Large-scale crop mapping from multi-source optical satellite imageries using machine learning with discrete grids

Abstract

The spatial distribution of crops is an important agricultural parameter, which is used to derive important information about crop productivity and food security. However, crop mapping on a large scale is challenging due to the low spatio-temporal information of satellite data, sparse sampling, and poor computational efficiency for massive data. To alleviate these problems, this study proposes a method based on discrete grids with machine learning to integrate GaoFen-1 and Sentinel-2 imagery. First, the proposed method fuses multi-source satellite data with similar observation characteristics to improve the spatial and temporal coverage of satellites. Second, a data augmentation technique based on a discrete grid framework was proposed to solve the problem of sparse samples. Finally, a machine learning algorithm in a discrete grid was introduced to improve processing efficiency and ensure the crop classification precision of large-scale remote sensing images. An experiment in the Sanjiang Plain area (approximately 108900 km2) of Northeast China showed that the proposed scheme benefited from a high spatio-temporal multi-source dataset and achieved good performance. Compared with a single data source, the accuracy of crop mapping using multi-source optical remote sensing data is higher, attaining up to 86 and 88 % in 2017 and 2018, respectively. Furthermore, the advantages of machine learning in discrete grids over large-scale areas are validated by evaluating the accuracy of different classifiers, which indicates the suitability of discrete grids in data augmentation and large-scale crop mapping. Finally, discrete grid technology offers a possibility for crop mapping over large-scale areas, and improves the processing efficiency of remote sensing big data. The findings in this study can contribute to studies on large-scale crop classification and serve as a reference to them.