A Multi-Sensor Approach for VHR Vegetation Monitoring

Abstract

The Hawaiian Islands are a showcase of biological diversity. With a myriad of vegetation communities, the tropical forests of Hawaii support a rich assemblage of endemic species, some of which are critically endangered. However, much of the Hawaiian forests are degraded and are subject to disturbance by invasive plants. Monitoring the response of Hawaiian forests to management efforts and tracking how vegetation changes over time is a key component of conservation and restoration efforts. Traditional “on-the-ground” vegetation monitoring techniques are time consuming and costly, and can vary in accuracy and consistency. Recent advances in remote sensing technology hold potential for providing an accurate and timely assessment of vegetation at a set point in time. Until recently, the available satellite sensors lacked the spatial resolution required to differentiate individual tree crowns, and thus, classification was limited to the stand or community level. Several new very high resolution (VHR) platforms have emerged in the field of remote sensing that can differentiate individual tree crowns and, thus, have the potential to change the paradigm of vegetation monitoring and its efficacy. VHR, sub-meter imaging platforms are now readily available for public use with commercial VHR satellite and aircraft imaging, unmanned aerial system (UAS) digital imaging, and the Gigapan system. The primary objective of this thesis was to determine the utility of new high spatial resolution remote sensing technologies for vegetation mapping and monitoring in Hawaiian forests. The strengths of the three platforms were evaluated and then combined, to produce an effective synthesis to implement remote sensing-based mapping to the species level and an OBIA procedural workflow was outlined. WV-3 imagery was classified with object based image analysis in eCognition into 7 vegetation classes and validated with UAS and Gigapan imagery. The dense vegetation of the Hawaiian mixed-mesic forest presents a challenging task to separate vegetation classes to the species level. Validation results yielded an overall user’s accuracy of 65% with Sparse Veg representing the highest user’s accuracy of 94% and Strawberry guava representing the lowest user’s accuracy of 38%. Kukui=75%, Christmas berry=73%, Koa=50% and Native Complex=42%. Grouping native and non-native vegetation classes yielded an overall accuracy of 72% with non-native=94% and native=69%. The high accuracy of mapping sparse veg shows great potential for providing information towards fuel mapping via this method. Further work is needed to accurately separate native vs non-native vegetation to the species level. A stronger computer processer is needed to add additional geometric and textural features into the iterative classification process. The UAS VHR platform shows the greatest potential for integration of remotely sensed imagery into an operational vegetation monitoring method. UAS allow for low cost, repeatable, high resolution data collection without risk to field personnel. A recommended method could employ a UAS to fly transects in a target area with visual or deep/machine learning analysis of random plots along the transects. Further advancements in multispectral sensors and longer lasting batteries will serve to allow for greater utility in monitoring, and management applications. Vertical takeoff and landing UAS may be of great use in areas without suitable landing area for typical fixed wing UAS. The costs associated with the implementation of remote sensing based monitoring protocols were determined as compared to traditional ground based monitoring methods. Ultimately, if new imagery was obtained under contract, remote sensing based monitoring serves to be more expensive than traditional ground based methods. However, an operational comparison which factors in either prior acquisition of imagery or capacity to gather data without going out to contract, shows a lower cost associated with remote sensing based monitoring.