Search NASA⌕ Search

SEARCH · Search NASA

Results for “SCORE SATELLITE”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Panel-Segmentation: A Python Package for Automated Solar Array Metadata Extraction Using Satellite Imagery

The NREL Python Panel-Segmentation package is a toolkit that automates the process of extracting accurate and valuable metadata related to solar array installations, using publicly available Google Maps satellite imagery. Previously published work includes automated azimuth estimation for individual solar installations in satellite images. Our continued research focuses on automated detection and classification of solar installation mounting configuration (tracking or fixed-tilt; rooftop, ground, or carport). Specifically, a Faster-RCNN Resnet-50 feature pyramid network (FPN) model was trained and validated on 862 manually labeled satellite images. This model was used to perform object detection on satellite imagery, locating and classifying individual solar installations' mounting configuration and type. Model results showed a mean average precision score (mAP) of 77.79%, with the model strongest at detecting fixed-tilt ground mount and fixed-tilt carport installations. The object detection model and its outputs have been incorporated into the Panel-Segmentation package's automated metadata extraction pipeline, which returns the mounting configuration and azimuth for individual solar arrays in satellite imagery. The complete image data set with labels has been released on the U.S. Department of Energy (DOE) DuraMAT DataHub, to encourage further research in this area.

deep learning↗
DOE OSTI2021

A spatial evaluation of Arctic sea ice and regional limitations in CMIP6 historical simulations

The Arctic sea ice response to a warming climate is assessed in a subset of models participating in Phase 6 of the Coupled Model Intercomparison Project (CMIP6), using several metrics in comparison with satellite observations and results from the Pan-Arctic Ice Ocean Modeling and Assimilation System and the Regional Arctic System Model. Our study examines the historical representation of sea ice extent, volume, and thickness using spatial analysis metrics, such as the integrated ice-edge error, Brier score, and Spatial Probability Score. We find that the CMIP6 multi-model mean captures the mean annual cycle and 1979-2014 sea ice trends remarkably well. However, individual models experience a wide range of uncertainty in the spatial distribution of sea ice when compared against satellite measurements and reanalysis data. Our metrics expose common and individual regional model biases, which sea ice temporal analyses alone do not capture. We identify large ice edge and ice thickness errors in Arctic sub-regions, implying possible model specific limitations in or lack of representation of some key physical processes. We postulate that many of them could be related to the oceanic forcing, especially in the marginal and shelf seas, where seasonal sea ice changes are not adequately simulated. We therefore conclude that an individual model’s ability to represent the observed/reanalysis spatial distribution still remains a challenge. We propose the spatial analysis metrics as useful tools to diagnose model limitations, narrow down possible processes affecting them, and guide future model improvements critical to the representation and projections of Arctic climate change.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2022

Estimating snow cover from high-resolution satellite imagery by thresholding blue wavelengths

We report the extent and duration of snow cover, a critical component of the hydrologic cycle and the global climate system, is expected to shift dramatically under climate change. Therefore, developing high-resolution assessments of snow cover change is crucial for estimating the impact of changing snow cover on watershed and ecosystems processes in cold regions. Remote sensing tools provide a powerful method for mapping snow-covered area (SCA) across a landscape. The most common method for estimating SCA utilizes the normalized difference snow index (NDSI), which relies on spectral measurements in the shortwave-infrared wavelengths (SWIR). NDSI can effectively estimate catchment- to regional-scale SCA, but it cannot be used to assess fine-scale SCA because of current limitations on the spatial resolution of satellite-derived SWIR measurements. Here, we map SCA using a threshold of blue wavelengths and high-resolution satellite imagery. The thresholding method, which we call the Blue Snow Threshold algorithm (BST), has previously been used with digital camera imagery. We refine and automate the algorithm for use with cloud-free high-resolution satellite imagery and find that the BST can be used to assess fine-scale SCA. For validation, we compared BST-derived estimates of SCA to a) airborne lidar surveys, b) Landsat fractional SCA, and c) snow disappearance dates from Snow Telemetry (SNOTEL) stations. When compared to airborne lidar surveys of SCA, the BST predicted SCA had a range of F-scores between 0.81 and 0.94 in four study areas in California and Colorado. We also found general agreement between SCA and snow disappearance at multiple SNOTEL sites across the western United States. Given the relatively recent availability of high-resolution satellite imagery with spectral measurements in the visible wavelengths but lacking in SWIR, the BST offers a reliable and easy-to-apply tool for examining fine-scale snow-related processes.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2023

