Accurate translation of aerial imagery to maps is a direction of great value and challenge in mapping, a method of generating maps that does not require using vector data as traditional mapping methods do. The tremendous progress made in recent years in image translation based on generative adversarial networks has led to rapid progress in aerial image-to-map translation. Still, the generated results could be better regarding quality, accuracy, and visual impact. This paper proposes a supervised model (SAM-GAN) based on generative adversarial networks (GAN) to improve the performance of aerial image-to-map translation. In the model, we introduce a new generator and multi-scale discriminator. The generator is a conditional GAN model that extracts the content and style space from aerial images and maps and learns to generalize the patterns of aerial image-to-map style transformation. We introduce image style loss and topological consistency loss to improve the model’s pixel-level accuracy and topological performance. Furthermore, using the Maps dataset, a comprehensive qualitative and quantitative comparison is made between the SAM-GAN model and previous methods used for aerial image-to-map translation in combination with excellent evaluation metrics. Experiments showed that SAM-GAN outperformed existing methods in both quantitative and qualitative results.

Forest fire regimes are changing as a function of increasing global weather extremes, socioeconomic development, and land use change. It is appropriate to use long-term time series satellite observations to better understand forest fire regimes. However, many studies that have analyzed the spatiotemporal characteristics of forest fires based on fire frequency have been inadequate. In this study, a set of metrics was derived from the VIIRS active fire data in China, from 2012 to 2021, through spatial extraction, spatiotemporal clustering, and spread reconstruction to obtain the frequency of forest fire spots (FFS), the frequency of forest fire events (FFE), the frequency of large forest fire events (LFFE), duration, burned area, and spread rate; these metrics were compared to explore the characteristics of forest fires at different spatiotemporal scales. The experimental results include 72.41 × 10⁴ forest fire spots, 7728 forest fire events, 1118 large forest fire events, and a burned area of 58.4 × 10⁴ ha. Forest fires present a significant spatiotemporal aggregation, with the most FFS and FFE in the Southern Region and the most severe LFFE and burned area in the Southwest Region. The FFS, FFE, and LFFE show a general decreasing trend on an annual scale, with occasional minor rebounds. However, the burned area had substantial rebounds in 2020. The high incidence of forest fires was concentrated from March to May. Additionally, 74.7% of the forest fire events had a duration of less than 5 days, while 25.3% of the forest fire events lasted more than 5 days. This helps us to understand the characteristics of more serious or higher risk forest fires. This study can provide more perspectives for exploring the characteristics of forest fires, and more data underpinning for forest fire prevention and management. This will contribute towards reasonable forest protection policies and a sustainable environment.

As a core task of computer vision perception, 3D scene understanding has received widespread attention. However, the current research mainly focuses on the semantic understanding task at the level of entity objects and often neglects the semantic relationships between objects in the scene. This paper proposes a 3D scene graph prediction model based on deep learning methods for scanned point cloud data of indoor scenes to predict the semantic graph about the class of entity objects and their relationships. The model uses a multi-scale pyramidal feature extraction network, MP-DGCNN, to fuse features with the learned category-related unbiased meta-embedding vectors, and the relationship inference of the scene graph uses an ENA-GNN network incorporating node and edge cross-attention; in addition, considering the long-tail distribution effect, a category grouping re-weighting scheme is used in the embedded prior knowledge and loss function. For the 3D scene graph prediction task, experiments on the indoor point cloud 3DSSG dataset show that the model proposed in this paper performs well compared with the latest baseline model, and the prediction effectiveness and accuracy are substantially improved.