Multi-Year Crop Type Mapping Using Sentinel-2 Imagery and Deep Semantic Segmentation Algorithm in the Hetao Irrigation District in China

• Published in: Remote sensing (Basel, Switzerland) Vol. 15; no. 4; p. 875
• Main Authors: Li, Guang, Han, Wenting, Dong, Yuxin, Zhai, Xuedong, Huang, Shenjin, Ma, Weitong, Cui, Xin, Wang, Yi
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.02.2023
• Subjects: Accuracy; Agricultural management; Agricultural production; Agriculture; Algorithms; Analysis; Cash crops; Cereal crops; Clouds; Construction; Corn; Crops; Cultivated lands; Data mining; Deep learning; deep semantic segmentation; global separability index; Image acquisition; Image processing; Image segmentation; Irrigation; Local government; Machine learning; Mapping; Methods; multi-year crop mapping; Multiphase; Neural networks; Regression analysis; Remote sensing; Satellites; Semantic segmentation; Semantics; Sentinel-2 imagery; Spatial distribution; Time series; U-Net; Wheat; China; Mongolia
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs15040875

Accurately obtaining the multi-year spatial distribution information of crops combined with the corresponding agricultural production data is of great significance to the optimal management of agricultural production in the future. However, there are still some problems, such as low generality of crop type mapping models and susceptibility to cloud pollution in large-area crop mapping. Here, the models were constructed by using multi-phase images at the key periods to improve model generality. Multi-phase images in key periods masked each other to obtain large-area cloud-free images, which were combined with the general models to map large areas. The key periods were determined by calculating the global separation index (GSI) of the main crops (wheat, maize, sunflower, and squash) in different growth stages in the Hetao Irrigation District (HID) in China. The multi-phase images in the key period were used to make the data set and were then combined with a variety of deep learning algorithms (U-Net, U-Net++, Deeplabv3+, and SegFormer) to construct general models. The selection of the key periods, the acquisition of regional cloud-free images, and the construction of the general crop mapping models were all based on 2021 data. Relevant models and methods were respectively applied to crop mapping of the HID from 2017 to 2020 to study the generality of mapping methods. The results show that the images obtained by combining multi-phase images in the key period effectively avoided the influence of clouds and aerosols in large areas. Compared with the other three algorithms, U-Net had better mapping results. The F1-score, mean intersection-over-union, and overall accuracy were 78.13%, 75.39% and 96.28%, respectively. The crop mapping model was applied to images in 2020, and its average overall accuracy was more than 88.28%. When we applied the model to map crops (county food crops, cash crops, and cultivated land area) from 2017 to 2019, the regression analysis between the mapping areas obtained by the model and the ground measurements was made. The R2 was 0.856, and the RMSE was 17,221 ha, which reached the application accuracy, indicating that the mapping method has certain universality for mapping in different years.