Land cover classification for Siberia leveraging diverse global land cover datasets

Abstract

Abstract Understanding the land cover is crucial to comprehending the functioning of the Earth’s system. The land cover of Siberia is characterized by uncertainty because it is wide-ranging and comprises various classification types. However, comparisons among land cover products reveal substantial discrepancies and uncertainties. Therefore, a reliable land cover product for Siberia is necessary. In this study, we generated new land cover data for Siberia using random forest (RF) classifiers with global land cover datasets. To assess their accuracy and characteristics, we individually validated global land cover products in Siberia using multi-source sample datasets. We trained the RF classifiers with multiple land cover products to produce a more precise land cover product for Siberia. The validations showed that: (a) the generated new land cover data achieved the highest overall accuracy (85.04%) and kappa coefficient (82.62%); (b) the classifications of mixed forest (user accuracy: 97.85%) and grasses (user accuracy: 94.85%) demonstrated improvements, showing higher performance compared to most other types; and (c) by comparing the distribution of land cover across climate zones, we discovered that temperature is a critical factor throughout Siberia. However, in warm summer climates, precipitation plays a critical role in vegetation distribution. The more accurate and detailed land cover created in this study enhances the reliability of analyses in Siberia and fosters a deeper understanding of the impact of the carbon cycle.

Keywords

• Land cover
• Global land cover
• Siberia
• Random forest classifiers
• Cross-validation