A number of recent studies showed that digital footprints around built environments, such as geo-located tweets, are promising data sources for characterizing urban land use. However, challenges for achieving this purpose exist due to the volume and unstructured nature of geo-located social media. Previous studies focused on analyzing Twitter data collectively resulting in coarse resolution maps of urban land use. We argue that the complex spatial structure of a large collection of tweets, when viewed through the lens of individual-level human mobility patterns, can be simplified to a series of key locations for each user, which could be used to characterize urban land use at a higher spatial resolution. Contingent issues that could affect our approach, such as Twitter users’ biases and tendencies at locations where they tweet the most, were systematically investigated using 39 million geo-located Tweets and two independent datasets of the City of Chicago: 1) travel survey and 2) parcel-level land use map. Our results support that the majority of Twitter users show a preferential return, where their digital traces are clustered around a few key locations. However, we did not find a general relation among users between the ranks of locations for an individual—based on the density of tweets—and their land use types. On the contrary, temporal patterns of tweeting at key locations were found to be coherent among the majority of users and significantly associated with land use types of these locations. Furthermore, we used these temporal patterns to classify key locations into generic land use types with an overall classification accuracy of 0.78. The contribution of our research is twofold: a novel approach to resolving land use types at a higher resolution, and in-depth understanding of Twitter users’ location-related and temporal biases, promising to benefit human mobility and urban studies in general.

The lack of quality label data is considered one of the main bottlenecks for training machine and deep learning models. Weakly supervised learning using incomplete, coarse, or inaccurate data is an alternative strategy to overcome the scarcity of training data. We trained a U-Net model for segmenting Buildings’ footprints from a high-resolution digital elevation model, using existing label data from the open-access Microsoft building footprints data set. Comparison using an independent, manually labeled benchmark indicated the success of the weak supervision learning as the quality of the model prediction (IoU: 0.876) surpassed that of the original Microsoft data quality (IoU: 0.672) by approximately 20 percent. Moreover, adding extra channels such as elevation derivatives, slope, aspect, and profile curvatures did not enhance the weak learning process as the model learned directly from the original elevation data. Our results demonstrate the value of using existing data for training deep learning models even if they are noisy and incomplete.

The tumor microenvironment comprises different cell populations that interact, contributing to tumor heterogeneity and therapy response. Spatial transcriptomics offers valuable insights into the transcriptional complexity and heterogeneity of the tumor microenvironment. In this study, we established a geographic information system (GIS)–augmented in silico reconstruction of tumor architecture (GIS-ROTA), a biologically informed analytic framework that integrates pathway or cell type–based enrichment analysis with spatial autocorrelation measurement to uncover functional spatial domains. The approach considers biological functions prior to identifying any spatial domains, providing direct interpretability and minimizing the subjectivity of interpreting clusters observed from conventional analytic methods. The application of GIS-ROTA to a Visium spatial transcriptomics dataset of primary and metastatic estrogen receptor–positive breast tumor samples revealed extensive colocalization of estrogen response with metabolic pathway gene sets and mutual exclusivity with metastasis-related and specific immune-related pathway gene sets. Overall, the GIS-ROTA framework integrates biological knowledge first, yielding spatial patterns with functional relevance and enabling the identification of novel targets for the development of therapeutic strategies. Significance: GIS-ROTA, a spatial transcriptomics analytical framework that maps tissues with biologically curated pathways and signatures followed by spatial localization analysis, yields biologically relevant spatial domains, enabling cancer target and biomarker discovery.

Animal-borne cameras are increasingly used to study animal behavior. Here, we assessed the utility of three machine-learning models – Resnet-50 (3 epochs), Resnet-50 (10 epochs) and Vision Transformer (ViT) (3 epochs) – for identifying breathing behavior from animal-borne camera footage from green turtles (Chelonia mydas). The ViT model had mean Accuracy (97.2%), Precision (63.3%) and F1 (72.7%) scores that outperformed the Resnet-50 models, while all models had a Recall of >99.9%. Thus, the ViT model correctly identified almost all breathing frames although false positives (apnea frames labeled as breathing) were relatively common and led to an over-estimation of breathing rates. We conclude that ViT models are a promising solution for behavioral classification of animal-borne camera footage and even if not yet capable of the fully automated calculation of breathing events in sea turtles, they can still massively reduce the quantity of footage that needs to be manually checked and labeled.

This study integrates geospatial analysis with machine learning to examine the spatial dynamics of food insecurity among Syrian refugees in Lebanon. Using household survey data and novel geospatial indicators from 2018 to 2022 for 22,626 households, we investigate why certain food security measures are effective in specific contexts while others are not. Our findings reveal that geolocational factors play a significant role in shaping food insecurity outcomes, often overshadowing traditional factors like household sociodemographics and living conditions. This suggests a potential shift from labor-intensive socioeconomic survey methods toward more scalable, geospatially driven approaches. The study also highlights considerable variation in food insecurity across locations and subpopulations, raising concerns about the effectiveness of individual measures like the Food Consumption Score (FCS), Household Dietary Diversity Score (HDDS), and Reduced Coping Strategies Index (rCSI) in capturing localized needs. By disaggregating food insecurity dimensions and understanding their spatial distribution, humanitarian and development organizations can better tailor their strategies, directing resources to areas where refugees face the most severe food challenges. From a policy perspective, our insights call for a refined approach that improves the predictive power of food insecurity models, aiding organizations in efficiently targeting interventions.

Recent studies on human mobility show that human movements are not random and tend to be clustered. In this connection, the movements of Twitter users captured by geo-located tweets were found to follow similar patterns, where a few geographic locations dominate the tweeting activity of individual users. However, little is known about the semantics (landuse types) and temporal tweeting behavior at those frequently-visited locations. Furthermore, it is generally assumed that the top two visited locations for most of the users are home and work locales (Hypothesis A) and people tend to tweet at their top locations during a particular time of the day (Hypothesis B). In this paper, we tested these two frequently cited hypotheses by examining the tweeting patterns of more than 164,000 unique Twitter users whom were residents of the city of Chicago during 2014. We extracted landuse attributes for each geo-located tweet from the detailed inventory of the Chicago Metropolitan Agency for Planning. Top-visited locations were identified by clustering semantic enriched tweets using a DBSCAN algorithm. Our results showed that although the top two locations are likely to be residential and occupational/educational, a portion of the users deviated from this case, suggesting that the first hypothesis oversimplify real-world situations. However, our observations indicated that people tweet at specific times and these temporal signatures are dependent on landuse types. We further discuss the implication of confounding variables, such as clustering algorithm parameters and relative accuracy of tweet coordinates, which are critical factors in any experimental design involving Twitter data.