Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Labeled Datasets”

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 37 records · Page 2

Expanding NeMO-Net Machine Learning Capabilities for Citizen Science

NASA NeMO-Net, the neural multi-modal observation and training network for global coral reef assessment, is an open-source deep convolutional neural network and interactive active learning training software aiming to accurately assess the present and past dynamics of coral reef ecosystems through determination of percent living cover and morphology as well as mapping of spatial distribution. We present an interactive citizen science video game, released this April, for desktop and iOS devices where users interactively label morphology classifications over mm-scale 3D coral reef imagery captured using diver photomosaic imagery, the UAV enabled NASA FluidCam instrument, and satellite datasets. To date, the application has had over 40,000 downloads and over60,000 unique coral reef classifications, each filtered through a user-based rating and expert evaluation system. We also present results from NeMO-Net’s convolutional neural network (CNN) models used to semantically segment 2D satellite imagery as well as projections of 3D coral reconstructions using user input data as training datasets. Fusing datasets using machine learning from multiple remote sensing platforms presents novel methodologies for assessing the health of coral ecosystems, which are critically endangered by a changing climate. In partnering with Mission Blue, the National Oceanic and Atmospheric Administration (NOAA), and the Living Oceans Foundation (LOF), NeMO-Net leverages an international consortium of subject matter experts to provide both proper training for citizen scientists and the generation of a labeled datasets to ingest into machine learning algorithms for global coral reef identification.

NeMO-Net↗

Multi-Mission Terrain Classifier for Safe Rover Navigation and Automated Science

We previously presented Soil Property and Object Classification (SPOC), a machine learning-based terrain classifier for Mars rovers, for automatically segmenting rover images by its surface type such as sand and bedrock. This paper presents a number of practical improvements to pave the way for potential future onboard deployment. First, we achieved 97.0% overall pixel accuracy, evaluated against the classification generated by human experts on images from Mars Science Laboratory (MSL) missions. The substantial increase in accuracy was primarily enabled by the sheer volume of data used for training; we created a new large-scale dataset of Martian terrain labels, namely AI4Mars, which contains more than 400k labels contributed by citizen scientists for 50k images taken by the Mars Exploration Rovers (MER) and Mars Science Laboratory (MSL) rover. Second, we demonstrated that SPOC can quickly adapt to a new mission landed on a previously unseen site. Specifically, we pretrained a model with MER and MSL data from the AI4Mars dataset and then adapted to the Mars 2020 Rover (M2020) by feeding a small volume of data between Sol 0 and 157; the adapted model was tested on Sol 200-203 and resulted in 84.2% overall pixel accuracy and 93.4% reliability (recall) for detecting sand, the most concerning class for rover’s traversability. Third, we found that pretraining can substantially mitigate the decline of accuracy over time. We showed that the performance of a SPOC model pretrained with the ImageNet dataset and then trained by MSL images only up to Sol 390 remains comparable to a model trained by images up to Sol 1689 on the test data after Sol 1689. Fourth, we reimplemented SPOC with a light-weight convolutional neural network (CNN), MobileNetV2, which typically runs within tens of milliseconds (ms) on mobile processors such as Qualcomm’s Snapdragon. Finally, we released the AI4Mars dataset to the public to encourage open innovation.

Ono, Masahiro↗

Cloud Mask Intercomparison eXercise (CMIX): An evaluation of cloud masking algorithms for Landsat 8 and Sentinel-2

