Spatial domain bridge transfer : An automated paddy rice mapping method with no training data required and decreased image inputs for the large cloudy area

• Published in: Computers and electronics in agriculture Vol. 181; p. 105978
• Main Authors: Zhang, Chengkang, Zhang, Hongyan, Zhang, Liangpei
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.02.2021; Elsevier BV
• Subjects: agriculture; Automation; Availability; electronics; food security; Machine Learning; Mapping; markets; Multi-temporal analysis; Paddy rice mapping; Phenological metrics; Phenology; Remote sensing; rough rice; Satellite imagery; satellites; time series analysis; Training; Phenological metrics; Paddy rice mapping; Multi-temporal analysis; Machine Learning
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105978

•We proposed a flexible operational framework to automatically map paddy rice.•Our experiments have revealed how many cloud-free images in cloudy regions were at least needed for automatic paddy rice mapping.•How to overcome the strong reliance on training data for machine learning-based methods for paddy rice extraction was fully explored.•Feature analysis revealed what kinds of phenological indicators were more effective in paddy rice mapping. Paddy rice mapping in an accurate and timely manner was of significance as the explicitly spatial paddy rice maps played a critical role in food security and market stability. Over the decades, remote sensing (RS) techniques offered great potential to extract paddy rice from the optical satellite images. Two main categories of paddy rice mapping methods were commonly utilized. The first one was the phenology-oriented methods by utilizing the distinctive phenology of paddy rice, and the machine learning-based approaches were the other mainstream for paddy rice extraction by being recognized as a classification task. However, neither of the above methods were actually applicable to automatically map paddy rice for the large cloudy areas. The phenological methods were subject to the availability of time-series cloud-free RS inputs due to cloud contamination, and machine learning-based methods were often confined to the availability of high-quality collected training samples. Therefore, this study proposed a new flexible framework to integrate the above two methods, in an attempt to achieve the goal of automatic paddy rice mapping for large cloudy area. Firstly, phenological paddy rice mapping was implemented for the areas with cloud-free time series satellite images available to generate the paddy rice potentials. Following that, the derived paddy rice potentials were purified through filtering operations. Finally, the machine learning classifier was trained using the purified samples, and spatially transferred to other regions for the extraction of paddy rice. Several experiments were carried out to assess our proposed framework and the subsequent mapping results. The experimental results were promising, indicating that our framework contributed to overcoming the strong reliance on training data for machine learning methods, and reducing the number of multi-temporal inputs for automated paddy rice mapping, which potentially provided a new paradigm for the purpose of paddy rice extraction.