Extraction of complex crop structure in the Hetao Irrigation District of Inner Mongolia using GEE and machine learning

Abstract: Hetao Irrigation District (HTID) has been the largest self-flowing irrigation district with one water intake in Asia, serving an important commercial grain and oil production base in China. The annual grain production in the HTID reached 2.55 million tons in 2018, accounting for 3.9‰ of the total crop cultivation area in China. Therefore, accurate and rapid extraction of crop structure can be of great practical significance in agricultural production for the food security of the HTID. However, it is difficult to distinguish the pixels of major crops in the remote sensing images, due to the severe soil salinization, fragmented and scattered crop distribution, as well as the same crop with the different spectrum of various crops. Moreover, there are the close growth periods of major crops in the HTID, which can mix the elements in the images. In this study, Sentinel-2 high-resolution remote sensing images and the GlobeLand30 dataset were used to extract the crop planting structure of the HTID using the Google Earth Engine cloud computing platform. Nearly 1 200 sample points were filtrated using the OTSU algorithm and Google Earth visual interpretation. The features of spectra, frequently-used vegetation, red-edge vegetation, and crop texture were input into four classifiers, including the Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), Classification and Regression Tree (CART). The Overall Accuracy (OA) and Kappa coefficient were used to evaluate the performances of model for the extraction of crop planting structure. Firstly, the impacts of classification features and classifiers combinations on the classification accuracy were explored to identify the classifier with the highest classification accuracy. Then, the feature optimization was performed on the five irrigation sub-districts using out-of-bag error rates for each irrigation sub-district. Finally, the optimal classifier and feature combinations were achieved to derive the cropping structure of four crops in the HTID in 2018. The results show that the RF classifier presented the highest classification accuracy using all feature bands, where the average OA of the HTID (81%) was 6 percentage points and 11 percentage points higher than that of the SVM and NB classifier, respectively. The Kappa coefficient reached 0.68, which was much higher than the rest. Furthermore, the importance of feature bands filtered by the RF was ranked first for the spectral features, the second for the vegetation features, and last for the gray texture features. The indexes were calculated using red-edge bands, indicating the better performance over the other commonly-used remote sensing vegetation indices in crop recognition. In addition, the feature-optimized scheme was the combination with the highest average OA of 86% and Kappa coefficient of 0.78, while the scheme containing 25 bands of spectral, vegetation and texture features presented an OA of 85% and Kappa coefficient of 0.75. Therefore, the new sights can be offered for extracting crop spatial distribution using remote sensing cloud computing platform in complex planting structure area. The finding can provide a strong reference to adjust the agricultural production structure, and further formulate the food macro-control policies in the Hetao Irrigation District.