Cloud cover is a major limiting factor in exploiting time-series data acquired by optical spaceborne remote sensing sensors. Multiple methods have been developed to address the problem of cloud detection in satellite imagery and a number of cloud masking algorithms have been developed for optical sensors but very few studies have carried out quantitative intercomparison of state-of-the-art methods in this domain. This paper summarizes results of the first Cloud Masking Intercomparison eXercise (CMIX) conducted within the Committee Earth Observation Satellites (CEOS) Working Group on Calibration & Validation (WGCV). CEOS is the forum for space agency coordination and cooperation on Earth observations, with activities organized under working groups. CMIX, as one such activity, is an international collaborative effort aimed at intercomparing cloud detection algorithms for moderate-spatial resolution (10–30 m) spaceborne optical sensors. The focus of CMIX is on open and free imagery acquired by the Landsat 8 (NASA/USGS) and Sentinel-2 (ESA) missions. Ten algorithms developed by nine teams from fourteen different organizations representing universities, research centers and industry, as well as space agencies (CNES, ESA, DLR, and NASA), are evaluated within the CMIX. Those algorithms vary in their approach and concepts utilized which were based on various spectral properties, spatial and temporal features, as well as machine learning methods. Algorithm outputs are evaluated against existing reference cloud mask datasets. Those datasets vary in sampling methods, geographical distribution, sample unit (points, polygons, full image labels), and generation approaches (experts, machine learning, sky images). Overall, the performance of algorithms varied depending on the reference dataset, which can be attributed to differences in how the reference datasets were produced. The algorithms were in good agreement for thick cloud detection, which were opaque and had lower uncertainties in their identification, in contrast to thin/semi-transparent clouds detection. Not only did CMIX allow identification of strengths and weaknesses of existing algorithms and potential areas of improvements, but also the problems associated with the existing reference datasets. The paper concludes with recommendations on generating new reference datasets, metrics, and an analysis framework to be further exploited and additional input datasets to be considered by future CMIX activities.

Sergii Skakun↗

Predicting the Seawater Chemistry of an Ocean World Using Machine Learning on Isotopic Measurements of Volatile CO2

Introduction: Given the long time intervals required for data transmission to and from ocean worlds targets, low bandwidth for data transmission, time required for data processing and analysis, and potentially extreme radiation environments (e.g., Europa), it is clear that ocean worlds missions will need more autonomous flight instruments and software in order to achieve established science goals. Protracted time intervals for data analysis (e.g., Europa Lander) strongly motivates the development of rapid, consistent and streamlined methods for interpreting data from flight mass spectrometers to e.g., determine how mass spectra from a plume or surface liquid/ice relates to the surface/subsurface. Since mass spectrometry also has the potential to correctly identify biosignatures[1], it is imperative that such methods for interpreting data are consistent and accurate. We used 848 isotope ratio mass spectra from laboratory analyses of CO2 that interacted with ocean worlds-relevant seawaters as a ‘training’ dataset for ‘unsupervised’ machine learning. In unsupervised learning, characteristics of the data are not labeled or linked, and any similarities found only result from the neural network. CO2 isotopologues analyzed for this dataset mimic the remote measurements of CO2 by a flight mass spectrometer, and are detailed in Theiling [2]. From this dataset, we used measured features of the spectra, such as retention time, intensity, and (isotopologue) mass ratios as inputs for our autoencoder neural network. Our neural network was trained to find similarities in these and other spectral features for seawaters of a particular composition and amount of initial CO2. Successful training then created an output of these similarities for various seawaters, which included MgSO4, Na2SO4, NaCl, MgCl2, KCl, and NaHCO3, and combinations of these salts. We then applied dimensionality reduction techniques such as Principal Component Analysis (PCA), T-Distributed Stochastic Neighbor Embedding (TSNE), and Uniform Manifold Approximation and Projection (UMAP) to demonstrate latent data features as a two-dimensional projection in a unitless, high-dimensional space. In this projection, a data point represents the combined effect of spectral features such as intensity, retention time, and isotope ratio. Our initial UMAP demonstrates data clustering (organization of the data by the neural network) based on the amount of CO2 that had initially interacted with each seawater. Further training using more ‘supervised’ learning techniques demonstrate strong clustering of preliminary data based on initial CO2 concentration, seawater chemical composition, and ionic strength (salinity). Our preliminary work therefore suggests that machine learning has the potential to identify compositional variants of an ocean world seawater based on mass spectra from volatile CO2 measurements. Acknowledgments: This work was funded through a Strategic Task Group at NASA Goddard Space Flight Center. The training dataset was collected through funding from the Oklahoma Space Grant Consortium. References: [1] Pappalardo, R. et al. (2013) Astrobiology, 13, 740–773. [2] Theiling (2020) Icarus, 114216.

