Evaluating the potential of multi-view data extraction from small Unmanned Aerial Systems (UASs) for object-based classification for Wetland land covers

Abstract

Unmanned Aerial Systems (UASs) have the potential to provide multi-view data, but the approaches used to extract the multi-view data from UAS and investigation of their use in image classification are currently unavailable in publications to our best knowledge. This study presents a method that combines collinearity equations and a two-phase optimization procedure to automatically project a point from real world coordinate system of an orthoimage to UAS image coordinate system (row and column numbers) to be used in multi-view data extraction. The results show average errors for the computed UAS column and row numbers were 1.6 and 1.8 pixels respectively evaluated with leave-one-out method. Based on this algorithm, it’s also for the first time that object-based multi-view data were extracted and presented, and the potential of using the multi-view data to aid Geographic Object-Based Image Analysis(GEOBIA) through bidirectional reflectance distribution function (BRDF) modelling was evaluated with two representatives of BRDFs, the Rahman-Pinty-Verstraete(RPV) and Ross-Thick-LiSparse (RTLS). Our results indicate the RPV model tends to overestimate the bidirectional reflectance for land cover types with high reflectance, while perform well for those with relatively low reflectance in our study area. To test the impact of using multi-view data on image classification, we extracted parameters from BRDF models and used these parameters as object features for object-based classification. The 10-fold cross validation results show that the 3-parameter RTLS significantly improved overall accuracy compared to the classifications relying only on the orthoimage features, while other BRDF models did not show significant improvements, raising the needs to develop new methods to better utilize the multi-view information in GEOIBA in the future.