@article{WANG2023, 
author = {JiaYue WANG and ZhiWen CAI and WenJing WANG and HaoDong WEI and Cong WANG and ZeXuan LI and XiuNi LI and Qiong HU},
title = {Integrating Multi-Source Gaofen Images and Object-Based Methods for Crop Type Identification in South China},
year = {2023},
journal = {Scientia Agricultura Sinica},
volume = {56},
number = {13},
pages = {2474-2490},
keywords = {GF data, object-oriented, random forest, remote sensing identification, crop classification},
url = {https://www.sciopen.com/article/10.3864/j.issn.0578-1752.2023.13.004},
doi = {10.3864/j.issn.0578-1752.2023.13.004},
abstract = {【Objective】Due to frequent cloudy and rainy weather, it is challenging to map crop types over South China with fragmented agricultural landscapes. The Gaofen (GF) series satellites developed by China have high spatial-temporal resolution and good image quality. This study exploited the spatial and temporal advantages of multi-source GF images for fine crop classification in heterogeneous agricultural areas with frequent clouds and rain.【Method】This study characterized the spatial geometry of agricultural fields based on GF-2 data with high spatial resolution, and cooperated with the encrypted observations from GF-1 and GF-6 time series to fully characterize the spectral seasonal variations of crops. By constructing three-dimensional classification features of spectrum-time-space, the crop classification based on random forest classifier were conducted, and the importance scores of different features were calculated. Furthermore, several classification scenarios were set based on different satellite combinations and different classification units, for further analyzing the performances of integrating different GF datasets on crop type mapping.【Result】The overall accuracy of synergistic GF-1, GF-2 and GF-6 by object-oriented crop classification was 95.49% with Kappa of 0.94 in Qianjiang city; the overall accuracy in Zaoyang city was 93.78% with Kappa of 0.92. The accuracy of crop classification by integrating GF-2 and GF-6 was higher than that by GF-2 and GF-1. In addition, the object-oriented crop classification based on GF-2 outperformed the pixel-oriented one, in which the overall accuracy improved by 1.4% and 1.32% in Qianjiang and Zaoyang, respectively. Compared with the spectral and spatial features of GF-1 and GF-2, the GF-6 spectral bands had the largest contribution to crop type identification, and the cumulative importance score accounted for 82% (Qianjiang) and 77% (Zaoyang) of all spectral bands. Among them, the four new spectral bands of GF-6, namely, red-edge I band (B5), red-edge Ⅱ band (B6), purple band (B7) and yellow band (B8), accounted for 47% (Qianjiang) and 33% (Zaoyang) of all spectral bands for crop type identification.【Conclusion】Integrating multi-source GF images by taking advantages of their spectral, spatial and temporal features could not only alleviate the "mixed pixel" problem caused by the fragmented agricultural landscape, but also reduce the uncertainty of crop type identification in cloudy and rainy areas, providing great potential for accurate crop mapping in South China.}
}