The spatiotemporal pattern of non-agriculturalization can be accurately and rapidly captured for the rational formulation of land use and cultivated land protection. Remote sensing imagery has been more precise and efficient for the large-scale dynamic monitoring of cultivated land, with the advancements in feature extraction and classification. There are also different and feature variables during dynamic monitoring of cultivated land. In this study, Sentinel-2 satellite imagery was employed to monitor the non-agriculturalization of cultivated land. The experimental and validation zones were selected from the northern part of Xinjiang, a vital agricultural production base in China. The reclaimed areas were also in the Eighth Division of the Xinjiang Production and Construction Corps in Shihezi City. The data source was comprised of Sentinel-2 remote sensing images, covering the research area as of July 17, 2022, and August 16, 2020. The images were improved to facilitate the automatic extraction of information, according to the non-agricultural land use on cultivated land. Initially, the spectral indices and gray-level co-occurrence matrix were calculated to obtain 31 features, including spectral, spectral index, and texture features. Subsequently, Pearson correlation coefficients were computed for the eight band features, seven non-red edge spectral indices, and eight red edge spectral indices, leading to the removal of redundant features with weak classification and high correlation. Principal component analysis (PCA) was applied to extract the first three components of eight texture features. Feature importance scores were ranked using the random forest (RF) with the Gini index, resulting in the selection of optimal features: red, green, blue, narrow NIR (B8A), normalized difference vegetation index (NDVI), modified soil adjusted vegetation index (MSAVI), red-edge chlorophyll index (Clre), red-edge vegetation index (REDVI), triangular vegetation index (TVI), PCA1, PCA2 and PCA3 of texture features. 12 features and the original were selected to construct five classification schemes, including four optimal combinations (Scheme 1-3 and Scheme 5) and the original (Scheme 4). Seven classification algorithms were then utilized to evaluate: RF, artificial neural network (ANN), minimum distance (MID), maximum likelihood (MAL), mahalanobis distance (MAD), support vector machine (SVM), and Softmax. The results included: 1) The importance ranking of feature variables was as follows: band features, red edge spectral index features, non-red edge spectral index features, texture features. 2) Accuracy improvement depended mainly on the effective selection of multiple feature variables and algorithms. The two-stage model was constructed using Softmax. The highest accuracy was achieved with the optimal combination of feature variables, namely band + spectral index + texture features. The dimensionality of the feature variable was reduced to 12. The overall accuracy, average user's accuracy, average producer's accuracy, and Kappa coefficient were 94.92%, 95.16%, 93.15%, and 0.88, respectively. 3) A comparison of 2020 and 2022 data revealed that there was an area of 146.153 km² transitioning from cultivated land to non-agricultural use, while an area of 123.074 km² was transitioned from nonagricultural to cultivated land. Overall, the cultivated land decreased by 23.079 km² . In conclusion, the non-agriculturalization of cultivated land can provide technical support and references for the extraction of land cover information, particularly for the resource protection and sustainable utilization of cultivated land. [ABSTRACT FROM AUTHOR]