A General Framework for Human-Machine Digitization of Geographic Regions from Remotely Sensed Imagery

Digitization of geographic regions, such as bodies of water, from remotely sensed imagery is a highly demanded yet arduous task. Though many automatic approaches for such digitization exist, they are typically geared towards specific regions of interest and sensing platforms, requiring complex workflow specification, parameter tuning, and/or manual training set creation. Moreover, the automatic digitization is performed all at once, making correction overwhelming when the result is not fully accurate. In this study, a general-purpose methodology that aims to greatly reduce the burden of geographic region digitization is proposed. This methodology specifies an interface for a human-machine team that exploits an incremental approach via online and active learning. With no prior training data or workflow definition, an effective pixel-level classifier is built in stride with a human user's adjustments to the machine's automatic vertex placement via novel interactive piecewise-linear contours. Several contour implementations specifically tailored towards geographic regions are presented along with results showing the effectiveness of each implementation. These contours work by spatially constraining the placement of vertices such that a machine implementation may effectively classify and train on pixel-level data for vertex placement after human verification or correction. An implementation of the methodology that uses a K nearest neighbors classifier and a simple instance-based estimation of uncertainty is presented along with several automatic vertex insertaion and placement strategies. Results show that the presented implementation succeeds in helping the user effectively annotate with no prior training data, yielding a vertex placement accuracy of 84% overall. Finally, conclusions and current research directions initiated from our findings are discussed.