Comparing quantile regression forest and mixture density long short-term memory models for probabilistic post-processing of satellite precipitation-driven streamflow simulations

Abstract. Deep learning (DL) and machine learning (ML) are widely used in hydrological modelling, which plays a critical role in improving the accuracy of hydrological predictions. However, the trade-off between model performance and computational cost has always been a challenge for hydrologists when selecting a suitable model, particularly for probabilistic post-processing with large ensemble members. This study aims to systematically compare the quantile regression forest (QRF) model and countable mixtures of asymmetric Laplacians long short-term memory (CMAL-LSTM) model as hydrological probabilistic post-processors. Specifically, we evaluate their ability in dealing with biased streamflow simulations driven by three satellite precipitation products across 522 nested sub-basins of the Yalong River basin in China. Model performance is comprehensively assessed using a series of scoring metrics from both probabilistic and deterministic perspectives. Our results show that the QRF model and the CMAL-LSTM model are comparable in terms of probabilistic prediction, and their performances are closely related to the flow accumulation area (FAA) of the sub-basin. The QRF model outperforms the CMAL-LSTM model in most sub-basins with smaller FAA, while the CMAL-LSTM model has an undebatable advantage in sub-basins with FAA larger than 60 000 km2 in the Yalong River basin. In terms of deterministic predictions, the CMAL-LSTM model is preferred, especially when the raw streamflow is poorly simulated and used as input. However, setting aside the differences in model performance, the QRF model with 100-member quantiles demonstrates a noteworthy advantage by exhibiting a 50 % reduction in computation time compared to the CMAL-LSTM model with the same ensemble members in all experiments. As a result, this study provides insights into model selection in hydrological post-processing and the trade-offs between model performance and computational efficiency. The findings highlight the importance of considering the specific application scenario, such as the catchment size and the required accuracy level, when selecting a suitable model for hydrological post-processing.

Geology↗
DOE OSTI2021

Detecting Arsenic Contamination Using Satellite Imagery and Machine Learning

Arsenic, a potent carcinogen and neurotoxin, affects over 200 million people globally. Current detection methods are laborious, expensive, and unscalable, being difficult to implement in developing regions and during crises such as COVID-19. This study attempts to determine if a relationship exists between soil’s hyperspectral data and arsenic concentration using NASA’s Hyperion satellite. It is the first arsenic study to use satellite-based hyperspectral data and apply a classification approach. Four regression machine learning models are tested to determine this correlation in soil with bare land cover. Raw data are converted to reflectance, problematic atmospheric influences are removed, characteristic wavelengths are selected, and four noise reduction algorithms are tested. The combination of data augmentation, Genetic Algorithm, Second Derivative Transformation, and Random Forest regression (R 2 =0.840 and normalized root mean squared error (re-scaled to [0,1]) = 0.122) shows strong correlation, performing better than past models despite using noisier satellite data (versus lab-processed samples). Three binary classification machine learning models are then applied to identify high-risk shrub-covered regions in ten U.S. states, achieving strong accuracy (=0.693) and F1-score (=0.728). Overall, these results suggest that such a methodology is practical and can provide a sustainable alternative to arsenic contamination detection.

