Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells

Abstract

Unsupervised text encoding models have recently fueled substantial progress in Natural Language Processing (NLP). The key idea is to use neural networks to convert words in texts to vector space representations (embeddings) based on word positions in a sentence and their contexts, which are suitable for end-to-end training of downstream tasks. We see a strikingly similar situation in spatial analysis, which focuses on incorporating both absolute positions and spatial contexts of geographic objects such as Points of Interest (POIs) into models. A general-purpose representation model for space is valuable for a multitude of tasks. However, no such general model exists to date beyond simply applying discretization or feedforward nets to coordinates, and little effort has been put into jointly modeling distributions with vastly different characteristics, which commonly emerges from GIS data. Meanwhile, Nobel Prize-winning Neuroscience research shows that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding and is critical for recognizing places and for path-integration. Therefore, we propose a representation learning model called Space2Vec to encode the absolute positions and spatial relationships of places. We conduct experiments on two real-world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Results show that because of its multiscale representations, Space2Vec outperforms well-established ML approaches such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification tasks. Detailed analysis shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec 's multi-scale representation can handle distributions at different scales. 1

Encoding Geospatial Vector Data for Deep Learning: LULC as a Use Case

“…Different backbones were used in related work for machine learning of geospatial vector data. In the case of [2], a multilayer perceptron (MLP) was used, [28] used an attention-based network (a Transformer), and [29] used a graph convolutional neural network (GCNN) to learn a feature representation of geo-object geometries. While GCNN have gained in popularity within the domain of GIScience, they come with two disadvantages.…”

Dynamic MLP for Fine-Grained Image Classification by Leveraging Geographical and Temporal Information

“…Then we summarize in fine-grained works that incorporate [12,18,26]. (b) Concatenation: Concatenate the image features with the multimodal features in a channel-by-channel fashion [29,35,37,38]. (c) Addition: Add both predictions from the last layer of the image and extra information for a joint prediction [8].…”

“…Methods using Additional Information: Besides visual information, researchers have included additional information to improve classification performance. Many existing works [29,31,35,37,38] combine the image feature with the additional multimodal feature directly through channelwise concatenation. [37] is the first to introduce multimodal features, e.g., images, ages, and dates, extracted from MLP backbone network by concatenating them together to make a joint prediction.…”

“…Several works have proposed the use of additional information in fine-grained image classification. [29,31,35,37,38] directly concatenate the image feature with the multimodal feature before the final classification head. Furthermore, the addition [8] and multiplication [28,38] operations are adopted to fuse the features or predictions from the last network layer.…”

Learning Large-scale Location Embedding From Human Mobility Trajectories with Graphs

“…The training labels are derived from large-scale geotagged documents such as tweets, check-ins, and images that are available from social sharing platforms. Mai et al [11] proposed Space2vec using an encoder-decoder framework to encode the absolute positions and spatial relationships of places based on POI information. But for large-scale location embedding studies, labels for all locations are expensive to get and the attributes are difficult to define because some locations may have multiple labels and unpopular locations have no label.…”