An interactive and iterative method for crop mapping through crowdsourcing optimized field samples

Abstract

Remote sensing appears as an essential approach for crop mapping, yet the interpretation of satellite imageries requires for a large amount of labeled data as ground truth information. Traditional approaches for ground truth data collection are costly, inefficient and are mostly one-off effort, thus the up-to-date ground truth data are extremely scarce in existing crop mapping research and applications. Here, we address the challenge of ground truth data scarcity by implementing an interactive and iterative crowdsourcing framework. We developed a crowdsourcing platform, named as FarmWatch, which helped process one Sentinel-2 imagery for an unsupervised clustering and instantly for a stratified random sampling. Sample tasks and collected information are synchronized across web and mobile applications, which enables the field information to be immediately applied for crop mapping. Results of the iterative process show that: firstly, a total of 95 samples have been collected in the initial round and the overall accuracy of crop mapping was 83.33%; secondly, after four rounds of sample collection (a total of 279 samples), the classification accuracy reached to 96.30% and such accuracy did not improve even though an additional 23 samples were added in the fifth round. Such interactive and iterative mechanism indicates that ground truth data has been sufficiently collected for the current crop mapping activity. It not only promotes the opportunity of up-to-date and in-season crop mapping, but also helps to inform users where to collect field samples and how many samples are basically required for producing an accurate crop map.

Keywords

• Crowdsourcing
• Remote sensing
• Classification
• Crop mapping
• Sampling