Europa↗

A Quantitative Analysis on the Use of Supervised Machine Learning in Earth Science

Recent review papers (Ball et al., 2017; Reichstein et al., 2019) have investigated the opportunities and challenges in applying supervised machine learning (ML) techniques to Earth science problems. A common challenge is the lack of training (or labeled) data. Supervised ML, and especially deep learning (DL), require large training datasets. While there are large, open access Earth science archives, the data typically require preprocessing in preparation for supervised ML, frequently including manual labeling. Our objective is to understand the landscape of supervised ML in the Earth sciences, including which research communities have most rapidly adopted supervised ML, which algorithms are applied, and what data are used to train these algorithms. We conducted a literature survey of Earth science papers published during the last 10 years in journals from the American Geophysical Union (AGU), American Meteorological Society (AMS), the Institute of Electrical and Electronics Engineers(IEEE), and the Society of Photo-Optical Instrumentation Engineers (SPIE). We identified papers containing the terms ML, DL, or the names of individual supervised ML algorithms. "Earth science" is an additional required search term for IEEE and SPIE. We investigate trends in supervised ML usage during the 10-year study period, and manually analyzed AGU papers from 2018-2019 to enable deep-dive statistics.

Katrina S Virts↗

A Statistical Model to Predict the Extratropical Transition of Tropical Cyclones

This paper introduces a logistic regression model for the extratropical transition(ET) of tropical cyclones in the North Atlantic and the Western North Pacific, using elastic net regularization to select predictors and estimate coefficients.Predictors are chosen from the 1979-2017 best track and reanalysis datasets, and verification is done against the tropical/extratropical labels in the best track data. In an independent test set, the model skillfully predicts ET at lead times up to two days, with latitude and sea surface temperature as its most important predictors. At a lead time of 24 h, it predicts ET with a Matthews correlation coefficient of 0.4 in the North Atlantic, and 0.6 in the Western North Pacific. It identifies 80% of storms undergoing ET in the North Atlantic, and 92% of those in the Western North Pacific. 90% of transition time errors are less than 24 h. Select examples of the model's performance on individual storms illustrate its strengths and weaknesses.Two versions of the model are presented: an "operational model" that may provide baseline guidance for operational forecasts, and a "hazard model"that can be integrated into statistical TC risk models. As instantaneous diagnostics for tropical/extratropical status, both models' zero lead time predictions perform about as well as the widely used Cyclone Phase Space (CPS) in the Western North Pacific and better than the CPS in the North Atlantic, and predict the timings of the transitions better than CPS in both basins.

Melanie Bieli↗

Identification of Security related Bug Reports via Text Mining using Supervised and Unsupervised Classification

This paper is focused on automated classification of software bug reports to security and non-security related, using both supervised and unsupervised approaches. For both approaches, three types of feature vectors are used. For supervised learning, we experiment with multiple learning algorithms and training sets with different sizes. Furthermore, we propose a novel unsupervised approach based on anomaly detection. The evaluated is based on three NASA datasets. The results show that supervised classification is affected more by the learning algorithms than by feature vectors and using only 25% of the data for training provides as good results as if 90% of data are used for training. Both supervised and unsupervised learning can be used for identification of security bug reports; the former slightly outperforms the latter at the expense of labeling the testing set. In general, the performance differs across datasets, mainly due to the different amounts of security related information.

Goseva-Popstojanova, Katerina↗

Supervised Machine Learning Approach for Classifying Earth Science Publications

The data collections archived and distributed by the GES DISC NASA data center are widely utilized for various Earth Science studies. As these collections are created, many research works are published regarding these collections' algorithms, their validation, and their applications. As NASA data centers collect these publications for public use, it is helpful to categorize them based on how they relate to their associated datasets. Specifically, whether the publication linked to the GES DISC dataset is using it for applicational research, describing the algorithm used for the dataset creation, validating the dataset, or providing a general overview of the data collection. Currently, this process requires simple manual labeling, and as such, it may be possible to solve via automation. To approach this problem, machine learning classifiers were developed to predict a publication's category. Manually labeled publications were used as the training data for the supervised machine learning algorithms, specifically Random Forest and Multinomial Naïve Bayes. After balancing the dataset and implementing the Multinomial Naïve Bayes algorithm, the classification accuracy achieved was substantially higher than the baseline accuracy, thus significantly improving the efficiency of publication labeling.

