. The present doctoral thesis, developed within the framework of the HiveTechSpace project in collaboration with 3Bee S.r.l. and the CNR Institute for Sustainable Plant Protection (IPSP), proposes a multi-level methodological framework for the characterisation, quantification, and prediction of trophic resources available to honey bee in contrasting agroecological landscapes of southern Italy, integrating GIS, multi-source satellite remote sensing, IoT precision beekeeping, and machine learning algorithms. The first component of the framework addresses the landscape-scale quantification of honey and pollen potential across four study sites representing contrasting human impact gradients in Campania and Molise (Carditello, Presenzano, Cuccaro Vetere, Castel del Giudice). A purpose-built database of melliferous potential was developed and integrated with the InVEST Crop Pollination model, specifically recalibrated for honey bee. An original indicator of effective honey potential (PMeff) was introduced, incorporating distance-weighted accessibility from the apiary and topographic correction based on DEM data. The results demonstrate that topographic factors profoundly alter the landscape suitability ranking: Cuccaro Vetere, which presents the highest distance-weighted potential (38,556 kg), drops to the lowest effective potential (8.8 kg·ha⁻¹) due to the energetic cost of uphill return flights, while Castel del Giudice is the only site achieving an overall efficiency factor above unity (EFFtotal = 1.077). A comparative analysis of Copernicus-derived cartography against field-validated land use maps at Presenzano documents a systematic overestimation of 24.2% in PMeff, attributable to insufficient thematic resolution in discriminating melliferous tree species within mixed orchard classes. The second component evaluates the feasibility of training a Random Forest classifier for the identification of melliferous plant species at genus level from multitemporal Sentinel-2 satellite features, using heterogeneous training datasets derived from the LUCAS 2022 survey (filtered for the Mediterranean bioclimatic region) and iNaturalist citizen-science observations. A feature extraction pipeline adapted from Ghassemi et al. (2024) was implemented in Google Earth Engine, integrating optical indices (Sentinel-2), radar backscatter (Sentinel-1), land surface temperature (MODIS), and elevation (ASTER GDEM). A preliminary test on iNaturalist observations for the entire Italian territory (8 classes, n = 17,716) yielded an overall accuracy of 48% (κ = 0.30), with performance severely degraded for minority classes due to a maximum class imbalance ratio of 11:1. The main test on the combined LUCAS + iNaturalist dataset filtered for the Mediterranean bioclimatic region (23 classes, n = 32,363) produced an overall accuracy of 43% (κ = 0.37). The higher Kappa despite lower overall accuracy reflects a more balanced class distribution when LUCAS data are incorporated. The results identify three main methodological bottlenecks: class imbalance, GPS positional accuracy of citizen-science observations, and spectral overlap between deciduous species with similar spring phenology. These findings motivate the adoption of continuous satellite vegetation indices as a phenology-based alternative to discrete taxonomic classification for trophic resource estimation. The third component develops and comparatively validates a multiclass machine learning classifier for the daily prediction of honey bee foraging quality, defined on the basis of daily hive weight dynamics continuously recorded via IoT sensors (target variable: 0 = poor, 1 = neutral, 2 = good; thresholds: ±0.3 kg·day⁻¹). The integrated dataset (4,321 daily observations, eight hives, three sites, July 2022 – November 2025) combines ERA5-Land meteorological reanalysis, Sentinel-2 spectral indices as phenological proxies for floral resource availability, IoT hive weight variables, and trigonometric seasonality encoding. The pipeline incorporates systematic data leakage prevention, PCA dimensionality reduction (95% variance threshold), SMOTE class balancing, and Bayesian hyperparameter optimisation via Optuna (50 trials per model). Seven algorithms from distinct families were benchmarked. In the baseline configuration, Random Forest achieved the best overall performance on the test set (F1 = 0.716; AUC = 0.812). Following Bayesian optimisation, XGBoost emerged as the best-performing optimised model (F1 = 0.714), followed by LightGBM (F1 = 0.713), Random Forest (F1 = 0.711), and Gradient Boosting (F1 = 0.710). The optimisation primarily improved the performance of Gradient Boosting (+0.042 in F1), but did not surpass the baseline performance of Random Forest. These results confirm the suitability of ensemble tree-based methods for multi-variate ecological classification problems characterised by non-linear interactions between heterogeneous features. Collectively, the three components constitute a scalable, reproducible framework for multi-level environmental suitability assessment for honey bee foraging, with direct applicability to apiary management, landscape planning for pollinator conservation, and the development of decision support systems integrating remote sensing and precision beekeeping data.

Downloads

Actions (login required)

Modifica documento