A Multi-Source Artificial Intelligence Framework for Non-Invasive Peanut Maturity Mapping Using Aerial Spectral Imagery and Weather Data

Abstract

Virginia-type peanut (Arachis hypogaea L.) maturity assessment is a critical challenge in precision agriculture because peanut pods develop belowground and mature asynchronously within the same plant and field. Conventional methods such as pod blasting, hull scrape evaluation, and maturity profile board classification remain important for harvest decisions, but they are destructive, labor intensive, spatially limited, and dependent on expert judgment. This dissertation developed a complete research chain for Virginia-type peanut maturity assessment, beginning with field sampling and ground reference data collection, progressing through the evaluation of field and surrounding agroclimatic factors, advancing into remote sensing and machine learning based maturity prediction, and culminating in the implementation of trained models as a deployable decision support tool for precision harvest planning. The dissertation was organized around six connected objectives. First, current traditional, sensing based, and artificial intelligence driven approaches for peanut maturity assessment were reviewed to identify major limitations and future needs. Second, field maturity dynamics were quantified across cultivars, locations, years, growth regulator treatments, research trials, and commercial grower fields. Third, growing degree day relationships with peanut maturity were evaluated under humid subtropical environments to understand the influence of accumulated agroclimatic conditions. Fourth, aerial multispectral imagery, vegetation indices, and accumulated growing degree days were integrated into a single-view machine learning framework to establish a baseline for non-destructive maturity prediction. Fifth, a multi-view stacked ensemble learning framework was developed to preserve and combine spectral, vegetation, and weather information for improved cultivar specific prediction and spatial mapping. Sixth, the resulting top models were implemented in a web based expert decision support system to generate peanut maturity index maps, days to maturity estimates, harvest readiness summaries, and automated reports. Field studies were conducted across Virginia and North Carolina during the 2022 to 2024 growing seasons using breeder seed trials, peanut variety quality evaluation (PVQE) trials, and commercial grower fields. Virginia-type cultivars including Bailey-II, Emery, NC-20, Sullivan, and Walton were evaluated through systematic pod sampling, pod blasting, mesocarp color classification, and peanut maturity index quantification. Maturity progression was strongly influenced by days after planting, cultivar, location, and environmental factors. Optimal maturity generally occurred between 140 and 150 days after planting for Bailey-II, Emery, NC-20, and Sullivan, while Walton showed later and more variable maturity. Growing degree day (GDD) analyses showed moderate to strong relationships with maturity, and the inclusion of additional weather factors such as soil temperature, relative humidity, rainfall, and photosynthetically active radiation provided broader agroclimatic context for interpreting maturity progression under humid subtropical conditions. Modeling efforts translated field and agroclimatic understanding into predictive maturity estimation. Single-view machine learning framework demonstrated that combined spectral reflectance, vegetation indices, and accumulated growing degree days could estimate peanut maturity, although performance varied by cultivar and data split. The multi-view stacked ensemble learning framework improved this approach by modeling each information domain separately and then combining their predictions through meta learning. This enabled more interpretable and spatially explicit prediction of peanut maturity index and days to maturity. Finally, these models were operationalized in the Peanut Readiness Evaluation Platform, a web based R Shiny expert decision support system that supports data ingestion, preprocessing, vegetation masking, feature extraction, model inference, interactive spatial visualization, and report generation. Overall, this dissertation advances peanut maturity assessment from manual field sampling alone to an integrated, data driven, and deployable decision support framework for precision harvest management in Virginia-type peanut production systems.