A novel edge-based accuracy assessment framework for target detection from remotely sensed images

Accuracy assessment is critical for evaluating map quality and improving algorithms for target detection in remotely sensed images. Traditional methods for assessing a target-detection task were often performed by estimating the relative pixels of correct and incorrect identifications (e.g., F 1 score). However, this approach only considers the pixel-by-pixel agreement and ignores important aspects of positional errors and target object quality. To provide comprehensive diagnostic information, we propose a novel edge-based accuracy assessment framework. This framework firstly applied a newly proposed assessment tool, Dynamic Epsilon-band Accuracy Function (DEAF), which is a linear fitting function from a series of edge accuracies associated with expanded buffers on the object edges, and then extracted the level-off and the middle point from the DEAF curve respectively to derive edge-based accuracy metrics and to assess thematic and positional error components. We presented an experiment on synthetic maps to reveal that the new edge-based accuracy metrics better reflects under- and over-segmentation errors compared to traditional measurements (e.g., F 1 score, over-/under-segmentation indices). Moreover, we applied the new framework for two practical target-detection tasks, agricultural fields and building rooftops. The results demonstrate the two innovative aspects of the edge-based framework, which are that it provides: 1) a more rigorous assessment of object quality, and 2) error attribution for both thematic and positional errors. The proposed framework therefore provides a new capability for assessing the accuracy of target detection maps, which is of vital importance for evaluating and improving state-of-the-art remote sensing products.