Remote Sensing Of Large Wood In High-Resolution Satellite Imagery: Design Of An Automated Classification Work-Flow For Multiple Wood Deposit Types

Wood researchers increasingly rely on remote-sensing products to augment field information about wood deposits in river corridors. The availability of very high-resolution (<1 m) satellite imagery makes capturing wood over greater spatial extents possible, but previous studies have found difficulty in automatically extracting wood deposits due to the challenge in distinguishing wood from spectrally similar corridor features such as sand. We also lack knowledge on the spectral properties of different wood deposit types in multiple depositional environments. In this work, we explore image classification work-flows for four wood deposit types in three North American environments: in-channel jams deposited in the Tatshenshini River in Alaska, USA; a wood raft on the Slave River in Northwest Territories, Canada; and wood deposited along a lakeshore and coastal embayment in the Mackenzie River Delta in Northwest Territories, Canada. We compare classification results of object-based and pixel-based image analysis with supervised [support vector machine (SVM)] and unsupervised (ISO clustering) classifiers. We evaluate several accuracy assessment parameters and achieve overall classification accuracies of 65-99%, showing automated image classification is a possible approach for analysing wood across larger areas. We also find that wood sensitivity in the classification ranged from 0 to 95%, indicating that some techniques are better suited to wood capture than others. We find that supervised classification produced more accurate wood maps, though there is large variation in classification outcomes across environments related to spatial arrangement of wood in the landscape. We discuss the influence of depositional environment on classification and provide recommendations for designing a wood classification work-flow.