Refining historical forest cover mapping and change analysis with time series algorithm-based samples transfer

Abstract

Accurate forest cover maps are crucial for ecosystem monitoring and sustainable land use planning. Traditionally, supervised learning methods usually rely on high-quality ground-truth data, which makes it difficult to apply when historical data collection is costly. This study investigates the effectiveness of the LandTrendr algorithm in long-term sample transfer management. First, we assessed the performance of four commonly used machine learning models (random forest, artificial neural network, support vector machine, and XGBoost) in classifying land cover in the reference year. Then, by integrating the best model's classification results with the LandTrendr algorithm, we simplified the process of manually selecting sample points by identifying stable pixel samples, and generated a long-term dataset consistent across time scales, facilitating land cover mapping for historical periods and providing benchmark data for forest cover change analysis. The results show that, among the four models, XGBoost achieved the highest classification accuracy (overall accuracy OA = 0.87, Kappa coefficient = 0.85). By comparing the sample transfer strategy to the traditional model transfer method, we found that the LandTrendr based sample transfer method consistently achieved over 85% accuracy in forest classification throughout the study period, outperforming the direct model transfer approach. Although accuracy decreased over time, the sample transfer method demonstrated stronger consistency. Overall, this study simplifies the process of continuous field data collection by utilizing LandTrendr to transfer stable pixel samples across multiple years, providing methodological support for the dynamic analysis of forest cover changes in Shaanxi Province over time.