Accurate and timely prediction of tea quality and yield at parcel level is a key enabler for precision management, climate adaptation, and economic optimization in tea production systems. Previous work has shown that a cadastral and tea production management system that integrates a geographic information system (GIS) with ground-based near-infrared (NIR) cameras and satellite imagery can estimate biochemical properties such as total nitrogen and fiber content using simple regression models on spectral indices. However, such approaches treat each observation as independent, ignore spatial and temporal dependencies, and provide limited support for uncertainty-aware decision making. In this paper, we propose TeaDeep-ST, a deep spatio-temporal learning framework that fuses multi-year IoT sensor data, multi-spectral satellite imagery, parcel-level GIS attributes, and weather records to predict tea quality and yield at fine spatial and temporal resolutions. Our model combines (i) temporal sequence encoders for parcel-wise time series, (ii) a graph neural network over the cadastral adjacency graph to capture spatial interactions, and (iii) a multi-task prediction head that outputs both point estimates and uncertainty intervals for multiple targets including total nitrogen, fiber, water content, and harvestable yield. We describe the system architecture, data integration pipeline, model design, and evaluation protocol using realistic tea plantation data, and we outline how TeaDeep-ST can be integrated into a GIS-based decision support system for harvest scheduling and resource allocation. Although our empirical evaluation is necessarily constrained by data availability, the results and ablation studies indicate that deep spatio-temporal learning substantially improves predictive accuracy over traditional index-based regression baselines and provides more informative uncertainty estimates for operational management.
Accurate and timely prediction of tea quality and yield at parcel level is a key enabler for precision management, climate adaptation, and economic optimization in tea production systems. Previous work has shown that a cadastral and tea production management system that integrates a geographic information system (GIS) with ground-based near-infrared (NIR) cameras and satellite imagery can estimate biochemical properties such as total nitrogen and fiber content using simple regression models on spectral indices. However, such approaches treat each observation as independent, ignore spatial and temporal dependencies, and provide limited support for uncertainty-aware decision making. In this paper, we propose TeaDeep-ST, a deep spatio-temporal learning framework that fuses multi-year IoT sensor data, multi-spectral satellite imagery, parcel-level GIS attributes, and weather records to predict tea quality and yield at fine spatial and temporal resolutions. Our model combines (i) temporal sequence encoders for parcel-wise time series, (ii) a graph neural network over the cadastral adjacency graph to capture spatial interactions, and (iii) a multi-task prediction head that outputs both point estimates and uncertainty intervals for multiple targets including total nitrogen, fiber, water content, and harvestable yield. We describe the system architecture, data integration pipeline, model design, and evaluation protocol using realistic tea plantation data, and we outline how TeaDeep-ST can be integrated into a GIS-based decision support system for harvest scheduling and resource allocation. Although our empirical evaluation is necessarily constrained by data availability, the results and ablation studies indicate that deep spatio-temporal learning substantially improves predictive accuracy over traditional index-based regression baselines and provides more informative uncertainty estimates for operational management.
