Fine-resolution spatial analytics are essential for urban planning and policy-making, yet traditional small-area estimation often struggles with sparse, hierarchical, or imbalanced data. This paper introduces a Spatially Regularised Bayesian Heterogeneous Graph Neural Network (SR-BHGNN) that integrates multiple census tract levels within a unified framework. The model builds a heterogeneous graph where nodes represent spatial units at different scales, edges encode adjacency or membership, and Bayesian inference quantifies uncertainty in parameters and predictions. A spatial regularisation term, inspired by Tobler’s First Law of Geography, penalises large discrepancies between neighbouring nodes, reducing errors in imbalanced datasets and ensuring coherent local estimates. We evaluate SR-BHGNN through two London case studies, population estimation and PM 2.5 prediction, comparing it against random forests, single-level GNNs, and spatial hierarchical Bayesian estimation. SR-BHGNN achieves strong performance gains, with classification accuracies of 0.85 for population estimation and 0.81 for PM 2.5 prediction. Its Bayesian design produces posterior distributions that capture uncertainty, enabling policy-relevant insights into vulnerable neighbourhoods or priority intervention zones (e.g. low-emission areas). These results demonstrate that SR-BHGNN advances the state of the art in small-area estimation, offering a flexible, uncertainty-aware framework for diverse urban analytics applications.

Volunteered geographic information (VGI) often contains rich geolocations that are crucial for disaster response and post-disaster assessment. However, existing studies on VGI geolocalization have not fully used the potential of multi-source and multimodal data. In this paper, we constructed a multimodal disaster dataset (MultiIan) and developed two novel methods (i.e. StaGeo and TriGeo) to enhance the cross-view geolocalization accuracy of disaster-related VGI. MultiIan comprised VGI texts and images, street view imagery (SVI) and remote sensing imagery (RSI). Large language models (LLMs) were used to extract the implicit geoinformation from VGI texts for geotagging. StaGeo was developed using staged training with ConvNeXt and vision transformer (ViT), while TriGeo used VGI ↔ SVI ↔ RSI triple-objective joint training of the ViT based on DINOv2. Using SVI to link VGI and RSI, our methods significantly improved the geolocalization accuracy of VGI across various train–test splits in MultiIan. With a typical 8:2 data split, StaGeo achieved Recall@1, Recall@5, Recall@10 and Recall@1% of 54.93%, 71.27%, 77.93% and 80.33%, respectively. TriGeo further improved these metrics, achieving 62.87%, 85.55%, 90.54% and 90.89%, respectively. These findings demonstrate significant advancements in our cross-view geolocalization methods, enabling timely geolocation to support rapid decision-making in emergency response and promoting the broader application of GeoAI in geospatial analysis.

Representation learning (RL) techniques are widely adopted in areas such as natural language processing and computer vision, with prominent examples such as attention and ConvNet architectures. In comparison, many GeoAI works still rely on feature engineering or data conversion to represent spatial data (e.g., points, polylines, polygons, 3D building models, etc.) as features in formats that are easier for neural networks to handle. The neural network architectures remain unchanged, and the need for feature engineering has become a bottleneck for applying deep learning to new tasks in the age of big data. In this paper, we advocate the idea of developing learnable spatial representation modules, which not only enable spatial reasoning but also enable neural nets to directly consume (i.e., encoding) or generate (i.e., decoding) spatial data. We propose Spatial Representation Learning (SRL), a new general-purpose representation learning framework for spatial reasoning. We discuss the key challenges of spatial representation learning including multi-scale RL, continuous RL, shape-centric RL, noise-robust RL, heterogeneity-aware RL, and fairness-aware RL. We also discuss the critical role and potential of SRL in various geospatial subdomains and how this technique can lead to a new generation of GeoAI.

