Machine learning in cropland dynamics: Evidence from China

The scientific identification of cropland dynamics patterns is essential for developing effective protection policies. Utilizing satellite remote sensing data, this study applies DTW (Dynamic Time Warping) K-means algorithms to classify cropland dynamics in China, providing a solid foundation for targeted policy design. The findings reveal four distinct patterns: growth, fluctuation, late-stage shrinkage, and early-stage shrinkage. An ensemble learning analysis further identifies key predictors of cropland dynamics, including water irrigation capacity, agricultural mechanization, industrialization, and urbanization. Notably, the primary predictors vary across different cropland dynamics patterns. The growth pattern is characterized by robust increases in agricultural mechanization and irrigation capacity, while the fluctuation pattern exhibits slower progress in both mechanization and irrigation. Meanwhile, the early-stage shrinkage pattern is distinguished by rapid non-farm economic expansion and higher levels of urbanization. Based on these insights, we propose a differentiated approach to cropland protection policies, ensuring strategies are tailored to specific regional dynamics. Overall, this study offers valuable guidance for future cropland conservation and management efforts. • We use DTW algorithms to classify cropland dynamics in China and find four distinct patterns for cropland use. • The growth pattern is characterized by rapid increases in mechanization and irrigation. • The fluctuation pattern exhibits slower progress in mechanization and irrigation. • The early-stage shrinkage pattern is distinguished by non-farm economic expansion and high urbanization. • Policy making should take into account pattern differences.