Extraction of Cotton Information with Optimized Phenology-Based Features from Sentinel-2 Images

Abstract

The spatial distribution of cotton fields is primary information for national farm management, the agricultural economy and the textile industry. Therefore, accurate cotton information at the regional scale is required with a rapid increase due to the chance provided by the huge amounts of satellite images accumulated in recent decades. Research has started to introduce the phenology characteristics shown at special growth phases of cotton but frequently focuses on limited vegetation indices with less consideration on the whole growth period. In this paper, we investigated a set of phenological and time-series features with optimization depending on each feature permutation’s importance and redundancy, followed by its performance evaluation through the cotton extraction using the Random Forest (RF) classifier. Three sets of 31 features are involved: (1) phenological features were determined by the biophysical and biochemical characteristics in the spectral space of cotton during each of its five distinctive phenological stages, which were identified from 2307 representative cotton samples using 21,237 Sentinel-2 images; (2) three typical vegetation indices were functionalized into time-series features by harmonic analysis; (3) three terrain factors were derived from the digital elevation model. Our analysis of feature determination revealed that the most valuable discriminators for cotton involve the boll opening stage and harmonic coefficients. Moreover, both qualitative and quantitative validation were performed to evaluate the retrieval of the optimized features-based cotton information. Visual examination of the map exhibited high spatial consistency and accurate delineation of the cotton field. Quantitative comparison indicates that classification of RF-coupled optimized features achieves improved overall accuracy 5.53% higher than that which works with either the limited vegetation indices. Compared with all 31 features, the optimized features realized greater identification accuracy while using only about half the number of features. Compared with test samples, the cotton map achieved an overall accuracy greater than 98% and a kappa more than 0.96. Further comparison of the cotton map area at the county-level showed a high level of consistency with the National Bureau of Statistics data from 2020, with R2 over 0.96, RMSE no more than 14.62 Kha and RRMSE less than 17.78%.