The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLVIII-1-2024
https://doi.org/10.5194/isprs-archives-XLVIII-1-2024-799-2024
https://doi.org/10.5194/isprs-archives-XLVIII-1-2024-799-2024
11 May 2024
 | 11 May 2024

Crop Phenology Estimation in Rice Fields Using Sentinel-1 GRD SAR Data and Machine Learning-Aided Particle Filtering Approach

Jie Yang, Hongtao Shi, Qinghua Xie, Juan M. Lopez-Sanchez, Xing Peng, Jianghao Yu, and Lei Chen

Keywords: Crop Phenology, Particle Filtering, Sentinel-1, Real-time, Machine-learning

Abstract. Monitoring crop phenology is essential for managing field disasters, protecting the environment, and making decisions about agricultural productivity. Because of its high timeliness, high resolution, great penetration, and sensitivity to specific structural elements, synthetic aperture radar (SAR) is a valuable technique for crop phenology estimation. Particle filtering (PF) belongs to the family of dynamical approach and has the ability to predict crop phenology with SAR data in real time. The observation equation is a key factor affecting the accuracy of particle filtering estimation and depends on fitting. Compared to the common polynomial fitting (POLY), machine learning methods can automatically learn features and handle complex data structures, offering greater flexibility and generalization capabilities. Therefore, incorporating two ensemble learning algorithms consisting of support vector machine regression (SVR), random forest regression (RFR), respectively, we proposed two machine learning-aided particle filtering approaches (PF-SVR, PF-RFR) to estimate crop phenology. One year of time-series Sentinel-1 GRD SAR data in 2017 covering rice fields in Sevilla region in Spain was used for establishing the observation and prediction equations, and the other year of data in 2018 was used for validating the prediction accuracy of PF methods. Four polarization features (VV, VH, VH/VV and Radar Vegetation Index (RVI)) were exploited as the observations in modeling. Experimental results reveals that the machine learning-aided methods are superior than the PF-POLY method. The PF-SVR exhibited better performance than the PF-RFR and PF-POLY methods. The optimal outcome from PF-SVR yielded a root-mean-square error (RMSE) of 7.79, compared to 7.94 for PF-RFR and 9.1 for PF-POLY. Moreover, the results suggest that the RVI is generally more sensitive than other features to crop phenology and the performance of polarization features presented consistent among all methods, i.e., RVI>VV>VH>VH/VV. Our findings offer valuable references for real-time crop phenology monitoring with SAR data.