**1. Introduction**

The Xinjiang Uygur Autonomous Region is a major agricultural region in the arid and semi-arid areas of Northwest China. Due to the dry climate, almost all agriculture in Xinjiang depends on irrigation, leading to water shortage. This region relies on large-area cotton cultivation for profit, with cotton production accounting for over 70% of the national total. Cotton planting consumes large amounts of water, exacerbating the problem of water scarcity. Some areas in Xinjiang have undergone structural adjustments of agriculture, reducing the cultivation area of cotton and expanding the planting scale of two other cash crops, i.e., chili pepper and tomato. These adjustments resulted in a more complex cropping structure, requiring the timely and accurate mapping of crop distribution. Crop type distribution is vital information for estimating water availability and environmental carrying capacity. This is especially important in the arid and semi-arid areas in Northwest China, where oasis agriculture is the economic pillar, while the ecological environment is relatively fragile.

Optical remote sensing has been widely used in agricultural area mapping and crop classification in recent years. Approaches utilizing MODIS (Moderate Resolution Imaging Spectroradiometer) vegetation indices for crop type discrimination only suit for large-open fields, due to the low resolution (250–500 m) of MODIS data [1–3]. A number of studies used Landsat spectrum and vegetation indices for crop mapping, but the data availability is heavily limited by cloud cover due to Landsat's 16-day revisit interval [4]. Landsat data also encounters mixed-pixel problems in heterogeneous smallholder farming areas. In addition, crop type discrimination places a higher demand on the spectral resolution. The increased temporal, spatial, and spectral resolution of Sentinel-2 A/B imagery provides new opportunities for improving crop type classification over heterogeneous cultivated land compared with other optical sensors [5,6].

As pointed out by previous studies, due to cloud cover, the optical data discontinuity in key growth stages of crops can still happen [7]. Furthermore, for crop types with similar phenological cycles, only using spectral information is still challenging for reliable discrimination of crop types. As synthetic aperture radar (SAR) can reflect the structure of vegetation, and optical imagery captures the multi-spectral information of crops, it has been indicated that the synergetic use of SAR and optical data can be complementary to each other [8,9].

Space-borne SAR, due to its all-day, all-weather capability, wide coverage, and strong penetrating ability, has been increasingly used in crop classification, to complement with the use of optical imagery. It was found that considerable improvement can be achieved by increasing polarization channels [10,11]. Some studies suggested that using multi-temporal acquisitions can improve the accuracy of crop type mapping, and cross-polarized backscatter outperforms other polarization modes [12,13]. The launches of Sentinel-1 A and B satellites dramatically increased the volume of freely available SAR data, with dual-polarization modes, a 12-day revisit time, and 20 m spatial resolution. Thus, Sentinel-1 data is more desirable for medium to high-resolution crop mapping. However, affected by speckles inherent in SAR imaging systems, crop type mapping using SAR data alone yields quite limited accuracy [14]. The work by Ban et al. suggested that apart from speckles, the single parameter, high incidence angle SAR system used in their study did not provide sufficient differences to differentiate some crop species [15]. Thus, due to the limited viewing angles and orbits of available SAR data in most study cases, the sole use of SAR data may not be sufficient for crop type classification, especially in complex cropping systems.

By the synergic use of microwave Sentinel-1 features and optical Sentinel-2 features, the accuracy of crop discrimination can be potentially improved [15–18]. However, despite the existing studies to combine SAR and optical images for crop classification, few studies (1) explored the performance of individual InSAR products (such as coherence, amplitude dispersion, and master versus slave intensity ratio) in crop type identification. Regarding the de-speckling of SAR intensity, the conventional procedure presented in previous studies uses a regular-shaped window (e.g., boxcar filter, Lee filter, refined Lee filter, etc.) to reduce speckle effects, but in the meantime blurs the image especially over textural areas [19].

The combination of SAR and optical imagery resulted in hundreds of input features (also known as input variables) for the classification model. We adopted a supervised random forest classifier in the crop classification due to its high capacity to deal with a large number of input features. Nevertheless, it has been reported that the classification accuracy can be considerably increased by removing redundant features [20]. Feature selection is a crucial step to improve the performance of a classifier. From an operational perspective, the manual selection of such high-dimensional features is not desirable. Many approaches used separability criterion or hypothetical tests to select features based on assumptions of the sample distribution. In some cases, these assumptions were not satisfied, particularly when using SAR features. Moreover, crop samples can even break the assumption of unimodal distribution. For example, the sowing and harvesting dates are usually farmer customized; crop growth stages are affected by local weather conditions and soil conditions; therefore, they are site-specific. Thus, we prefer not to use separability criterion or hypothetical tests to select features. In the literature, several methods based on machine learning algorithms have been proposed for feature selection [21,22]. Some studies indicated the built-in attribute of random forest, the feature importance score, can be utilized as a ranking criterion to aid feature selection.

The objective of this research is to develop a method to integrate time series Sentinel-1 and Sentinel-2 features for the mapping of typical oasis crop distribution in heterogeneous smallholder farming areas. Firstly, in addition to SAR backscatter intensity, a number of InSAR products were extracted from time series Sentinel-1 data, such as the interferometric coherence, amplitude dispersion, and master versus slave intensity ratio. A statistically homogeneous pixel (SHP) distributed scatterer interferometry (DSI) [23,24] algorithm, originally proposed in the interferometric SAR (InSAR) community to identify distributed scatters (DSs), was adopted for the de-speckling of backscatter intensity and bias mitigation of coherence coefficient, so as to improve the quality of SAR features. To the best of our knowledge, this is the first time the use of amplitude dispersion and bias mitigated coherence is explored in crop type discrimination. Secondly, optical features were extracted from multi-temporal Sentinel-2 images. In particular, red-edge spectral bands and 11 indices were derived and included as input features for the oasis crop classification. Thirdly, a recursive feature increment (RFI) approach, on the basis of random forest feature importance, was proposed to obtain the optimal combination of S1 and S2 features for crop type discrimination. Finally, a random forest classifier was applied to the optimal feature set to produce a crop type distribution map. This study aims to answer the following questions: (1) Does the integration of Sentinel-1 and Sentinel-2 features achieve better performance than using SAR or optical features alone in the oasis crop type mapping? (2) If yes, which SAR feature has the most significant contribution? Are there any InSAR products that are capable of distinguishing oasis crop types? (3) To what extent can the inclusion of red-edge spectral bands and indices improve the accuracy of the oasis crop type identification? Which red-edge band or indices contribute most?
