Chia-Yu Hsu

Chia-Yu Hsu, Wenwen Li, Sizhe Wang

Research on geospatial foundation models (GFMs) has become a trending topic in geospatial artificial intelligence (AI) research due to their potential for achieving high generalizability and domain adaptability, reducing model training costs for individual researchers. Unlike large language models, such as ChatGPT, constructing visual foundation models for image analysis, particularly in remote sensing, encountered significant challenges such as formulating diverse vision tasks into a general problem framework. This paper evaluates the recently released NASA-IBM GFM Prithvi for its predictive performance on high-level image analysis tasks across multiple benchmark datasets. Prithvi was selected because it is one of the first open-source GFMs trained on time-series of high-resolution remote sensing imagery. A series of experiments were designed to assess Prithvi's performance as compared to other pre-trained task-specific AI models in geospatial image analysis. New strategies, including band adaptation, multi-scale feature generation, and fine-tuning techniques, are introduced and integrated into an image analysis pipeline to enhance Prithvi's domain adaptation capability and improve model performance. In-depth analyses reveal Prithvi's strengths and weaknesses, offering insights for both improving Prithvi and developing future visual foundation models for geospatial tasks.

Wenwen Li, Chia-Yu Hsu, Sizhe Wang et al.

This paper introduces a real-time GeoAI workflow for large-scale image analysis and the segmentation of Arctic permafrost features at a fine-granularity. Very high-resolution (0.5m) commercial imagery is used in this analysis. To achieve real-time prediction, our workflow employs a lightweight, deep learning-based instance segmentation model, SparseInst, which introduces and uses Instance Activation Maps to accurately locate the position of objects within the image scene. Experimental results show that the model can achieve better accuracy of prediction at a much faster inference speed than the popular Mask-RCNN model.

Wenwen Li, Hyunho Lee, Sizhe Wang et al.

Vision foundation models are a new frontier in Geospatial Artificial Intelligence (GeoAI), an interdisciplinary research area that applies and extends AI for geospatial problem solving and geographic knowledge discovery, because of their potential to enable powerful image analysis by learning and extracting important image features from vast amounts of geospatial data. This paper evaluates the performance of the first-of-its-kind geospatial foundation model, IBM-NASA's Prithvi, to support a crucial geospatial analysis task: flood inundation mapping. This model is compared with convolutional neural network and vision transformer-based architectures in terms of mapping accuracy for flooded areas. A benchmark dataset, Sen1Floods11, is used in the experiments, and the models' predictability, generalizability, and transferability are evaluated based on both a test dataset and a dataset that is completely unseen by the model. Results show the good transferability of the Prithvi model, highlighting its performance advantages in segmenting flooded areas in previously unseen regions. The findings also indicate areas for improvement for the Prithvi model in terms of adopting multi-scale representation learning, developing more end-to-end pipelines for high-level image analysis tasks, and offering more flexibility in terms of input data bands.

Chia-Yu Hsu, Wenwen Li

Improving the interpretability of geospatial artificial intelligence (GeoAI) models has become critically important to open the "black box" of complex AI models, such as deep learning. This paper compares popular saliency map generation techniques and their strengths and weaknesses in interpreting GeoAI and deep learning models' reasoning behaviors, particularly when applied to geospatial analysis and image processing tasks. We surveyed two broad classes of model explanation methods: perturbation-based and gradient-based methods. The former identifies important image areas, which help machines make predictions by modifying a localized area of the input image. The latter evaluates the contribution of every single pixel of the input image to the model's prediction results through gradient backpropagation. In this study, three algorithms-the occlusion method, the integrated gradients method, and the class activation map method-are examined for a natural feature detection task using deep learning. The algorithms' strengths and weaknesses are discussed, and the consistency between model-learned and human-understandable concepts for object recognition is also compared. The experiments used two GeoAI-ready datasets to demonstrate the generalizability of the research findings.

Wenwen Li, Sizhe Wang, Hyunho Lee et al.

Landslides cause severe damage to lives, infrastructure, and the environment, making accurate and timely mapping essential for disaster preparedness and response. However, conventional deep learning models often struggle when applied across different sensors, regions, or under conditions of limited training data. To address these challenges, we present a three-axis analytical framework of sensor, label, and domain for adapting geospatial foundation models (GeoFMs), focusing on Prithvi-EO-2.0 for landslide mapping. Through a series of experiments, we show that it consistently outperforms task-specific CNNs (U-Net, U-Net++), vision transformers (Segformer, SwinV2-B), and other GeoFMs (TerraMind, SatMAE). The model, built on global pretraining, self-supervision, and adaptable fine-tuning, proved resilient to spectral variation, maintained accuracy under label scarcity, and generalized more reliably across diverse datasets and geographic settings. Alongside these strengths, we also highlight remaining challenges such as computational cost and the limited availability of reusable AI-ready training data for landslide research. Overall, our study positions GeoFMs as a step toward more robust and scalable approaches for landslide risk reduction and environmental monitoring.

