Abstract: The Hetao Irrigation District served as a critical hub for grain and oilseed production in China. Accurately and rapidly identifying its cropping structure and elucidating its spatiotemporal evolution were of significant value for understanding agricultural development trends in Northern China. However, due to the high fragmentation arising from mixed cultivation and the overlapping phenological periods of various crops in this region, precisely distinguishing crop types within diverse fragmented patches via remote sensing imagery remains a formidable challenge. To address this issue, this study leveraged the Google Earth Engine (GEE) cloud platform to obtain 9,673 sample points encompassing four crop categories—wheat, maize, sunflower, and others—through visual interpretation of imagery spanning 2000 to 2023; subsequently, the number of decision trees in the random forest algorithm was optimized using a grid search method within a range of 0 to 300 at a step size of 10. By utilizing variables such as crop spectral characteristics, vegetation indices, harmonic coefficients, and texture features, the Gini coefficient was introduced to analyze feature importance, and hierarchical clustering was executed with a correlation threshold of 0.9 to derive the optimal feature set for the model. Concurrently, tailored to the structural characteristics of farmland cropping in the Hetao Irrigation District, a Pixel-based Cropland Fragmentation Index (PCFI) was proposed to evaluate the spatiotemporal evolution of the cropping structure, integrating the proportion of neighboring pixels distinct from the central pixel within a 3 × 3 window (heterogeneity, H) and the distance from a given pixel to the nearest pixel of the same crop type (proximity, N). The results demonstrate that following feature optimization, the overall accuracy of the random forest model generally improved by approximately 1% (with only a minor decline in simulation accuracy in 2021), and specifically achieved overall accuracies of 97.04% and 95.76% in 2022 and 2023, respectively; meanwhile, the model outputs exhibited high consistency with actual survey data, yielding an overall accuracy of 89.09% and a Kappa coefficient of 0.81. Regarding crop identification, vegetation indices contributed most significantly to the classification, followed by crop spectral characteristics, whereas texture features exhibited relatively low importance. From 2000 to 2023, the cultivation area of spring wheat in the Hetao Irrigation District continuously decreased, with particularly pronounced declines in sub-irrigation areas such as Wulanbuhe and Yongji; conversely, the area dedicated to spring maize steadily expanded, showcasing a stable growth trend across all sub-irrigation regions especially after 2010; the sunflower planting area exhibited a marked expansion after 2005 but tended to stabilize after 2020. Furthermore, the PCFI displayed an overall downward trend during the study period, wherein the planting fragmentation of spring maize and sunflowers exhibited a synergistic decline and converged, although the fragmentation level of spring wheat specifically exhibited an upward trajectory; meanwhile, a high concentration of identical crop cultivation was observed in the Wulanbuhe sub-irrigation area, whereas fragmented crop plots remained prominent in peri-urban zones within the Jiefangzha sub-irrigation area. Ultimately, the findings of this study provide an optimized technical approach and methodology for the efficient and precise extraction of spatial distribution patterns and temporal dynamics of crops in regions characterized by complex planting structures.