Application of big data analytics in remote sensing supporting sustainable forest management

Sustainable forest management requires detailed forest information for planning accurate treatments. The information is expected to be accurate enough and preferably obtained at a low cost and with periodic updates. Such spatial scale information is nowadays provided by remote sensing data. On the one hand, the development and use of aerial laser scanning for estimating forest variables has been a game-changer in recent decades for forest management. On the other hand, satellite remote sensing technologies, generated a constant flow of data from different platforms, in different formats and with different purposes. Combined with this ongoing remote sensing data stream, the development of computer technology has provided forest management with many new tools for data capture, data representation, data visualization, and management planning applications. Today, new computing power makes it possible to tackle the complex problem of managing and processing big data from remote sensing with new strategies that have revolutionized the way of understanding the use of these data sources. This thesis is aimed at assessing big data analytics for practical cases of forest monitoring, especially in the Italian context, where large-scale aggregated forest remote sensing data have always been a structural lack. Four main studies were covered in the thesis. Study I involved the review and aggregation of remotely sensed forestry data at the national scale. The available Italian airborne laser scanning data were aggregated to develop a consistent mosaic of canopy heigh model, while different local forest maps were used to develop for the first time a high-resolution forest mask of Italy which was validated against the official statistics of the Italian National Forest Inventory. An online geographic forest information system was implemented to store and facilitate the access and analysis of both spatial datasets. The two information layers were explored in operational cases, through the integration of remote sensing and inventory data in studies II and III. In the former, the forest mask produced mosaicking the Italian regional local forest maps was compared with four other forest masks available for the entire area of Italy to examine their effects on the estimation of growing stock volume and to clarify which product is best suited for this purpose. Non-forest pixels in each forest mask were removed from a national wall-to-wall growing stock volume map constructed using inventory and remote sensing data. The estimated Growing stock volume from each mask was compared with the official national forest inventory estimates. In the III study, airborne laser scanning coverage and the forest mask were used in combination with Landsat spectral data for large-scale volume estimation. Estimates were performed considering different proportions between airborne laser scanning and Landsat coverage. The integration between satellite spectral data and airborne laser scanning information is particularly critical in countries like Italy, where wall-to-wall airborne laser scanning coverage is still lacking. In the last study (IV), Sentinel-2 multitemporal data were used to identify poplar plantations, which are the primary source of Italian industrial timber. The study area was the dynamic agricultural area of Pianura Padana where most of the Italian poplar plantations are concentrated. The capabilities of the Sentinel-2 data were integrated with a deep learning approach that provided better results compared to traditional logistic regression. The map we produced can allow the poplar plantation monitoring, which requires frequent updating, not feasible with traditional forest inventories. In so doing, these studies, aimed at enhancing knowledge about missing information layers at the national scale, attempting to close the gaps underlined by previous studies.