63 RADIATION, THERMAL, AND OTHER ENVIRON. POLLUTAN↗
DOE OSTI2024

A globally sampled high-resolution hand-labeled validation dataset for evaluating surface water extent maps

Effective monitoring of global water resources is increasingly critical due to climate change and population growth. Advancements in remote sensing technology, specifically in spatial, spectral, and temporal resolutions, are revolutionizing water resource monitoring, leading to more frequent and high-quality surface water extent maps using various techniques such as traditional image processing and machine learning algorithms. However, satellite imagery datasets contain trade-offs that result in inconsistencies in performance, such as disparities in measurement principles between optical (e.g., Sentinel-2) and radar (e.g., Sentinel-1) sensors and differences in spatial and spectral resolutions among optical sensors. Therefore, developing accurate and robust surface water mapping solutions requires independent validations from multiple datasets to identify potential biases within the imagery and algorithms. However, high-quality validation datasets are expensive to build, and few contain information on water resources. For this purpose, we introduce a globally sampled, high-spatial-resolution dataset labeled using 3 m PlanetScope imagery. Our surface water extent dataset comprises 100 images, each with a size of 1024×1024 pixels, which were sampled using a stratified random sampling strategy covering all 14 biomes. We highlighted urban and rural regions, lakes, and rivers, including braided rivers and coastal regions. We evaluated two surface water extent mapping methods using our dataset – Dynamic World, based on Sentinel-2, and the NASA IMPACT model, based on Sentinel-1. Dynamic World achieved a mean intersection over union (IoU) of 72.16 % and F1 score of 79.70 %, while the NASA IMPACT model had a mean IoU of 57.61 % and F1 score of 65.79 %. Performance varied substantially across biomes, highlighting the importance of evaluating models on diverse landscapes to assess their generalizability and robustness. Our dataset can be used to analyze satellite products and methods, providing insights into their advantages and drawbacks. Our dataset offers a unique tool for analyzing satellite products, aiding the development of more accurate and robust surface water monitoring solutions. The dataset can be accessed via https://doi.org/10.25739/03nt-4f29.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Realistic galaxy image simulation via score-based generative models

ABSTRACT We show that a denoising diffusion probabilistic model (DDPM), a class of score-based generative model, can be used to produce realistic mock images that mimic observations of galaxies. Our method is tested with Dark Energy Spectroscopic Instrument (DESI) grz imaging of galaxies from the Photometry and Rotation curve OBservations from Extragalactic Surveys (PROBES) sample and galaxies selected from the Sloan Digital Sky Survey. Subjectively, the generated galaxies are highly realistic when compared with samples from the real data set. We quantify the similarity by borrowing from the deep generative learning literature, using the ‘Fréchet inception distance’ to test for subjective and morphological similarity. We also introduce the ‘synthetic galaxy distance’ metric to compare the emergent physical properties (such as total magnitude, colour, and half-light radius) of a ground truth parent and synthesized child data set. We argue that the DDPM approach produces sharper and more realistic images than other generative methods such as adversarial networks (with the downside of more costly inference), and could be used to produce large samples of synthetic observations tailored to a specific imaging survey. We demonstrate two potential uses of the DDPM: (1) accurate inpainting of occluded data, such as satellite trails, and (2) domain transfer, where new input images can be processed to mimic the properties of the DDPM training set. Here we ‘DESI-fy’ cartoon images as a proof of concept for domain transfer. Finally, we suggest potential applications for score-based approaches that could motivate further research on this topic within the astronomical community.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Systematic Validation of Ensemble Cloud-Process Simulations using Polarimetric Radar Observations and Simulator over the NASA Wallops Flight Facility

