Advanced Airborne Hyperspectral Remote Sensing to Support Forest Management - HYPERFOREST

Coordinating Institution: CRP Gabriel Lippmann
Contracting Partner(s): University of Louvain (B) , Flemish Institute of Technological Research (B) , University of Ghent (B) , University of Zürich (CH) , Research Institute for Nature and Forest - INBO (B)
From: 01/01/2010
To: 31/12/2013
Budget: 176,000.00€
Contact(s): Udelhoven Thomas

Summary

Forest management should encompass the many functions related to forest recourses. It requires detailed data to execute current operations, to build-up records of past activities and to predict the long-term impacts of management decisions. In support of this, the HYPERFOREST proposal aims at providing foresters with detailed spatial explicit information on forest vitality, species composition, canopy closure, etc, based on airborne hyperspectral remote sensing data.

Due to the complex nature of hyperspectral remote sensing data sets, a complete imagery preprocessing chain must be set up to perform standard corrections for radiometric, geometric and atmospheric effects which might corrupt the data. Moreover, bidirectional effects caused by the heterogeneous character of terrestrial targets are influencing the captured (airborne) hyperspectral signal. Forests are such heterogeneous surfaces and might have pronounced vegetation structure which will also affect the accuracy of hyperspectral derived thematic products, useful in forest management practices.

What is more, forest management plans cannot be effective without proper knowledge on forest vegetation structure. Hence, this project targets (i) the development of an advanced airborne hyperspectral imagery pre-processing chain (e.g. APEX) that considers vegetation structure effects and hence bidirectional effects on the captured signal, (ii) the delivery of a robust methodology to extract forest thematic products from this pre-processed imagery, and (iii) intensive interactions with end-users by considering their feedback facilitating the supply of tuned and more end-user oriented forest thematic products. First of all, this requires the determination of forest structure parameters (for instance crown density, vertical LAI distribution, etc) at the forest test sites (three plot locations in Flanders: Wijnendalebos, Aelmoeseneiebos, Kersselaerspleyn as indicated on the map of form 7) derived from full dendrometric inventories, fine spatial scale terrestrial and coarser scale airborne LiDAR measuring campaigns.

In order to identify the most contributing structure parameters to the hyperspectral signal, radiative transfer models will be used. Reference forest canopy spectral data will be collected using field spectroradiometers in the Aelmoeseneiebos (where a measuring tower is available). Canopy leaf picking and leaf biochemical analysis (chlorophyll, dry matter and water content) will be conducted since they are crucial inputs in these radiative transfer models. The analysis of the effects of vegetation structure on hyperspectral signatures will be accomplished using a bottom-up (frog’s eye view with the terrestrial LiDAR) and top-down (bird’s eye view with the airborne LiDAR and the terrestrial LiDAR mounted on the measuring tower) approach.

The bottom-up approach initiates with implementing gradually coarser vegetation structure data (from high to less detail) - the structure which is most affecting the hyperspectral signal - in the radiative transfer models. From this analysis, the minimum required level of canopy structure info that can also be obtained from airborne LIDAR data is assessed at spatially explicit scale. Once the most contributing structure parameters are available from airborne LiDAR data, combined with its quantified effect on hyperspectral signals, a procedure can be developed to build an advanced hyperspectral imagery pre-processing chain (for APEX data) that considers the impact of vegetation structure and its bidirectional effects on the captured signal.

This procedure will be based on comparison of the original APEX signals with simulated ones from radiative transfer models with airborne derived vegetation structure parameters as inputs. Finally, a methodology based on deep belief neural networks will be developed to produce forest parameters from the remote sensing thematic data derived from the traditional and advanced hyperspectral imagery pre-processing chains. Basically, this project aims at identifying forest canopies components that contribute the most to the captured reflectance values of airborne sensors. Or stated otherwise, adding a third and vertical dimension to the horizontal and thus two dimensional character of hyperspectral imagery. With this knowledge, monitoring and detecting changes in the vegetation state can be differentiated in contributions of physiologic changes like growth or senescence and contributions due to vegetation structure affecting the captured airborne hyperspectral signal.