Daniela Szwarcman, Sujit Roy, Paolo Fraccaro et al.

This technical report presents Prithvi-EO-2.0, a new geospatial foundation model that offers significant improvements over its predecessor, Prithvi-EO-1.0. Trained on 4.2M global time series samples from NASA's Harmonized Landsat and Sentinel-2 data archive at 30m resolution, the new 300M and 600M parameter models incorporate temporal and location embeddings for enhanced performance across various geospatial tasks. Through extensive benchmarking with GEO-Bench, the 600M version outperforms the previous Prithvi-EO model by 8\% across a range of tasks. It also outperforms six other geospatial foundation models when benchmarked on remote sensing tasks from different domains and resolutions (i.e. from 0.1m to 15m). The results demonstrate the versatility of the model in both classical earth observation and high-resolution applications. Early involvement of end-users and subject matter experts (SMEs) are among the key factors that contributed to the project's success. In particular, SME involvement allowed for constant feedback on model and dataset design, as well as successful customization for diverse SME-led applications in disaster response, land use and crop mapping, and ecosystem dynamics monitoring. Prithvi-EO-2.0 is available on Hugging Face and IBM terratorch, with additional resources on GitHub. The project exemplifies the Trusted Open Science approach embraced by all involved organizations.

Wenwen Li, Chia-Yu Hsu, Maosheng Hu

Recent interest in geospatial artificial intelligence (GeoAI) has fostered a wide range of applications using artificial intelligence (AI), especially deep learning, for geospatial problem solving. However, major challenges such as a lack of training data and the neglect of spatial principles and spatial effects in AI model design remain, significantly hindering the in-depth integration of AI with geospatial research. This paper reports our work in developing a deep learning model that enables object detection, particularly of natural features, in a weakly supervised manner. Our work makes three contributions: First, we present a method of object detection using only weak labels. This is achieved by developing a spatially explicit model based on Tobler's first law of geography. Second, we incorporate attention maps into the object detection pipeline and develop a multistage training strategy to improve performance. Third, we apply this model to detect impact craters on Mars, a task that previously required extensive manual effort. The model generalizes to both natural and human-made features on the surfaces of Earth and other planets. This research advances the theoretical and methodological foundations of GeoAI.

Hibiki Ito, Chia-Yu Hsu, Hiroaki Ogata

While secondary use of real-world data (RWD) in education offers substantial research opportunities, data sharing is often limited by privacy constraints. Differentially private synthetic data generation (DP-SDG) has emerged as a possible solution. However, educational RWD is fragmented across platforms and institutions and stored in different formats, so DP-SDG must be tailored to each dataset, requiring substantial engineering effort. In addition, such data are often small-sample and high-dimensional, making deep learning (DL)-based methods common but difficult to implement without specialist expertise. In this setting, it is also hard to achieve practically useful downstream utility. As a result, despite its theoretical promise, DP-SDG remains far from a practical solution in education. To address this issue, we propose a more practical two-stage method: (1) training-free, LLM-based DP-SDG is performed for sharing synthetic data and (2) on-demand real-data validation, where researchers submit code for remote validation of results. This simple method is designed for individual data custodians without extensive DP-SDG expertise. It can also be adapted to multi-shot synthesis, where data from different learner cohorts are synthesised regularly. We evaluate this method experimentally in both the one-shot and multi-shot synthesis settings using RWD collected over three years and conduct a case study with real researchers. Results show that LLM-based DP-SDG performs comparably to a DL-based baseline while greatly reducing engineering costs, and that non-DP validation causes measurable but moderate privacy leakage. Nonetheless, in the case study researchers reported that on average only 36% of synthetic findings are validated on real data. Overall, the paper provides a practical method for sharing educational RWD, while highlighting challenges in risk mitigation and epistemic precision.

Hibiki Ito, Chia-Yu Hsu, Hiroaki Ogata

The rapid adoption of digital technologies has greatly increased the volume of real-world data (RWD) in education. While these data offer significant opportunities for advancing learning analytics (LA), secondary use for research is constrained by privacy concerns. Differentially private synthetic data generation is regarded as the gold-standard approach to sharing sensitive data, yet studies on the private synthesis of educational data remain very scarce and rely predominantly on large, low-dimensional open datasets. Educational RWD, however, are typically high-dimensional and small in sample size, leaving the potential of private synthesis underexplored. Moreover, because educational practice is inherently iterative, data sharing is continual rather than one-off, making a traditional one-shot synthesis approach suboptimal. To address these challenges, we propose the Cyclic Adaptive Private Synthesis (CAPS) framework and evaluate it on authentic RWD. By iteratively sharing RWD, CAPS not only fosters open science, but also offers rich opportunities of design-based research (DBR), thereby amplifying the impact of LA. Our case study using actual RWD demonstrates that CAPS outperforms a one-shot baseline while highlighting challenges that warrant further investigation. Overall, this work offers a crucial first step towards privacy-preserving sharing of educational RWD and expands the possibilities for open science and DBR in LA.

