Forest dynamics in Mediterranean forest using landsat imagery and LiDAR

Abstract

Twelve fungal strains including Lecanicillium muscarium (Petch.) Zare and Gams, Isaria farinosa (Holmsk.) Mediterranean forests are complex ecosystems and most aspects of their functioning are unknown. Remote sensing applications are used to monitoring the effect of forest management. The study area counted with an old management plan from 1981 which was reviewed in 2012. Six different databases built with Landsat bands, vegetation, tasseled cap transformations, LiDAR-based variables and topographic variables were used to assess the performance of Random Forest algorithm. Best of the models were used to produce forest-no forest and multiclass land cover maps for the years 1981, 1990, 2000, 2009, 2011 and 2014. Processes and patterns were analyzed using landscape metrics. Also RF regression models were assessed for prediction of canopy height model, canopy cover and maximum height for the same years. All classification models presented a very good performance (Accuracies > 85% with kappa > 0.80). For multiclass classification, the best performance was achieved by FUSION with 95.37% of accuracy and kappa 0.94. For binary classification LT, LVT and FUSION achieved more than 98.5% in accuracy and kappa index higher than 0.96. Results showed that the silvicultural activities focused on pine tree species for biomass production modify the landscape by recovering Holm oak species. The landscape in the study area became fragmented over the study period, because of the increase in the Number of Patches and the decrease in Mean Patch Area.Estimation of canopy height and canopy cover with the use of RF did not offer such a robust variable for explaining forest dynamics since the accuracies ranged about 0.74 with RMSE of 0.563m, 6.99% and 2.3m, for CHM, CC and HMAX, respectively.

Twelve fungal strains including Lecanicillium muscarium (Petch.) Zare and Gams, Isaria farinosa (Holmsk.) Mediterranean forests are complex ecosystems and most aspects of their functioning are unknown. Remote sensing applications are used to monitoring the effect of forest management. The study area counted with an old management plan from 1981 which was reviewed in 2012. Six different databases built with Landsat bands, vegetation, tasseled cap transformations, LiDAR-based variables and topographic variables were used to assess the performance of Random Forest algorithm. Best of the models were used to produce forest-no forest and multiclass land cover maps for the years 1981, 1990, 2000, 2009, 2011 and 2014. Processes and patterns were analyzed using landscape metrics. Also RF regression models were assessed for prediction of canopy height model, canopy cover and maximum height for the same years. All classification models presented a very good performance (Accuracies > 85% with kappa > 0.80). For multiclass classification, the best performance was achieved by FUSION with 95.37% of accuracy and kappa 0.94. For binary classification LT, LVT and FUSION achieved more than 98.5% in accuracy and kappa index higher than 0.96. Results showed that the silvicultural activities focused on pine tree species for biomass production modify the landscape by recovering Holm oak species. The landscape in the study area became fragmented over the study period, because of the increase in the Number of Patches and the decrease in Mean Patch Area.Estimation of canopy height and canopy cover with the use of RF did not offer such a robust variable for explaining forest dynamics since the accuracies ranged about 0.74 with RMSE of 0.563m, 6.99% and 2.3m, for CHM, CC and HMAX, respectively.