PSeqNet: A crop phenology monitoring model accounting for phenological associations

Abstract

Variations of crop phenology are critical indicators of growth conditions and are essential for scheduling irrigation and fertilization activities to mitigate climate risks. Accurately characterizing carry-over climate impacts and phenological associations, especially the delayed influence of earlier stages development on later stages, is key to understanding crop phenological dynamics under changing climate. However, current remote sensing methods face challenges in matching extracted phenological metrics to crop phenological stages and in exploring complex climate interactions. To address these challenges, we propose a novel data-driven phenology monitoring algorithm named Phenology Seq2Seq Network (PSeqNet) to account for underlying phenological associations using fused remote sensing and meteorological data. A two-stream encoder processes and fuses temporal changes of remotely sensed and meteorological information during the growing season, followed by an autoregressive phenological decoder that utilizes the hierarchical structure of phenological development to learn associations among stages. PSeqNet is applied in Northeastern China as a case study to detect and forecast multiple rice phenological stages at the station level. The results indicate that PSeqNet with a two-stream encoder effectively utilizes fused information and extracts distinct associations among stages through its autoregressive decoder. PSeqNet outperformed generic curve fitting and shape model fitting methods in terms of overall accuracy and the degree of correlation, with overall mean absolute error (MAE) ranging from 3.3 to 4.0 days and correlation coefficient (r) ranging from 0.56 to 0.66. Further analysis highlights that the PSeqNet’s ability to capture unique phenological associations such as the weaker correlation between heading and tillering, and the stronger linear correlation between milking and heading. These distinct associations among stages cannot be adequately characterized by curve-fitting and shape model fitting methods. By utilizing the partial seasonal observations, the PSeqNet model also exhibits an outstanding performance in within-season forecasting in a progressive manner (overall MAE decreased from 4.8 to 4.1 days for maturity). Our findings indicate that the PSeqNet presents a promising approach for representing the phenological associations and provides a flexible approach for integrating multi-source information. This robust phenological monitoring approach holds great potential for identifying crop phenological association patterns and their driving climate factors across broader regions.