The BiLateral Operational Storm-Scale Observation and Modeling (BLOSSOM) project was initiated to establish a long-term supersite to improve understanding of cloud physical states and processes as well as to support satellite and climate model programs over the WFF site via a bilateral approach of storm-scale observations and process modeling. This study highlights a noble systematic validation framework of the BLOSSOM ensemble cloud-process simulations through mixed-phase, light-rain, and deep-convective precipitation cases. The framework consists of creating a domain-shifted ensemble of large-scale forcing datasets, and configuring and performing cloud-process simulations with three different bulk microphysics schemes. Validation uses NASA S-band dual-POLarimetric radar (NPOL) observations in the form of statistical composites and skill scores via a polarimetric radar simulator and newly developed CfRad Data tool (CfRAD). While the simulations capture the overall structures of the reflectivity composites, polarimetric signals are still poorly simulated, mainly due to a lack of representation of ice microphysics diversity in shapes, orientation distributions, and their complex mixtures. Despite the limitation, this new ensemble-based validation framework demonstrates that 1) no particular forcing or microphysics scheme outperforms the rest and 2) the skill scores of coarse- and fine-resolution ensemble simulations with different domain-shifted forcing and microphysics schemes are highly correlated with each other with no clear improvement. On the other hand, this suggests that coarse-resolution ensemble simulations are relevant for selecting the best meteorological forcing and microphysics scheme before conducting computationally demanding large eddy simulations (LESs) in support of aircraft and satellite instrument development as well as cloud-precipitation-convection parameterizations.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2023

Pyrocumulonimbus Events Over British Columbia in 2017: An Ensemble Model Study of Parameter Sensitivities and Climate Impacts of Wildfire Smoke in the Stratosphere

Abstract The Pyrocumulonimbus (pyroCb) events over British Columbia in 2017 were observed in the lower stratosphere for about 8–10 months after the smoke injections. Several previous studies used global climate models to investigate the physical parameters for the 2017 pyroCb events, but the conclusions show strong model dependency. In this study, we use the Energy Exascale Earth System Model (E3SM) and complete an ensemble of runs exploring three injection parameters: smoke aerosol mass, the percentage of black carbon within the smoke aerosols, and plume injection height. Additionally, we consider the heterogeneous reaction of ozone and primary organic matter. According to the satellite daily observed aerosol optical depth (AOD), we find that the best ensemble member is the simulation with 0.4 Tg of smoke, 3% of which is black carbon, a 13.5 km smoke injection height, and a 10 −5 probability factor of the heterogeneous reaction. Besides AOD, we examine the ensemble score based on the metrics used in previous studies: the metrics of the extinction coefficient at 18 km altitude and the maximum plume height. The conclusion of the best estimate of the injection parameters for 2017 pyroCb events shows strong not only model but also evaluation metric dependency. We use the Random Forest machine learning technique to quantify the relative importance of each parameter in accurately simulating the 2017 pyroCb events and find that the injection height is the most critical feature, no matter which metric is used to score the ensemble members.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

A machine learning pipeline for identifying infiltration managed aquifer recharge locations from satellite imagery in the San Joaquin Valley, California

This study focuses on an agricultural region in California’s Central Valley, USA, where Managed Aquifer Recharge (MAR) is widely implemented to mitigate groundwater depletion under increasing water demand and climate variability. A deep learning and machine learning framework was developed to identify infiltration-MAR locations using satellite imagery and environmental data. The framework integrates surface water detection from Sentinel-2 imagery, geospatial delineation of water bodies, spatiotemporal tracking of water body dynamics, and supervised classification using meteorological, environmental, and topographic variables. The framework was applied to a 2379 km² study area southwest of Fresno, where 765 water bodies were detected, including 139 identified MAR sites based on publicly available datasets and expert knowledge. The classification model achieved an accuracy of 0.94 and an F1 score of 0.85. Feature importance analysis indicates that cropland, normalized difference vegetation index (NDVI), and evaporation are among the most influential predictors for infiltration-MAR. Notably, the framework suggests that engineered water management in infiltration-MAR systems can disrupt or even reverse the expected positive correlation between surface water extent and precipitation. These findings provide physically interpretable insights into the characteristics of existing infiltration-MAR facilities and demonstrate the potential of the proposed framework as a reproducible, interpretable, and potentially transferable tool for data-driven infiltration-MAR identification and inventory development under growing climatic and hydrological uncertainty.

