It is of great significance to apply the object detection methods to automatically detect boulders from planetary images and analyze their distribution. This contributes to the selection of candidate landing sites and the understanding of the geological processes. This paper improves the state-of-the-art object detection method of YOLOv5 with attention mechanism and designs a pyramid based approach to detect boulders from planetary images. A new feature fusion layer has been designed to capture more shallow features of the small boulders. The attention modules implemented by combining the convolutional block attention module (CBAM) and efficient channel attention network (ECA-Net) are also added into YOLOv5 to highlight the information that contribute to boulder detection. Based on the Pascal Visual Object Classes 2007 (VOC2007) dataset which is widely used for object detection evaluations and the boulder dataset that we constructed from the images of Bennu asteroid, the evaluation results have shown that the improvements have increased the performance of YOLOv5 by 3.4% in precision. With the improved YOLOv5 detection method, the pyramid based approach extracts several layers of images with different resolutions from the large planetary images and detects boulders of different scales from different layers. We have also applied the proposed approach to detect the boulders on Bennu asteroid. The distribution of the boulders on Bennu asteroid has been analyzed and presented.

The classification and recognition of the shapes of buildings in map space play an important role in spatial cognition, cartographic generalization, and map updating. As buildings in map space are often represented as the vector data, research was conducted to learn the feature representations of the buildings and recognize their shapes based on graph neural networks. Due to the principles of graph neural networks, it is necessary to construct a graph to represent the adjacency relationships between the points (i.e., the vertices of the polygons shaping the buildings), and extract a list of geometric features for each point. This paper proposes a deep point convolutional network to recognize building shapes, which executes the convolution directly on the points of the buildings without constructing the graphs and extracting the geometric features of the points. A new convolution operator named TriangleConv was designed to learn the feature representations of each point by aggregating the features of the point and the local triangle constructed by the point and its two adjacency points. The proposed method was evaluated and compared with related methods based on a dataset consisting of 5010 vector buildings. In terms of accuracy, macro-precision, macro-recall, and macro-F1, the results show that the proposed method has comparable performance with typical graph neural networks of GCN, GAT, and GraphSAGE, and point cloud neural networks of PointNet, PointNet++, and DGCNN in the task of recognizing and classifying building shapes in map space.

Buildings are important entity objects of cities, and the classification of building shapes plays an indispensable role in the cognition and planning of the urban structure. In recent years, some deep learning methods have been proposed for recognizing the shapes of building footprints in modern electronic maps. Furthermore, their performance depends on enough labeled samples for each class of building footprints. However, it is impractical to label enough samples for each type of building footprint shapes. Therefore, the deep learning methods using few labeled samples are more preferable to recognize and classify the building footprint shapes. In this paper, we propose a relation network based method for the recognization of building footprint shapes with few labeled samples. Relation network, composed of embedding module and relation module, is a metric based few-shot method which aims to learn a generalized metric function and predict the types of the new samples according to their relation with the prototypes of these few labeled samples. To better extract the shape features of the building footprints in the form of vector polygons, we have taken the TriangleConv embedding module to act as the embedding module of the relation network. We validate the effectiveness of our method based on a building footprint dataset with 10 typical shapes and compare it with three classical few-shot learning methods in accuracy. The results show that our method performs better for the classification of building footprint shapes with few labeled samples. For example, the accuracy reached 89.40% for the 2-way 5-shot classification task where there are only two classes of samples in the task and five labeled samples for each class.

Pretrained large-scale protein language models, such as ESM-1b and ProtTrans, are becoming the fundamental infrastructure for various protein-related biological modeling tasks. Existing works use mainly pretrained protein language models in feature extraction. However, the knowledge contained in the embedding features directly extracted from a pretrained model is task-agnostic. To obtain task-specific feature representations, a reasonable approach is to fine-tune a pretrained model based on labeled datasets from downstream tasks. To this end, we investigate the fine-tuning of a given pretrained protein language model for protein secondary structure prediction tasks. Specifically, we propose a novel end-to-end protein secondary structure prediction framework involving the lightweight fine-tuning of a pretrained model. The framework first introduces a few new parameters for each transformer block in the pretrained model, then updates only the newly introduced parameters, and then keeps the original pretrained parameters fixed during training. Extensive experiments on seven test sets, namely, CASP12, CASP13, CASP14, CB433, CB634, TEST2016, and TEST2018, show that the proposed framework outperforms existing predictors and achieves new state-of-the-art prediction performance. Furthermore, we also experimentally demonstrate that lightweight fine-tuning significantly outperforms full model fine-tuning and feature extraction in enabling models to predict secondary structures. Further analysis indicates that only a few top transformer blocks need to introduce new parameters, while skipping many lower transformer blocks has little impact on the prediction accuracy of secondary structures.

Remote sensing image scene classification takes image blocks as classification units and predicts their semantic descriptors. Because it is difficult to obtain enough labeled samples for all classes of remote sensing image scenes, zero-shot classification methods which can recognize image scenes that are not seen in the training stage are of great significance. By projecting the image visual features and the class semantic features into the latent space and ensuring their alignment, the variational autoencoder (VAE) generative model has been applied to address remote-sensing image scene classification under a zero-shot setting. However, the VAE model takes the element-wise square error as the reconstruction loss, which may not be suitable for measuring the reconstruction quality of the visual and semantic features. Therefore, this paper proposes to augment the VAE models with the generative adversarial network (GAN) to make use of the GAN’s discriminator in order to learn a suitable reconstruction quality metric for VAE. To promote feature alignment in the latent space, we have also proposed cross-modal feature-matching loss to make sure that the visual features of one class are aligned with the semantic features of the class and not those of other classes. Based on a public dataset, our experiments have shown the effects of the proposed improvements. Moreover, taking the ResNet models of ResNet18, extracting 512-dimensional visual features, and ResNet50 and ResNet101, both extracting 2048-dimensional visual features for testing, the impact of the different visual feature extractors has also been investigated. The experimental results show that better performance is achieved by ResNet18. This indicates that more layers of the extractors and larger dimensions of the extracted features may not contribute to the image scene classification under a zero-shot setting.