Abstract: Driven by the expansion of agricultural production, remote sensing technology has gradually replaced traditional manual survey methods in crop monitoring. Notably, the time-series remote sensing data provide new opportunities for high-precision crop mapping and spatiotemporal evolution analysis. However, existing remote sensing methods often face input feature redundancy and the curse of dimensionality, significantly influencing the accuracy and efficiency of crop mapping and spatiotemporal evolution analysis. To address this problem, by combining random forest and hierarchical clustering algorithms, this study proposed a new crop mapping method based on feature screening. By evaluating the importance of spectral and temporal features, this study eliminated redundant features while retaining the most discriminative features. Then, combined with machine learning technology, it significantly improved the efficiency of crop mapping and spatiotemporal evolution analysis. Using this method, this study generated a planting distribution map of major grain crops in Heilongjiang Province from 2019 to 2023 based on an optimized feature set and a random forest classifier, with a crop classification accuracy of 89.39% and a Kappa coefficient of 0.85. Compared with the method using full time-series features, the proposed method reduced the classification time by 85% at the expense of an accuracy loss of only 0.11 percentage points, underscoring its significant strengths in crop mapping. On this basis, the spatiotemporal evolution analysis indicates a year-on-year decline in rice planting area, an upward trend in corn planting area, and a stable soybean planting area. In summary, this study provides precise support for agricultural remote sensing monitoring and a reliable tool for the spatiotemporal evolution analysis of major grain crops, holding important application potential in precision agriculture management and food security monitoring.