کد مقاله | کد نشریه | سال انتشار | مقاله انگلیسی | نسخه تمام متن |
---|---|---|---|---|
6348458 | 1621804 | 2016 | 10 صفحه PDF | دانلود رایگان |
- Remote sensing of canopy cover (CC) is essential in forest monitoring and ecology.
- Large area CC prediction in boreal forests using Landsat was shown feasible.
- High quality ground CC reference (NÂ =Â 250) across diverse forest structure and sites.
- Simple beta regression model with red reflectance optimal to predict CC.
- First validation of Landsat Tree Cover Continuous global product in boreal forests.
Large area prediction of continuous field of tree cover i.e., canopy cover (CC) using Earth observation data is of high interest in practical forestry, ecology, and climate change mitigation activities. We report the accuracy of using Landsat images for CC prediction in boreal forests validated with field reference plots (NÂ =Â 250) covering large variation in latitude, forest structure, species composition, and site type. We tested two statistical models suitable for estimating CC: the beta regression (BetaReg) and random forest (RanFor). Landsat-based predictors utilized include individual bands, spectral vegetation indices (SVI), and Tasseled cap (Tass) features. Additionally, we tested an alternative model based on spectral mixture analysis (SMA). Finally, we carried out a first validation in boreal forests of the recently published Landsat Tree Cover Continuous (TCC) global product.Results showed simple BetaReg with red band reflectance provided the highest prediction accuracy (leave-site-out RMSECV 13.7%; R2CV 0.59; biasCV 0.5%). Spectral transformations into SVI and Tass did not improve accuracy. Including additional predictors did not significantly improve accuracy either. Nonlinear model RanFor did not outperform BetaReg. The alternative SMA model did not outperform the empirical models. However, empirical models cannot resolve the underestimation of high cover and overestimation of low cover. SMA prediction errors appeared less dependent on forest structure, while there seemed to be a potential for improvement by accounting for endmember variability of different tree species. Finally, using temporally concurrent observations, we showed the reasonably good accuracy of Landsat TCC product in boreal forests (RMSE 13.0%; R2 0.53; bias â2.1%), however with a tendency to underestimate high cover.
Journal: International Journal of Applied Earth Observation and Geoinformation - Volume 53, December 2016, Pages 118-127