Wenwen Li, Chia-Yu Hsu, Sizhe Wang et al.

GeoAI has emerged as an exciting interdisciplinary research area that combines spatial theories and data with cutting-edge AI models to address geospatial problems in a novel, data-driven manner. While GeoAI research has flourished in the GIScience literature, its reproducibility and replicability (R&R), fundamental principles that determine the reusability, reliability, and scientific rigor of research findings, have rarely been discussed. This paper aims to provide an in-depth analysis of this topic from both computational and spatial perspectives. We first categorize the major goals for reproducing GeoAI research, namely, validation (repeatability), learning and adapting the method for solving a similar or new problem (reproducibility), and examining the generalizability of the research findings (replicability). Each of these goals requires different levels of understanding of GeoAI, as well as different methods to ensure its success. We then discuss the factors that may cause the lack of R&R in GeoAI research, with an emphasis on (1) the selection and use of training data; (2) the uncertainty that resides in the GeoAI model design, training, deployment, and inference processes; and more importantly (3) the inherent spatial heterogeneity of geospatial data and processes. We use a deep learning-based image analysis task as an example to demonstrate the results' uncertainty and spatial variance caused by different factors. The findings reiterate the importance of knowledge sharing, as well as the generation of a "replicability map" that incorporates spatial autocorrelation and spatial heterogeneity into consideration in quantifying the spatial replicability of GeoAI research.

Wenwen Li, Chia-Yu Hsu, Sizhe Wang et al.

Retrogressive Thaw Slumps (RTS) in Arctic regions are distinct permafrost landforms with significant environmental impacts. Mapping these RTS is crucial because their appearance serves as a clear indication of permafrost thaw. However, their small scale compared to other landform features, vague boundaries, and spatiotemporal variation pose significant challenges for accurate detection. In this paper, we employed a state-of-the-art deep learning model, the Cascade Mask R-CNN with a multi-scale vision transformer-based backbone, to delineate RTS features across the Arctic. Two new strategies were introduced to optimize multimodal learning and enhance the model's predictive performance: (1) a feature-level, residual cross-modality attention fusion strategy, which effectively integrates feature maps from multiple modalities to capture complementary information and improve the model's ability to understand complex patterns and relationships within the data; (2) pre-trained unimodal learning followed by multimodal fine-tuning to alleviate high computing demand while achieving strong model performance. Experimental results demonstrated that our approach outperformed existing models adopting data-level fusion, feature-level convolutional fusion, and various attention fusion strategies, providing valuable insights into the efficient utilization of multimodal data for RTS mapping. This research contributes to our understanding of permafrost landforms and their environmental implications.

Chia-Yu Hsu, Shubhanshu Shekhar

We consider the problem of constructing sequential power-one tests where the null and alternative classes are specified indirectly through historical or offline data. More specifically, given an offline dataset consisting of observations from $L+1$ distributions $\{P_0, P_1, \ldots, P_L\}$, and a new unlabeled data stream $\{X_t: t \geq 1\} \overset{i.i.d}{\sim} P_θ$, the goal is to decide between the null $H_0: θ= 0$, against the alternative $H_1: θ\in [L]:=\{1,\ldots,L\}$. Our main methodological contribution is a general approach for designing a level-$α$ power-one test for this problem using a multi-class classifier trained on the given offline dataset. Working under a mild "separability" condition on the distributions and the trained classifier, we obtain an upper bound on the expected stopping time of our proposed level-$α$ test, and then show that in general this cannot be improved. In addition to rejecting the null, we show that our procedure can also identify the true underlying distribution almost surely. We then establish a sufficient condition to ensure the required separability of the classifier, and provide some converse results to investigate the role of the size of the offline dataset and the family of classifiers among classifier-based tests that satisfy the level-$α$ power-one criterion. Finally, we present an extension of our analysis for the training-and-testing distribution mismatch and illustrate an application to sequential change detection. Empirical results using both synthetic and real data provide support for our theoretical results.

Forouzan Fallah, Wenwen Li, Chia-Yu Hsu et al.

