Land surface phenology (LSP) characterizes the seasonal dynamics of vegetation communities that compose individual satellite pixels and its interannual and spatial variations have been widely associated with climate. However, increasing evidence shows an effect of land cover composition within a pixel on LSP, but it remains unclear the extent of impacts relative to other drivers. To fill this gap, this study quantitatively assessed the contributions of land cover composition, climate, and topography on the spatial and interannual variation in LSP throughout the 2002 Ponil Complex Fire in New Mexico, USA, using a machine learning approach of Boosted Regression Trees (BRT). As the fire mainly converted ponderosa pine and Douglas-fir (evergreen tree) to soil ground and Gambel Oak (deciduous shrub), we computed both the proportion of tree cover to all vegetation cover (PTV) and vegetation fractional cover (VFC) as the metrics of land cover composition from high-resolution images in 2018 and from MODIS growing season greenness from 2001-2018. Start (SOS) and end (EOS) of growing season were derived from 500-m MODIS data from 2001-2018 and 30-m Harmonized Landsat Sentinel-2 data in 2018. BRT models showed that PTV was the most important predictor of spatial variations in SOS and EOS in 2018, despite the different contributions (20.3% - 42.9%) at 30-m and 500-m spatial scales. Although the growing degree days (28.6%) and the first freeze date (19.6%) were the most important predictors of interannual variations in SOS and EOS from 2001-2018, respectively, VFC also presented an important contribution for SOS (8.4%) and EOS (12.2%). This study demonstrates the utility of machine learning in modeling phenology and highlights the essential role of land cover composition in understanding the spatial and interannual variations of LSP that have been widely associated with topography and climate.

地表物候学（LSP）表征组成单个卫星像素的植被群落的季节动态，其年际和空间变化与气候广泛相关。然而，越来越多的证据表明，像素内的土地覆盖成分对LSP有影响，但尚不清楚相对于其他驱动因素的影响程度。为了填补这一空白，本研究使用增强回归树（Boosted Regression Trees）的机器学习方法，定量评估了2002年美国新墨西哥州Ponil Complex Fire期间LSP的空间和年际变化对土地覆盖成分，气候和地形的贡献（ BRT）。由于大火主要将美国黄松和花旗松（常绿树）转化为土壤和甘贝橡树（落叶灌木），我们从2018年的高分辨率图像和2001-2018年的MODIS生长季绿色度中，计算出树木覆盖率占所有植被覆盖率（PTV）和植被分数覆盖率（VFC）的指标，作为土地覆盖成分的指标。生长季节的开始（SOS）和结束（EOS）来自2001-2018年的500 m MODIS数据和2018年的30 m Harmonized Landsat Sentinel-2数据。BRT模型显示PTV是空间变化的最重要预测因子尽管在30米和500米空间尺度上做出了不同的贡献（20.3％-42.9％），但2018年SOS和EOS的贡献仍然很大。尽管生长度天数（28.6％）和首次冻结日期（19.6％）分别是2001-2018年SOS和EOS年际变化的最重要预测指标，但VFC还为SOS（8.4％）和SOS做出了重要贡献EOS（12.2％）。