Abstract: To explore the practical role of enhanced vegetation index (EVI) time series data in improving the accuracy of forest type recognition could promote the deep application of optical remote sensing data in forest resources investigation and monitoring. With Cuigang Forest Farm of Xinlin Forestry Bureau in Daxing’anling as the object, we constructed six classification schemes, using random forest algorithm with spectral feature, texture feature and EVI time series feature. The data sources were 20-view Landsat 8 OLI time series data from 2014 to 2018, 56 fixed plots data from 2017-2019, and the 2016 Class II survey data. Our aims were to realize the classification of forest types in Cuigang Forest Farm and to evaluate the accuracy of different classification schemes. The results showed the EVI values of Larix gmelinii forest, Betula platyphylla forest, coniferous-broadleaved mixed forest, coniferous mixed forest and broadleaved mixed forest were significantly different in non-growing seasons (36-111 days and 287-367 days), with the EVI value of mixed conifer forest being significantly higher, and that of mixed broadleaf forest being always lower than the other four forest types. In the early growing season (111-143 days), the EVI value of B. platyphylla forest were higher than L. gmelinii forest, which could effectively distinguish the two forests. Among the six classification schemes, spectral feature, texture feature, and EVI time series feature had the highest classification accuracy, with a Kappa of 0.82 and a classification accuracy of 86.1%. The comparison results showed that the overall accuracy of adding vegetation index time series feature was improved by 14.3% compared with that of spectral feature. The random forest algorithm with combined spectral, texture and EVI time series features could effectively classify forest stand types in Cuigang Forest Farm, with good recognition accuracy and confidence.

Key words: time series data, Daxing’anling, forest type, random forest algorithm