Rohan Dayal↗

Analyzing EOSDIS Dataset Research Outputs using Knowledge Graphs and Large Language Models

Datasets, unlike publications, can be updated over time, with each new version receiving a DOI but not always being linked to previous ones. This complicates tracking citations across a dataset’s lifecycle. We address this by integrating dataset versions and citations into a knowledge graph (KG), which helps trace dataset citations and analyze dataset usage in applied research. To categorize publications from various journals, we fine-tuned NASA IMPACT INDUS Large Language Model (LLM) on a labeled publication set, assigning publications to one of twenty applied research areas. By linking datasets to these research areas, we improved dataset searchability and discovery through these domains.

open-source↗

Microstructure Segmentation With Deep Learning Encoders Pre-Trained on a Large Microscopy Dataset

This study examined the improvement of microscopy segmentation intersection over union accuracy by transfer learning from a large dataset of microscopy images called MicroNet. Many neural network encoder architectures were trained on over 100,000 labeled microscopy images from 54 material classes. These pre-trained encoders were then embedded into multiple segmentation architectures including UNet and DeepLabV3+ to evaluate segmentation performance on created benchmark microscopy datasets. Compared to ImageNet pre-training, models pre-trained on MicroNet generalized better to out-of-distribution micrographs taken under different imaging and sample conditions and were more accurate with less training data. When training with only a single Ni-superalloy image, pre-training on MicroNet produced a 72.2% reduction in relative intersection over union error. These results suggest that transfer learning from large in-domain datasets generate models with learned feature representations that are more useful for downstream tasks and will likely improve any microscopy image analysis technique that can leverage pre-trained encoders.

machine learning↗

Exploring Sentinel-1 and Sentinel-2 diversity for Flood inundation mapping using deep learning

Identification of flood water extent from satellite images has historically relied on either synthetic aperture radar (SAR) or multi-spectral (MS) imagery. MS sensors are limited to cloud free conditions, whereas SAR imagery is plagued by noise-like speckle. Prior studies that use combinations of MS and SAR data to overcome individual limitations of these sensors have not fully examined sensitivity of flood mapping performance to different combinations of SAR and MS derived spectral indices or band transformations in color space. This study explores the use of diverse bands of Sentinel 2 (S2) through well-established water indices and Sentinel 1 (S1) derived SAR imagery along with their combinations to assess their capability for generating accurate flood inundation maps. The robustness in performance of S-1 and S-2 band combinations was evaluated using 446 hand labeled flood inundation images spanning across 11 flood events from Sen1Floods11 dataset which are highly diverse in terms of land cover as well as location. A modified K-fold cross validation approach is used to evaluate the performance of 32 combinations of S1 and S2 bands using a fully connected deep convolutional neural network known as U-Net. Our results indicated that usage of elevation information has improved the capability of S1 imagery to produce more accurate flood inundation maps. Compared to a median F1 score of 0.62 when using only S1 bands, the combined use of S1 and elevation information led to an improved median F1 score of 0.73. Water extraction indices based on S2 bands have a statistically significant superior performance in comparison to S1. Among all the band combinations, HSV (Hue, Saturation, Value) transformation of S2 bands provides a median F1 score of 0.9, outperforming the commonly used water spectral indices owing to HSV’s transformation’s superior contrast distinguishing abilities. Additionally, U-Net algorithm was able to learn the relationship between raw S2 based water extraction indices and their corresponding raw S2 bands, but not of HSV owing to relatively complex computation involved in the latter. Results of the paper establishes important benchmarks for the extension of S1 and S2 data-based flood inundation mapping efforts over large spatial extents.

Goutam Konapala↗

Benchmark Models for Classification of Radiation Type Induced in Immune Cells

NASA Biological and Physical Sciences and the Science Mission Directorate have published a benchmark dataset of mouse immune cells subjected to radiation-induced DNA damage. The dataset comprises ML-ready microscopic imagery of said cells, including labels indicating radiation type and dose. The machine learning team at NASA Interagency Implementation and Advanced Concept Team (IMPACT) created multiple benchmark models. Initially, we conducted a preliminary analysis using thresholding. The algorithm used thresholds on average brightness of the available images to classify them into their respective radiation type. We also tested machine learning approaches. Convolutional Neural Networks (CNN) emerged as the best-performing model. This poster presents the benchmark scores obtained by the models.

