National scale tree canopy cover modelling using google earth engine and stacking ensemble: A case study of the Greek forests.

Forest ecosystems are major providers of multiple ecosystem services, yet they are increasingly impacted by various natural and human-induced disturbances, as well as by climate change, particularly evident in the territory of the Mediterranean basin. As a result, forests are under significant pressure, leading to their severe degradation and loss. Therefore, the systematic monitoring of forests and the retrieval of their attributes are crucial for effective forest management and conservation planning. In this study we produced a national scale tree canopy cover map for Greece, at 20 m resolution, incorporating visually interpreted sample plots, Sentinel-2 satellite data, elevation data, canopy height, and an ecosystem type map. The tree canopy cover model was developed using a stacked ensemble of four machine learning algorithms; Gradient Boosting Machines, Distributed Random Forest, Generalized Linear Model, and Deep Learning (Neural Networks). For the model's assessment, maps depicting per pixel prediction uncertainties were generated. Tree canopy cover ensemble models were developed for six forest types of Greece. According to the results, the tree canopy cover model, achieved an R2 value of 0.72 and RMSE of 18.846%. For the individual tree canopy models developed per forest ecosystem type, performance varied, with R2 value ranging from 0.165 to 0.535 and RMSE from 19.811% to 25.767%. The study highlights the potential of integrating advanced machine learning techniques with earth observation data to enhance forest mapping and provide a standartised approach for forest monitoring and accounting, simultaneously supporting sustainable forest management practices, toward the implementation of national and EU nature conservation, restoration and adaptation policies.