Cumulative Link Mixed-Effects Models in the Service of Remote Sensing Crop Progress Monitoring

Abstract: This study introduces an innovative Cumulative Link Modeling approach to monitor crop progress over large areas using remote sensing data. The models utilize the predictive attributes of calendar time, thermal time, and the Normalized Difference Vegetation Index (NDVI). Two distinct issues are tackled: real-time crop progress prediction, and completed season fitting. In the context of prediction, the study presents two model variations, the standard one based on the Multinomial distribution and a novel one based on the Multivariate Binomial distribution. In the context of fitting, random effects are incorporated to capture the inherent inter-seasonal variability, allowing the estimation of biological parameters that govern crop development and determine stage completion requirements. Theoretical properties in terms of consistency, asymptotic normality, and distribution-misspecification are reviewed. Model performance was evaluated on eight crops, namely corn, oats, sorghum, soybeans, winter wheat, alfalfa, dry beans, and millet, using in-situ data from Nebraska, USA, spanning a 20-year period. The results demonstrate the wide applicability of this approach to different crops, providing real-time predictions of crop progress worldwide, solely utilizing open-access data. To facilitate implementation, an ecosystem of R packages has been developed and made publicly accessible under the name Ages of Man.