Vishal Perekadan↗

GeoAI advances in specific landform mapping

Landform mapping (also referred to as geomorphology or geomorphometry) can be divided into two domains: general and specific (Evans 2012). Whereas general landform mapping categorizes all elements of the study area into landform classes, such as ridges, valleys, peaks, and depressions, the mapping of specific landforms requires the delineation (even if fuzzy) of individual landforms. The former is mainly driven by physical properties such as elevation, slope, and curvature. The latter, however, must consider the cognitive (human) reasoning that discriminates individual landforms in addition to these physical properties (Arundel and Sinha 2018). Both mapping forms are important. General geomorphometry is needed to understand geological and ecological processes and as boundary layer input to climate and environmental models. Specific geomorphometry supports such activities as disaster management and recovery, emergency response, transportation, and navigation. In the United States, individual landforms of interest are named in the U.S. Geological Survey (USGS) Geographic Names Information System, a point dataset captured specifically to digitize geographic names from the USGS Historical Topographic Map Collection (HTMC). Named landform extent is represented only by the name placement in the HTMC. Recent work has investigated CNN-based deep learning methods to capture these extents in machine-readable form. These studies first relied on physical properties (Arundel et al. 2020) and then included the HTMC as a band in RGB images in limited testing (Arundel et al. 2023). Results from the HTMC dataset surpassed those using just physical properties and using the HTMC alone performed best due to the hillshading and elevation (contour) data incorporated into the topographic maps. However, results fell short of an operational capacity to map all named landforms in the United States. Thus, our current work expands upon past research by focusing on the HTMC and physical information as inputs and the named landform label extents. Specifically, we propose to leverage pre-trained foundation models for segmentation and optical character recognition (OCR) models to jointly map landforms in the United States. Our approach aims to bridge the disparities among the independent information sources to facilitate informed decision-making. The modeling pipeline performs (1) segmentation using the physical information and (2) information extraction using OCR, in parallel. Then a computer vision approach merges the two branches into a labeled segmentation. References: Arundel, Samantha T., Wenwen Li, and Sizhe Wang. 2020. “GeoNat v1.0: A Dataset for Natural Feature Mapping with Artificial Intelligence and Supervised Learning.” Transactions in GIS 24 (3): 556–72. https://doi.org/10.1111/tgis.12633. Arundel, Samantha T, and Gaurav Sinha. 2018. “Validating GEOBIA Based Terrain Segmentation and Classification for Automated Delineation of Cognitively Salient Landforms BT - Proceedings of Workshops and Posters at the 13th International Conference on Spatial Information Theory (COSIT 2017).” In Proceedings of Workshops and Posters at the 13th International Conference on Spatial Information Theory (COSIT 2017), Lecture Notes in Geoinformation and Cartography, edited by Paolo Fogliaroni, Andrea Ballatore, and Eliseo Clementini, 9–14. Cham: Springer International Publishing. Arundel, Samantha T., Gaurav Sinha, Wenwen Li, David P. Martin, Kevin G. McKeehan, and Philip T. Thiem. 2023. “Historical Maps Inform Landform Cognition in Machine Learning.” Abstracts of the ICA 6 (August): 1–2. https://doi.org/10.5194/ica-abs-6-10-2023. Evans, Ian S. 2012. “Geomorphometry and Landform Mapping: What Is a Landform?” Geomorphology 137 (1): 94–106. https://doi.org/10.1016/j.geomorph.2010.09.029.

machine learning↗

Development of Machine Learning Algorithms to Segment and Study Images of Astromaterial Samples

