BTS: a binary tree sampling strategy for object identification based on deep learning

Object-based convolutional neural networks (OCNNs) have achieved great performance in the field of land-cover and land-use classification. Studies suggested that generation object positions (OCPs) largely determines OCNNs. Optimized distribution OCPs facilitates identification segmented objects with irregular shapes. In this study, we propose a morphology-based binary tree sampling (BTS) method provides reasonable, effective, robust strategy to generate evenly distributed OCPs. The proposed BTS algorithm consists three major steps: 1) calculating required number for each object, 2) dividing vector into smaller sub-objects, 3) generating based on sub-objects. Taking classification as case compare other competing methods. results suggest outperforms all methods, it yields more contribute better representation objects, thus leading higher accuracy. Further experiments efficiency can be improved when multi-thread technology is implemented.