High resolution cropland mapping integrating deep semantic segmentation with Planet satellites

Abstract:

Abstract: Satellite remote-sensing technology is an essential means of acquiring farmland information, thus a powerful tool for rapid and accurate monitoring of large-scale farmlands. Traditional methods, based on shape, spectrum and texture features, utilize pixel-based or object-oriented algorithm to achieve farmland and crop classification. However, due to the complexity of spectral feature and the lack of high-resolution satellite images, the traditional methods featured by relatively low classification accuracy and low resolution in farmland mapping. High-resolution farming mapping has not yet been achieved and adopted. In addition, the traditional methods are highly dependent on expert knowledge and could induce large uncertainty. By introducing end-to-end learning theory and automatically extracting the most remarkable feature of meter-scale satellite images, the recently developed deep neural network can largely reduce labor cost, and greatly improve resolution and accuracy. To fill the research gap, we build a deep semantic segmentation model on the Google Earth Engine platform, in which the high-resolution images provided by the Planet satellites are used. Our study is of significance for agricultural strategies and management, and may contribute to the alleviation of food security and the sustainable development of agricultural sector.