Classification↗
DOE OSTI2025

DeepAndes: A Self-Supervised Vision Foundation Model for Multispectral Remote Sensing Imagery of the Andes

By mapping sites at large scales usingremotely sensed data, archaeologists can generate unique insights into long-term demographic trends, interregional social networks, and human adaptations in the past. Remote sensing surveys complement field-based approaches, and their reach can be especially great when combined with deep learning and computer vision techniques. However, conventional supervised deep learning methods face challenges in annotating fine-grained archaeological features at scale. In addition, while recent vision foundation models have shown remarkable success in learning large-scale remote sensing data with minimal annotations, most off-the-shelf solutions are designed for RGB images rather than multispectral satellite imagery, such as the eight-band data used in our study. In this article, we introduce DeepAndes, a transformer-based vision foundation model trained on three million multispectral satellite images, specifically tailored for Andean archaeology. DeepAndes incorporates a customized DINOv2 self-supervised learning algorithm optimized for eight-band multispectral imagery, marking the first foundation model designed explicitly for the Andes region. We evaluate its image understanding performance through imbalanced image classification, image instance retrieval, and pixel-level semantic segmentation tasks. Our experiments show that DeepAndes achieves superior F1 scores, mean average precision, and Dice scores in few-shot learning scenarios, significantly outperforming models trained from scratch or pretrained on smaller datasets. This underscores the effectiveness of large-scale self-supervised pretraining in archaeological remote sensing.

Guo, Junlin [Vanderbilt Univ., Nashville, TN (Unit↗
DOE OSTI2022

The Global Atmosphere‐aerosol Model ICON‐A‐HAM2.3–Initial Model Evaluation and Effects of Radiation Balance Tuning on Aerosol Optical Thickness

Abstract The Hamburg Aerosol Module version 2.3 (HAM2.3) from the ECHAM6.3‐HAM2.3 global atmosphere‐aerosol model is coupled to the recently developed icosahedral nonhydrostatic ICON‐A (icon‐aes‐1.3.00) global atmosphere model to yield the new ICON‐A‐HAM2.3 atmosphere‐aerosol model. The ICON‐A and ECHAM6.3 host models use different dynamical cores, parameterizations of vertical mixing due to sub‐grid scale turbulence, and parameter settings for radiation balance tuning. Here, we study the role of the different host models for simulated aerosol optical thickness (AOT) and evaluate impacts of using HAM2.3 and the ECHAM6‐HAM2.3 two‐moment cloud microphysics scheme on several meteorological variables. Sensitivity runs show that a positive AOT bias over the subtropical oceans is remedied in ICON‐A‐HAM2.3 because of a different default setting of a parameter in the moist convection parameterization of the host models. The global mean AOT is biased low compared to MODIS satellite instrument retrievals in ICON‐A‐HAM2.3 and ECHAM6.3‐HAM2.3, but the bias is larger in ICON‐A‐HAM2.3 because negative AOT biases over the Amazon, the African rain forest, and the northern Indian Ocean are no longer compensated by high biases over the sub‐tropical oceans. ICON‐A‐HAM2.3 shows a moderate improvement with respect to AOT observations at AERONET sites. A multivariable bias score combining biases of several meteorological variables into a single number is larger in ICON‐A‐HAM2.3 compared to standard ICON‐A and standard ECHAM6.3. In the tropics, this multivariable bias is of similar magnitude in ICON‐A‐HAM2.3 and in ECHAM6.3‐HAM2.3. In the extra‐tropics, a smaller multivariable bias is found for ICON‐A‐HAM2.3 than for ECHAM6.3‐HAM2.3.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Quantifying physical parameterization uncertainties associated with land-atmosphere interactions in the WRF model over Amazon

