A novel approach to detect the spring corn phenology using layered strategy

Accurate and continuous crop phenology information at a regional scale is important for agronomic management and yield estimation. However, detecting continuous crop phenology remains challenging due to the low sensitivity of remote sensing signals to certain phenological stages and the limited of availability remote sensing images. Therefore, this study developed a layered strategy to detect continuous crop phenology. First, a novel Phenology Separability Index (PSI) is established to select features from the Gaussian probability density distribution. PSI quantifies the capability of optical vegetation indexes (VIs), Synthetic Aperture Radar (SAR) signals, and meteorological factors to distinguish between various phenological stages. Then, the multi-temporal sample is established to enhance training sample representation and quantity. Finally, a random forest model is trained using features extracted from multi-temporal samples to improve detection accuracy. This model effectively reduces phenological stage confusion due to redundant features and limited samples. In addition, this study validated its extensibility by mapping crop phenology in the cities of Acheng, Zhaozhou, Lishu, and Buxin and assessed its uncertainty using the Sobol approach. Results indicated that growing degree day has the highest separability among meteorological factors, surpassing both SAR singles and optical VIs. Moreover, the proposed layered strategy was robust, explaining 96% of spatial variation in crop phenology at the regional scale. The accuracy of the layered strategy method (total RMSE = 8.74 days) surpassed that of the multi-temporal sample method (total RMSE = 15.76 days) and the traditional method with a single-temporal sample (total RMSE = 17.21 days). In addition, this study indicated that optical VIs are prone to confuse with the early or late phenological stage of corn, whereas SAR singles are highly sensitive to jointing date.