Wheat phenology detection with the methodology of classification based on the time-series UAV images

Near real-time crop phenology information can offer significant guidance for the implementation of crop man-agement. Previous approaches to crop phenology detection have relied on time-series vegetation index curves, which can only be formed after the end of the whole phenology. To overcome the lag problem in phenology estimation, this study treats phenology detection as a classification problem based on imaging from an Un-manned Aerial Vehicle (UAV). Wheat field trials over two experimental seasons involved different sowing dates, nitrogen (N) rates, and wheat cultivars. A feature selection algorithm based on the compactness-separation principle (FS-CS) was used to filter the spectral and texture features extracted from time-series UAV images. The multi-level correlation vector machine (mRVM) was used to classify the principal phenological stages, including emergence, tillering, jointing, booting, and heading anthesis, filling, and maturity stages. The results showed that the classification accuracies of each stage were 0.86, 0.87, 0.31, 0.61, 0.22, 0.25, 0.77 and 0.93, respectively. Furthermore, the combination of spectral features and texture features has been proven to compensate for each other's deficiencies, and the overall accuracy obtained using two features together increased by 27 % and 13 %, respectively. Finally, the efficiency of the feature selection algorithm and classifier used in this study were discussed. The best estimation results were generated using FS-CS and mRVM when the optimal number of features was small. This research provides theoretical support for instantaneous detection of crop phenology based on remote sensing and imaging technology, and also provides technical guidance for efficient real-time discrimination of crop phenology using mono-temporal UAV imagery.

Automated summary

This study proposes a classification method based on UAV imagery for crop phenology detection. The results show that the combination of spectral and texture features can improve classification accuracy, providing technical guidance for real-time detection of crop phenology.