An improved self-training network for building and road extraction in urban areas by integrating optical and radar remotely sensed data

Enhancing urban management and planning necessitates the automated and accurate extraction of features from remotely sensed images. To address the challenges in urban feature extraction, including issues such as cloud cover, shadows, diverse building roof types, and similar spectral behaviors associated with various urban covers, a strategic approach involves the integration of different remote sensing data sources. In this study, a novel deep learning-based framework is proposed to integrate optical and synthetic aperture radar (SAR) remote sensing data for automatic extraction of buildings and roads by reducing the dependence on the training datasets. First, the proposed system based on the deep convolutional encoder/decoder architecture is trained on a small training dataset (user-provided) and the network labels the unlabeled dataset. In the next step, a decision tree has been used to prevent wrongly labeled pixels from entering the training process to increase the reliability of labeled pixels. Then the labels and training dataset are updated, and the network is retrained on this data. The efficiency and reliability of the proposed system were assessed across four distinct regions. Through an iterative refinement process designed to enhance system outcomes, a consistent improvement in the reliability of the training dataset was observed across different datasets. Notably, an accuracy evaluation (F1) score exhibited substantial enhancements, with an 11% improvement for building extraction and a 10% increase for road extraction in the first dataset. In the second dataset, a remarkable 32% improvement for building extraction and a notable 63% increase for road extraction were observed. The third dataset showcased an 11% boost for building extraction and a substantial 20% increase for road extraction. Furthermore, the performance of the proposed system was systematically benchmarked against three prevalent deep learning methods. Across all datasets, the proposed method consistently demonstrated comparable and superior performance, substantiating its efficacy in varied experimental contexts. These findings underscore the robustness and effectiveness of the proposed system in advancing feature extraction methodologies.