Identification and mapping of yellow-flowering rapeseed fields by combining social media data, Sentinel-2 imagery, deep learning algorithm, and Google Earth Engine

Abstract

Rapeseed cultivation on winter-fallow fields enables a seasonal switch between agricultural and bioenergy output. In view of the spectral features during the rapeseed flowering period (RFP), numerous remote sensing studies identified and produced the maps of rapeseed fields. The RFP is frequently identified by field surveys, visual imagery interpretation, or empirical crop knowledge in a specific region, none of which is appropriate for large-scale rapeseed identification and mapping. In this research, we combine social media data on the RFP fields, Sentinel-2 imagery, and deep learning algorithm to identify and produce maps of yellow-flowering rapeseed fields. In the context of citizen science and crowdsourcing, the social media data on the fields is regarded as reference data and utilized to generate the spatial distribution of the RFP and select temporal Setninel-2 imagery. We develop and evaluate the novel method in Anhui Province, China. The resultant rapeseed map in 2018 has a user’s accuracy and a producer’s accuracy of 0.93 and 0.99, respectively. To test the generalization of the knowledge-based method, we apply the deep neural network (DNN) model trained in Anhui Province to produce the maps of yellow-flowering rapeseed fields in Hubei Province and Shaanxi Province, China. The overall accuracy of the resultant rapeseed maps for Hubei Province and Shaanxi Province is 0.97 and 0.95, respectively. The proposed method provides a new way to produce rapeseed maps on a large scale, which could be used to support agricultural planning and ecological system management.

Keywords

• Rapeseed mapping
• Deep learning
• Sentinel-2
• Google Earth Engine