Abstract: The resilience of cropland refers to the ability of cropland to recovery after external disturbance, which is the basis for the determination of land reclamation and ecological restoration strategies. There is still a lack of effective methods to assess the resilience of cropland in coal mining subsidence area. In this paper, machine learning method is introduced to convert the resilience assessment into the measurement of the probability of the possibility of cropland restoration, and empirical research is carried out on the resilience level of cropland in coal mining subsidence area in north mining area of Xuzhou. The results show that random forest algorithm can be used as a new method of resilience assessment, and the accuracy of the model optimization is as high as 88.28%, which can make full use of the historical experience of cropland restoration, avoid subjectivity in the assessment process, and accurately reflect the ability of cropland restoration in coal mining subsidence area. The probability of resilience of cropland lies between 0.037 and 0.995, with high resilience in the peripheral zone of the subsidence area and low resilience in the central zone. The machine learning shows that irrigation guarantee rate is the core factor affecting the resilience of cropland in coal mining subsidence area, with an importance ratio of 20.88%; followed by the depth of waterlogging, with a ratio of 17.15%; the factors such as the content of soil organic matter, the fragmentation of cropland, and the accessibility to roads also have a great impact, with importance ratios of 15.18%, 12.58%, and 12.38%, respectively. The study shows that the training sample data and machine learning method can effectively assess the level of resilience of coal mining subsided cropland, which provides a scientific reference for the decision-making of land reclamation and ecological restoration in mining areas.