Bavaria Buildings is a large, analysis-ready dataset providing openly available co-registered 40cm aerial imagery of Upper Bavaria paired with building footprint information. The Bavaria Buildings dataset (BBD) contains 18205 orthophotos of 2500 × 2500 pixels, where each pixel covers 40cm × 40cm in space (Digitales Orthophoto 40cm - DOP40). The dataset has been pre-processed and co-registered and also provides a set of 5.5 million image tiles of 250 × 250 pixels ready for deep learning and image analysis tasks. For each image tile, we provide two segmentation masks; one based on the official building footprints (Hausumringe) data as published by the Free State of Bavaria and one based on a historic OpenStreetMap (OSM) extract dating to 2021. The dataset is ready for essential analysis tasks, such as detection, segmentation, instance extraction, footprint geometry extraction, multimodal localization, and multimodal data quality assessment of buildings in Bavaria. We plan to update the dataset with each major re-publication of the upstream data sources to foster change detection research in the future. The BBD is available at https://doi.org/10.14459/2023mp1709451.

The recent advance of adapting pre-trained task-agnostic artificial intelligence (AI) models leads to great successes in downstream tasks via fine-tuning, or low-resource (i.e., few-shot and zero-shot) learning. However, when adapting such pre-trained AI models to geographical applications, it is still challenging to find the "sweet spot" of the model's generalizability and specializability (e.g., geographic generalizability v.s. spatial heterogeneity). For instance, a building detection task may require vision models with different parameters across different geographic areas of the world. In this paper, we rethink this interesting topic, namely Geographical Generalizability of GeoAI models, with a case study of detecting OpenStreetMap (OSM) missing buildings across different countries in sub-Saharan Africa. We consider a real-world scenario, in which we first train a Single-Shot Multibox Detection (SSD) base model for OSM missing building detection in Kakola, Tanzania, where a previous humanitarian mapping project of OSM was organized to map all possible buildings. Then we extrapolate this base model using Few-Shot Transfer Learning (FSTL) to a set of areas in the proximity of the test area in Cameroon. Here, we develop a Geographical Weighted Model Ensemble (GWME) method to improve Geographical Generalizability of GeoAI models. Moreover, we compare four unsupervised model ensemble weighting strategies: 1) Average weighting, 2) Image similarity weighting, 3) Geographical distance weighting, and 4) Self-attention-based weighting. Experiments show promising results of the proposed GWME method, which implicitly generates model weights from their location embedding and image feature embedding in an unsupervised manner. More specifically, the self-attention-based model ensemble achieves the highest performance. The results shed inspiring light on improving the generalizability and replicability of GeoAI models across geographic areas. Data and code are available at https://github.com/tum-bgd/GWME.

The ACM SIGSPATIAL Cup 2023 proposed the challenge to identify and map supraglacial lakes in Greenland in satellite imagery. The peculiarities of supraglacial lakes pose a hard problem for semantic segmentation and object detection tasks because the definition of a lake is ill-fitted to the inner workings of such approaches. For example, lakes are often covered by ice and snow and narrow streams can connect distinct lakes, which is not directly translatable to the semantic segmentation of water. It is also not well-posed for object detection, especially the identity relation - what is a lake, what is not (yet) a lake, and what are two lakes is challenging. In this context, we worked on adapting semantic segmentation using the Segment Anything Model and instance segmentation using Mask R-CNN to the setting. The latter ended up superior in our own evaluation and even got ranked second among all participants. We are proud that our approach has led to competitive performance. The source code is available from https://github.com/tum-bgd/GISCup23.

Determining the location of a building’s entrance is crucial to location-based services, such as wayfinding for pedestrians. Unfortunately, entrance information is often missing from current mainstream map providers such as Google Maps. Frequently, automatic approaches for detecting building entrances are based on street-level images that are not widely available. To address this issue, we propose a more general approach for inferring the main entrances of public buildings based on the association between spatial elements extracted from OpenStreetMap. In particular, we adopt three binary classification approaches, weighted random forest, balanced random forest, and smooth-boost to model the association relationship. There are two types of features considered in the classification: intrinsic features derived from building footprints and extrinsic features derived from spatial contexts, such as roads, green spaces, bicycle parking areas, and neighboring buildings. We conducted extensive experiments on 320 public buildings with an average perimeter of 350 m. The experimental results showed that the locations of building entrances estimated by the weighted random forest and balanced random forest models have a mean linear distance error of 21 m and a mean path distance error of 22 m, ruling out 90% of the incorrect locations of the main entrance of buildings.