Monitoring Rice Crop with Dense Segmentation on Satellite Images: A case study in Vietnamese Mekong Delta

Rice is the major crop in Vietnam with the crucial roles in national food security and export. While the demand of this crop is increasing overtime, the land use for this purpose is restricted due to various factors such as urban expansion and seawater intrusion. Therefore, a sustainable and effective rice paddy area monitoring is highly desirable, as the traditional methods are costly, inaccurate and requires heavy human-related workload. In this paper, we propose a novel rice monitoring framework using the satellite image, thanks to its high quality and free-to-access. Our framework first preprocesses the raw satellite image to alleviate the effect of adversarial factors such as geometric discrepancy and solar radiation. Then we design a deep learning model inspiring from u-net architecture to exploit simultaneously the spatial, spectral and temporal nature of satellite images. The empirical result conducted over data collected in Vietnamese Mekong Delta shows that our framework outperforms others state-of-the-art baselines and achieves the accuracy of 92.7% without the need of any handcraft feature engineering or expert knowledge.