In this letter, we address the problem of road extraction in Wildland–urban interface (WUI) areas. In recent years, with the great success of convolutional neural networks (CNNs) in various vision-related tasks, researchers have developed many CNN-based methods for road extraction on remote sensing images. Nevertheless, these methods mostly treat road extraction as a binary classification problem on semantic labeling. In WUI areas, the road is narrower and tends to be occluded by trees, which may result in the serious discontinuous problem of inferred road maps. To address this issue, we propose transforming the input representation of the binary classification map into a continuous signed distance map. In this way, our model is forced to predict the continuous distance representations and, thus, improve the spatial continuities of inferred roads. In addition, a real-value regression task is designed to train along with the original binary classification task to generate spatially continuous and semantically accurate road maps. Then, we conduct experiments on the public Massachusetts road data set and a homemade data set collected from Yajishan Mountain, Beijing, China. Finally, our proposed method achieves intersection-over-unions (IoUs) of 64.11% and 65.92% for the Massachusetts and WUI-Yajishan data sets, respectively, without any postprocessing. In addition, the ablation analysis shows that introducing the regression task on the proposed signed distance representation can effectively alleviate the problem of discontinuous road prediction. Furthermore, comparing with the state-of-the-art methods demonstrates the superiority of our method for road extraction in WUI areas.

Features from multiple scales can greatly benefit the semantic edge detection task if they are well fused. However, the prevalent semantic edge detection methods apply a fixed weight fusion strategy where images with different semantics are forced to share the same weights, resulting in universal fusion weights for all images and locations regardless of their different semantics or local context. In this work, we propose a novel dynamic feature fusion strategy that assigns different fusion weights for different input images and locations adaptively. This is achieved by a proposed weight learner to infer proper fusion weights over multi-level features for each location of the feature map, conditioned on the specific input. In this way, the heterogeneity in contributions made by different locations of feature maps and input images can be better considered and thus help produce more accurate and sharper edge predictions. We show that our model with the novel dynamic feature fusion is superior to fixed weight fusion and also the na\"ive location-invariant weight fusion methods, via comprehensive experiments on benchmarks Cityscapes and SBD. In particular, our method outperforms all existing well established methods and achieves new state-of-the-art.

This letter addresses the issue of accurate object detection in large-area remote sensing images. Although many convolutional neural network (CNN)-based object detection models can achieve high accuracy in small image patches, the models perform poorly in large-area images due to the large quantity of false and missing detections that arise from complex backgrounds and diverse groundcover types. To address this challenge, this letter proposes a sample update-based CNN (SUCNN) framework for object detection in large-area remote sensing images. The proposed framework contains two stages. In the first stage, a base model—single-shot multibox detector—is trained with the training data set. In the second stage, artificial composite samples are generated to update the training set. The parameters of the first-stage model are fine-tuned with the updated data set to obtain the second-stage model. The first- and second-stage models are evaluated using the large-area remote sensing image test set. Comparison experiments show the effectiveness and superiority of the proposed SUCNN framework for object detection in large-area remote sensing images.