Introduction: Micrometer-scale chemical analyses of chondritic meteorites and mission-returned asteroid samples can reveal details of the physical and chemical processes operating in the early solar system, including processes that gave rise to planets, moons, and minor bodies. These primitive astromaterials are comprised of chondrules, calcium- and aluminum-rich inclusions (CAI), and many other silicates, oxides, metals, sulfides, and fine-grained materials. The chemical and mineralogical complexity of these samples, vast populations of different components, and heterogeneity across mm to km scales, all limit our understanding of the origin and evolution of these materials. Here, we describe recent efforts to use machine learning techniques to automate the segmentation of chemical maps of chondritic meteorites, designed to aid studies of asteroid samples returned by spacecraft. By automating the task of segmentation it will become possible to rapidly analyze and interpret the sizes, shapes, mineralogy, chemistry, and other properties of every chondrule, calcium- and aluminum-rich inclusion (CAI) and other clast within and between asteroid samples. Sample return missions significantly accelerate and heighten the need to develop such new data analysis techniques, and associated data repositories. Techniques: Neural networks require abundant training data, i.e. images which have been segmented by a human user. We have manually segmented data available from previous petrologic and chemical work at NASA Johnson Space Center and the American Museum of Natural History [1-4]. These data were derived from energy- and wavelength-dispersive X-ray spectroscopy (EDS, WDS) mapping of samples from many chondrite groups. The Deeplabv3+ [5] neural network architecture was trained on human-labeled masks and used to create machine-labeled masks. Several different algorithms were investigated, with inputs ranging from common RGB image formats through to hyperspectral datasets, with raw data comprising greyscale maps of Mg, Ca, and Al, with or without Si, Fe, Ti for both EDS and WDS data, and extending to other elements in EDS only. Each greyscale image was paired with a binary mask for each labelled particle type. Results: The trained algorithms can segment (Fig 1), classify, and measure the dimensions of thousands of particles in chemical maps of a standard 1-inch round petrographic section in seconds to minutes, rather than many hours needed by a human. Accuracy of the algorithms varied from chondrite to chondrite and across particle types. Further results and details of the algorithms will be presented at the workshop. Future directions: Machine learning has the potential to revolutionize our understanding of complex particle populations contained within primitive astromaterial, with segmentation being a critical first step. Example applications include better understanding of particle transport, nebular reservoirs, parent body accretion, and a deeper understanding of the relationships between particle populations and bulk rock elemental and isotopic compositions. In addition to benefits that machine learning can bring to individual researchers, building a community data repository of thousands to millions of particles across hundreds of samples will open up many other possibilities. For example, with a large enough dataset it will be possible to search for exceptionally closely matching particles across disparate samples. Such a capability would enable a single CAI from OSIRISREx or Hayabusa/II samples to be matched to chondritic CAIs that exhibit near-identical size, texture, and mineralogy, down to the level of similar core phenocrysts, zonation, and rim sequences. Such comparative analyses will help to disentangle precursor chemistry, chronology, gas/dust reservoirs during heating, and accretion. Such an endeavor would be impossible without machine learning and a large community data repository of astromaterial chemical/mineralogic maps.

Machine Learning↗

AI4MARS: A Dataset for Terrain-Aware Autonomy on Mars

Deep learning has quickly become a necessity for selfdriving vehicles on Earth. In contrast, the self-driving vehicles on Mars, including NASA’s latest rover, Perseverance, which is planned to land on Mars in February 2021, are still driven by classical machine vision systems. Deep learning capabilities, such as semantic segmentation and object recognition, would substantially benefit the safety and productivity of ongoing and future missions to the red planet. To this end, we created the first large-scale dataset, AI4Mars, for training and validating terrain classification models for Mars, consisting of ~326K semantic segmentation full image labels on 35K images from Curiosity, Opportunity, and Spirit rovers, collected through crowdsourcing. Each image was labeled by ~10 people to ensure greater quality and agreement of the crowdsourced labels. It also includes ~1.5K validation labels annotated by the rover planners and scientists from NASA’s MSL (Mars Science Laboratory) mission, which operates the Curiosity rover, and MER (Mars Exploration Rovers) mission, which operated the Spirit and Opportunity rovers. We trained a DeepLabv3 model on the AI4Mars training dataset and achieved over 96% overall classification accuracy on the test set. The dataset is made publicly available.1

Ono, Hiro↗

Towards an Aviation Large Language Model by Fine-tuning and Evaluating Transformers