The Weather Research and Forecasting (WRF) model can be used to diagnose regional land-atmosphere (L-A) coupling strength in the absence of sufficient observations but subjected to uncertainties associated with model physical parameterizations. In this study, we propose a framework to quantify and reduce model physical parameterization uncertainties associated with surface fluxes and L-A coupling. An ensemble of WRF simulations with different physical schemes is used to simulate surface fluxes and land-atmosphere coupling strength over the Amazon region. The physical parameterizations investigated include cloud microphysics (MP), land surface processes (LSM), planetary boundary layer (PBL), surface layer (SL), and cumulus (CU). We perform 120 ensemble simulations using the WRF model and different combinations of six MPs, three LSMs, six PBLs and SLs and three CUs. The measurements from the GoAMAZON field campaign and satellite data are used to evaluate model performance. A Multi-way analysis of variance (ANOVA) approach is applied to quantify the relative importance of different physics processes on L-A coupling. The Tukey’s test is used to sort schemes that have no significant differences into one group. The suite of physics that result in the best simulations of the corresponding variables are selected based on the Taylor skill score. Results show that the relative importance of processes and their interaction vary with the variables of interest. For example, CU was the most important process in modulating soil moisture, 2 m-humidity, latent heat, and net radiation. LSM showed dominant effects on 2 m-temperature and also has the largest impact on sensible heat and the lifting condensation level. The best physical parameterization ensembles show much narrower ranges of the variables of interest than the priori ensemble. Results of this study show the roles of different physical processes in modulating L-A interactions, quantify model uncertainties from physical processes, and provide insights for improving the model physics parameterizations.

Wang, Chen↗
DOE OSTI2023

Rapid Spaceborne Mapping of Wildfire Retardant Drops for Active Wildfire Management

Aerial application of fire retardant is a critical tool for managing wildland fire spread. Retardant applications are carefully planned to maximize fire line effectiveness, improve firefighter safety, protect high-value resources and assets, and limit environmental impact. However, topography, wind, visibility, and aircraft orientation can lead to differences between planned drop locations and the actual placement of the retardant. Information on the precise placement and areal extent of the dropped retardant can provide wildland fire managers with key information to (1) adaptively manage event resources, (2) assess the effectiveness of retardant slowing or stopping fire spread, (3) document location in relation to ecologically sensitive areas; and perform or validate cost-accounting for drop services. This study uses Sentinel-2 satellite data and commonly used machine learning classifiers to test an automated approach for detecting and mapping retardant application. We show that a multiclass model (retardant, burned, unburned, and cloud artifact classes) outperforms a single-class retardant model and that image differencing (post-application minus pre-application) outperforms single-image models. Compared to the random forest and support vector machine, the gradient boosting model performed the best with an overall accuracy of 0.88 and an F1 Score of 0.76 for fire retardant, though results were comparable for all three models. Our approach maps the full areal extent of the dropped retardant within minutes of image availability, rather than linear representations currently mapped by aerial GPS surveys. The development of this capability allows for the rapid assessment of retardant effectiveness and documentation of placement in relation to sensitive environments.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Harnessing Machine Learning and Data Fusion for Accurate Undocumented Well Identification in Satellite Images

This study utilizes satellite data to detect undocumented oil and gas wells, which pose significant environmental concerns, including greenhouse gas emissions. Three key findings emerge from the study. Firstly, the problem of imbalanced data is addressed by recommending oversampling techniques like Rotation–GaussianBlur–Solarization data augmentation (RGS), the Synthetic Minority Over-Sampling Technique (SMOTE), or ADASYN (an extension of SMOTE) over undersampling techniques. The performance of borderline SMOTE is less effective than that of the rest of the oversampling techniques, as its performance relies heavily on the quality and distribution of data near the decision boundary. Secondly, incorporating pre-trained models trained on large-scale datasets enhances the models’ generalization ability, with models trained on one county’s dataset demonstrating high overall accuracy, recall, and F1 scores that can be extended to other areas. This transferability of models allows for wider application. Lastly, including persistent homology (PH) as an additional input improves performance for in-distribution testing but may affect the model’s generalization for out-of-distribution testing. A careful consideration of PH’s impact on overall performance and generalizability is recommended. Overall, this study provides a robust approach to identifying undocumented oil and gas wells, contributing to the acceleration of a net-zero economy and supporting environmental sustainability efforts.

