Assessment of cultivated land productivity with time-series remote sensing and random forest classification

Cultivated land productivity directly influences agricultural sustainability and food security. This study proposes a remote sensing–based approach for assessing cultivated land productivity, integrating traditional assessment principles. A long-term NDVI time series was constructed using Landsat-8 and Sentinel-2 imagery, from which time-series NDVI features and phenological indicators were extracted. Correlation analysis identified variables with correlation coefficients above 0.4, which were then used to train a random forest classification model. The growing season length exhibited the strongest correlation with the integrated fertility index (R = 0.717). The initial classification achieved an accuracy of 75.30% and a Kappa coefficient of 0.69, with macro-averaged precision, recall, and F1-score around 75%. After merging similar productivity levels, overall accuracy improved to 92.22%, the Kappa coefficient increased to 0.88, and all macro-averaged metrics exceeded 91%. The spatial consistency between the remote sensing–based results and traditional assessments confirmed the method’s reliability. This research provides an efficient and accurate framework for mapping cultivated land productivity at the county scale, supporting scientific land management and agricultural decision-making.