Combining shape and crop models to detect soybean growth stages

Abstract

Accurately monitoring soybean growth stages (SGSs) is crucial for successful crop management and the development of agricultural information systems. This study focused on 18 states that accounted for over 95% of the soybean area in the United States from 2013 to 2020. We proposed an SMFs-APTT method that integrates crop data layers (CDLs), time-series of VIIRS data, and meteorological data. It combines a shape-model function in separate meteorological stages (SMFs) with a crop model that relies on an accumulated photothermal time (APTT). This approach provides both full-season and within-season monitoring of four soybean growth stages (SGSs); i.e., emergence, blooming, pod-setting, and leaf-dropping. Based on the results obtained from the SMFs-APTT method, a long short-term memory (LSTM) model was employed to predict SGSs early. The dates of the detected and predicted SGSs were compared with National Agricultural Statistics Service (NASS) Crop Progress and Condition Report (CPR) statistical data for analysis and verification. Our results show the following. (1) For the full-season extraction of SGSs, the average root mean square error (RMSE) derived using the SMFs–APTT method was 0.86–3.06 days and that obtained using the SMFs method was 1.56–3.26 days. (2) For the within-season monitoring of SGSs, the SMFs–APTT method was also able to accurately track the growth stages as early as ∼30 days after they reach 50% completion with an average RMSE of 2.1 days. (3) For the early prediction of SGSs, the LSTM model that was trained based on the SMFs–APTT results achieved an RMSE of ∼4.3 days at the state scale and could approximately predict SGSs ∼30 days in advance. Our findings suggest that the SMFs–APTT method provides accurate and reliable extraction of SGSs for full-season and within-season monitoring, which is of benefit to crop modeling and management. Furthermore, the LSTM model successfully forecast SGSs, indicating its potential for making early predictions.