Super-resolution (SR) for remote sensing imagery often fails under out-of-distribution (OOD) conditions, such as rare geomorphic features captured by diverse sensors, producing visually plausible but physically inaccurate results. We present RareFlow, a physics-aware SR framework designed for OOD robustness. RareFlow's core is a dual-conditioning architecture. A Gated ControlNet preserves fine-grained geometric fidelity from the low-resolution input, while textual prompts provide semantic guidance for synthesizing complex features. To ensure physically sound outputs, we introduce a multifaceted loss function that enforces both spectral and radiometric consistency with sensor properties. Furthermore, the framework quantifies its own predictive uncertainty by employing a stochastic forward pass approach; the resulting output variance directly identifies unfamiliar inputs, mitigating feature hallucination. We validate RareFlow on a new, curated benchmark of multi-sensor satellite imagery. In blind evaluations, geophysical experts rated our model's outputs as approaching the fidelity of ground truth imagery, significantly outperforming state-of-the-art baselines. This qualitative superiority is corroborated by quantitative gains in perceptual metrics, including a nearly 40\% reduction in FID. RareFlow provides a robust framework for high-fidelity synthesis in data-scarce scientific domains and offers a new paradigm for controlled generation under severe domain shift.

Amal S. Perera, David Fernandez, Chandi Witharana et al.

Accurate mapping of permafrost landforms, thaw disturbances, and human-built infrastructure at pan-Arctic scale using sub-meter satellite imagery is increasingly critical. Handling petabyte-scale image data requires high-performance computing and robust feature detection models. While convolutional neural network (CNN)-based deep learning approaches are widely used for remote sensing (RS),similar to the success in transformer based large language models, Vision Transformers (ViTs) offer advantages in capturing long-range dependencies and global context via attention mechanisms. ViTs support pretraining via self-supervised learning-addressing the common limitation of labeled data in Arctic feature detection and outperform CNNs on benchmark datasets. Arctic also poses challenges for model generalization, especially when features with the same semantic class exhibit diverse spectral characteristics. To address these issues for Arctic feature detection, we integrate geospatial location embeddings into ViTs to improve adaptation across regions. This work investigates: (1) the suitability of pre-trained ViTs as feature extractors for high-resolution Arctic remote sensing tasks, and (2) the benefit of combining image and location embeddings. Using previously published datasets for Arctic feature detection, we evaluate our models on three tasks-detecting ice-wedge polygons (IWP), retrogressive thaw slumps (RTS), and human-built infrastructure. We empirically explore multiple configurations to fuse image embeddings and location embeddings. Results show that ViTs with location embeddings outperform prior CNN-based models on two of the three tasks including F1 score increase from 0.84 to 0.92 for RTS detection, demonstrating the potential of transformer-based models with spatial awareness for Arctic RS applications.

Wenwen Li, Chia-Yu Hsu, Sizhe Wang et al.

This paper assesses trending AI foundation models, especially emerging computer vision foundation models and their performance in natural landscape feature segmentation. While the term foundation model has quickly garnered interest from the geospatial domain, its definition remains vague. Hence, this paper will first introduce AI foundation models and their defining characteristics. Built upon the tremendous success achieved by Large Language Models (LLMs) as the foundation models for language tasks, this paper discusses the challenges of building foundation models for geospatial artificial intelligence (GeoAI) vision tasks. To evaluate the performance of large AI vision models, especially Meta's Segment Anything Model (SAM), we implemented different instance segmentation pipelines that minimize the changes to SAM to leverage its power as a foundation model. A series of prompt strategies was developed to test SAM's performance regarding its theoretical upper bound of predictive accuracy, zero-shot performance, and domain adaptability through fine-tuning. The analysis used two permafrost feature datasets, ice-wedge polygons and retrogressive thaw slumps because (1) these landform features are more challenging to segment than manmade features due to their complicated formation mechanisms, diverse forms, and vague boundaries; (2) their presence and changes are important indicators for Arctic warming and climate change. The results show that although promising, SAM still has room for improvement to support AI-augmented terrain mapping. The spatial and domain generalizability of this finding is further validated using a more general dataset EuroCrop for agricultural field mapping. Finally, we discuss future research directions that strengthen SAM's applicability in challenging geospatial domains.

Chia-Yu Hsu, Wenwen Li

This paper reports a new solution of leveraging temporal classification to support weakly supervised object detection (WSOD). Specifically, we introduce raster scan-order techniques to serialize 2D images into 1D sequence data, and then leverage a combined LSTM (Long, Short-Term Memory) and CTC (Connectionist Temporal Classification) network to achieve object localization based on a total count (of interested objects). We term our proposed network LSTM-CCTC (Count-based CTC). This "learning from counting" strategy differs from existing WSOD methods in that our approach automatically identifies critical points on or near a target object. This strategy significantly reduces the need of generating a large number of candidate proposals for object localization. Experiments show that our method yields state-of-the-art performance based on an evaluation on PASCAL VOC datasets.