Abstract: Accurate identification of orchard spatial distribution is essential for precision management of forest fruit industry. Due to the similar spectral responses of orchards and croplands, high intra-class variability, and the requirement for large training samples in traditional classification methods, accurate discrimination between orchards and croplands remains challenging. This study developed a hierarchical classification framework and incorporated orchard phenological characteristics to achieve fine-scale discrimination between orchards and croplands within cultivated areas in Wensu County, Xinjiang, China. First, a random forest algorithm was applied to classify the first-level land use types by integrating spectral and texture features, thereby identifying the spatial extent of cultivated land. Based on Sentinel-2 multispectral imagery, normalized difference vegetation index (NDVI) time-series images were generated using a semi-monthly median compositing approach. Considering the phenological characteristics of fruit trees, including early leaf bud and late leaf senescence, the orchard phenological accumulation index (OPAI) and the orchard greenness difference index (OGDI) were constructed. By integrating phenological accumulation information and greenness variation characteristics, the orchard mapping index (OMI) was further proposed to enhance the spectral differences between orchards and croplands and improve their separability. The separability index (SI) method was used to determine the optimal temporal window for OPAI, and the corresponding time window for OGDI was subsequently derived. Finally, the separability and thresholds (SEaTH) method was employed to determine the optimal classification threshold of OMI for separating orchards from croplands. The results indicated that the overall accuracy (OA) and Kappa coefficient of the first-level land use classification were 87.63% and 0.85, respectively. The classification accuracy of cropland reached 93.48%, providing a reliable spatial extent for subsequent orchard extraction. The separability analysis of NDVI indicated that the SI values in May and October exceeded 1.40 and 1.00, respectively, demonstrating that these two periods were the optimal phenological windows for orchard identification. The differences in mean and median OMI values between orchards and croplands reached 8.37 and 9.46, respectively, indicating that OMI significantly enlarged the inter-class differences. This indicates that OMI significantly improved the separability between orchards and croplands, achieving better discrimination performance than OPAI or OGDI alone. When OPAI or OGDI was used separately, the OA values exceeded 88%, with a Kappa coefficient of 0.77 and a maximum orchard classification accuracy of 92.68%. In contrast, the classification using OMI achieved an OA greater than 90% and a Kappa coefficient of 0.81, with orchard classification accuracy reaching 94.83%. The approach based on OMI demonstrated superior accuracy and effectively enhanced the separability between orchards and croplands. The optimal OMI threshold determined by the SEaTH method was 4.06, and the resulting orchard distribution pattern was consistent with the actual local planting distribution, confirming the reliability of the classification results. Overall, orchards and croplands exhibited a fragmented spatial distribution pattern in Wensu, with orchards mainly concentrated in the western and central regions, whereas croplands were primarily distributed in the southeastern part of the study area. This study demonstrates that multi-temporal remote sensing data can effectively characterize phenological differences between orchards and croplands, identify key phenological windows for orchard extraction in arid regions, and validate the effectiveness of OMI for orchard mapping. The results provide technical support for regional orchard information extraction using remote sensing.