SMOTE↗
DOE OSTI2023

Automatic detection of ship-induced cloud features in satellite imagery

Ships crossing the ocean are known to produce long, curvilinear features called ship tracks visible in satellite imagery via the Twomey effect; however, there has been little exploitation of satellite imagery for broad atmospheric studies or global monitoring of ship emissions due to the difficulty of automated ship track detection. Prior studies are either proof-of-concept, qualitatively assessed, or restricted to a certain time of day. We propose a statistical method for the automated identification of ship tracks and demonstrate using GOES-West ABI data. We first present a human-assisted segmentation method, which we use to generate a ground truth data set of 529 annotated ship tracks in GOES-West ABI products. We then describe a two-stage automated approach comprising a detection stage to generate ship track proposals and a classification stage to reduce false positives. For detection, we present a novel pipeline based around a z-score filtering technique, and for classification, we demonstrate several classifiers from literature. In a final experiment, we quantitatively tune the detection parameters and train the classifier using the ground truth dataset, then test on a sequestered set of images; the detect-then-classify system had an overall Pd of 0.68 and 0.80 for daytime and nighttime data, respectively, and the classifier reduced false positive detections by 67% and 75%.

47 OTHER INSTRUMENTATION↗
DOE OSTI2023

Scaling Resolution of Gigapixel Whole Slide Images Using Spatial Decomposition on Convolutional Neural Networks

Gigapixel images are prevalent in scientific domains ranging from remote sensing, and satellite imagery to microscopy, etc. However, training a deep learning model at the natural resolution of those images has been a challenge in terms of both, overcoming the resource limit (e.g. HBM memory constraints), as well as scaling up to a large number of GPUs. In this paper, we trained Residual neural Networks (ResNet) on 22,528 x 22,528-pixel size images using a distributed spatial decomposition method on 2,304 GPUs on the Summit Supercomputer. We applied our method on a Whole Slide Imaging (WSI) dataset from The Cancer Genome Atlas (TCGA) database. WSI images can be in the size of 100,000 x 100,000 pixels or even larger, and in this work we studied the effect of image resolution on a classification task, while achieving state-of-the-art AUC scores. Moreover, our approach doesn't need pixel-level labels, since we're avoiding patching from the WSI images completely, while adding the capability of training arbitrary large-size images. This is achieved through a distributed spatial decomposition method, by leveraging the non-block fat-tree interconnect network of the Summit architecture, which enabled GPU-to-GPU direct communication. Finally, detailed performance analysis results are shown, as well as a comparison with a data-parallel approach when possible.

Tsaris, Aristeidis (aris)↗
DOE OSTI2021

Iterative self-organizing SCEne-LEvel sampling (ISOSCELES) for large-scale building extraction

Convolutional neural networks (CNN) provide state-of-the-art performance in many computer vision tasks, including those related to remote-sensing image analysis. Successfully training a CNN to generalize well to unseen data, however, requires training on samples that represent the full distribution of variation of both the target classes and their surrounding contexts. With remote sensing data, acquiring a sufficiently representative training set is a challenge due to both the inherent multi-modal variability of satellite or aerial imagery and the general high cost of labeling data. To address this challenge, we have developed ISOSCELES, an Iterative Self-Organizing SCEne LEvel Sampling method for hierarchical sampling of large image sets. Using affinity propagation, ISOSCELES automates the selection of highly representative training images. Compared to random sampling or using available reference data, the distribution of the training is principally data driven, reducing the chance of oversampling uninformative areas or undersampling informative ones. In comparison to manual sample selection by an analyst, ISOSCELES exploits descriptive features, spectral and/or textural, and eliminates human bias in sample selection. Using a hierarchical sampling approach, ISOSCELES can obtain a training set that reflects both between-scene variability, such as in viewing angle and time of day, and within-scene variability at the level of individual training samples. We verify the method by demonstrating its superiority to stratified random sampling in the challenging task of adapting a pre-trained model to a new image and spatial domain for country-scale building extraction. Using a pair of hand-labeled training sets comprising 1,987 sample image chips, a total of 496,000,000 individually labeled pixels, we show, across three distinct model architectures, an increase in accuracy, as measured by F1-score, of 2.2–4.2%.

42 ENGINEERING↗