In the aviation domain, there are many applications for machine learning and artificial intelligence tools that utilize natural language. For example, there is a desire to know the commonalities in written safety reports such as voluntary post incidents reports or create more accurate transcripts of air traffic management conversations. Another use-case is the possibility of extracting airspace procedures and constraints currently written in documents such as Letters of Agreement (LOA) which is used as the evaluation case in this paper. These applications can benefit from the use of state-of-the-art Natural Language Processing (NLP) techniques when adapted to the language/phraseology specific to the aviation domain. This paper evaluates the viability of transferring pre-trained large language models to the aviation domain by adapting transformer based models using aviation datasets. This paper utilized two datasets to adapt a ‘Robustly Optimized Bidirectional Encoder Representations from Transformers Approach’ (RoBERTa) model and two down-stream classification tasks to assess its performance. These datasets are all built upon Letters of Agreement which are Federal Aviation Administration (FAA) documents that formalize airspace operations across the national airspace system. The first two datasets are used for the adaptation of RoBERTa to the aviation domain and were of different sizes to assess the number of documents needed to adapt to the aviation domain. They contain many examples of ‘aviation English’ using domain specific terminology and phrasing which serves as a representative basis to perform the unsupervised adaptation. The second dataset is a separate set of LOA documents with two sets of classification labels to be used for evaluation; one at the document level and one at the line level. These down-stream evaluations allowed the measurement of improvement by adapting RoBERTa. The accuracy increased by 4-6% on both tasks and the F1 score on the class of interest increased by 4-8% from the adaptation.

Air Traffic Management↗

Towards an Aviation Large Language Model by Fine-tuning and Evaluating Transformers

In the aviation domain, there are many applications for machine learning and artificial intelligence tools that utilize natural language. For example, there is a desire to know the commonalities in written safety reports such as voluntary post incidents reports or create more accurate transcripts of air traffic management conversations. Another use-case is the possibility of extracting airspace procedures and constraints currently written in documents such as Letters of Agreement (LOA) which is used as the evaluation case in this paper. These applications can benefit from the use of state-of-the-art Natural Language Processing (NLP) techniques when adapted to the language/phraseology specific to the aviation domain. This paper evaluates the viability of transferring pre-trained large language models to the aviation domain by adapting transformer based models using aviation datasets. This paper utilized two datasets to adapt a ‘Robustly Optimized Bidirectional Encoder Representations from Transformers Approach’ (RoBERTa) model and two down-stream classification tasks to assess its performance. These datasets are all built upon Letters of Agreement which are Federal Aviation Administration (FAA) documents that formalize airspace operations across the national airspace system. The first two datasets are used for the adaptation of RoBERTa to the aviation domain and were of different sizes to assess the number of documents needed to adapt to the aviation domain. They contain many examples of ‘aviation English’ using domain specific terminology and phrasing which serves as a representative basis to perform the unsupervised adaptation. The second dataset is a separate set of LOA documents with two sets of classification labels to be used for evaluation; one at the document level and one at the line level. These down-stream evaluations allowed the measurement of improvement by adapting RoBERTa. The accuracy increased by 4-6% on both tasks and the F1 score on the class of interest increased by 4-8% from the adaptation.

Air Traffic Management↗

Microstructure Segmentation with Deep Learning Encoders Pre-Trained on a Large Microscopy Dataset

This study examined the improvement of microscopy segmentation accuracy by transfer learning from a large dataset of microscopy images called MicroNet. Many neural network encoder architectures, including VGG, Inception, and ResNet, were trained on over 100,000 labelled microscopy images from 54 classes. These pre-trained encoders were then embedded into multiple segmentation architectures including U-Net and DeepLabV3+ to evaluate segmentation performance on newly created benchmark microscopy datasets. Compared to ImageNet pre-training, models pre-trained on MicroNet generalized better to out-of-distribution micrographs taken under different imaging and sample conditions and were more accurate with less training data. When training with only a single Ni-superalloy image, pre-training on MicroNet produced a 72.2 percent reduction in relative segmentation error. These results suggest that transfer learning from large in-domain datasets generate models with learned feature representations that are more useful for downstream tasks and will likely improve any microscopy image analysis technique that can leverage pre-trained encoders.

machine learning↗