Identification of garlic based on active and passive remote sensing data and object-oriented technology

Abstract: Garlic has been one of the most widely planted cash crops in Henan and Shandong provinces, China. The garlic plantation varies greatly in the frequently fluctuated price over the past few years, further dominating the decision-making on the planting for the healthy and sustainable development of the garlic market. However, there is a great challenge on the rapid and accurate extraction of garlic with a high precision, where the garlic crop planting is often mixed with other ground crops. Taking the Four-Huaiqing Chinese medicine base in Kaifeng City, Henan Province of China as a study area, this research aims to improve the identification accuracy and efficiency of fragmented garlic with the complex planting structure using remote sensing data. An object-oriented model was also proposed to integrate the active and passive remote sensing data of Sentinel satellites using the Google Earth Engine (GEE) platform and random forest (RF). The normalized difference vegetation index (NDVI) time-series data was selected to separate the phenological features of garlic and other land cover types. The maximum difference of NDVI and the features of RF were used to screen the monthly mean Sentinel-2 optical data in March 2021. A time series of Sentinel-1 synthetic aperture radar (SAR) backscattering coefficients were selected for the monthly mean data from November 2020 to May 2021. Two steps were then implemented before classification. The first one was to integrate the data segmentation using the simple non-iterative clustering (SNIC), further to select the best segmentation scale with the highest classification accuracy and Kappa coefficient. The second one was to select the best within three neighborhood values (4, 8, and 16) using the gray-level co-occurrence matrix (GLCM), thereby calculating the seven texture features of synthetic optical data, and finally to reduce the data dimensions using the principal components analysis (PCA). The red edge bands of vegetation were integrated with the Sentinel passive and active remote sensing data in the10 or 20 m spatial resolution. As such, the identification accuracy of garlic was improved to combine various groups of spectral features, backscattering coefficients, vegetation indexes, and different principal component groups of texture features. The result showed that the highest overall accuracy of classification and Kappa coefficient reached 94.54%, and 0.93, respectively, using the active and passive Sentinel remote sensing data in the 10 m spatial resolution, where the producer's and user's accuracies of garlic were 97.83% and 96.38%, respectively, at the SNIC segmentation scale of 5, and the GLCM neighborhood value of 4, as well as the first and second principal components of the 7 texture features. In the case of active and passive Sentinel remote sensing data with three vegetation red edge bands in the 20 m spatial resolution, the highest overall accuracy of classification and Kappa coefficient reached 94.14%, and 0.92, respectively, where the producer's and user's accuracies of garlic were 95.72% and 98.81%, respectively, at the SNIC segmentation scale of 3, and the GLCM neighborhood value of 4, as well as the first and second principal components of the 7 texture features. The producer's and user's accuracies in the 20m spatial resolution Sentinel-2 data with three vegetation red edge bands were improved 0.49% and 4.38%, respectively, compared with the 10m spatial resolution. In the time series of Sentinel-1 SAR data, the producer's and user's accuracies in 10 m spatial resolution were improved 0.66% and 2.03%, respectively, compared with the 20m. Correspondingly, the Sentinel active and passive remote sensing data can be effectively integrated to fully represent the spectral and structural information. Moreover, the overall accuracy and Kappa coefficient were much higher than those using the optical or time series SAR data alone, particularly the maximum in the 10 m high spatial resolution. Therefore, the SNIC, GLCM, PCA, RF, and GEE platforms can be widely expected to accurately and efficiently extract the garlic planting areas using GF satellites data.