Amount Of Ground Truth Data Research Articles

A Machine Learning Framework for Assessing Experts’ Decision Quality

Expert workers make non-trivial decisions with significant implications. Experts’ decision accuracy is, thus, a fundamental aspect of their judgment quality, key to both management and consumers of experts’ services. Yet, in many important settings, transparency in experts’ decision quality is rarely possible because ground truth data for evaluating the experts’ decisions is costly and available only for a limited set of decisions. Furthermore, different experts typically handle exclusive sets of decisions, and thus, prior solutions that rely on the aggregation of multiple experts’ decisions for the same instance are inapplicable. We first formulate the problem of estimating experts’ decision accuracy in this setting and then develop a machine–learning–based framework to address it. Our method effectively leverages both abundant historical data on workers’ past decisions and scarce decision instances with ground truth labels. Using both semi-synthetic data based on publicly available data sets and purposefully compiled data sets on real workers’ decisions, we conduct extensive empirical evaluations of our method’s performance relative to alternatives. The results show that our approach is superior to existing alternatives across diverse settings, including settings that involve different data domains, experts’ qualities, and amounts of ground truth data. To our knowledge, this paper is the first to posit and address the problem of estimating experts’ decision accuracies from historical data with scarce ground truth, and it is the first to offer comprehensive results for this problem setting, establishing the performances that can be achieved across settings as well as the state-of-the-art performance on which future work can build. This paper was accepted by Anindya Ghose, information systems. Funding: T. Geva acknowledges research grants from the Jeremy Coller Foundation and from the Henry Crown Institute for Business Research. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2021.03357 .

Human selection bias drives the linear nature of the more ground truth effect in explainable deep learning optical coherence tomography image segmentation.

Supervised deep learning (DL) algorithms are highly dependent on training data for which human graders are assigned, for example, for optical coherence tomography (OCT) image annotation. Despite the tremendous success of DL, due to human judgment, these ground truth labels can be inaccurate and/or ambiguous and cause a human selection bias. We therefore investigated the impact of the size of the ground truth and variable numbers of graders on the predictive performance of the same DL architecture and repeated each experiment three times. The largest training dataset delivered a prediction performance close to that of human experts. All DL systems utilized were highly consistent. Nevertheless, the DL under-performers could not achieve any further autonomous improvement even after repeated training. Furthermore, a quantifiable linear relationship between ground truth ambiguity and the beneficial effect of having a larger amount of ground truth data was detected and marked as the more-ground-truth effect.

Pixel Diffuser: Practical Interactive Medical Image Segmentation without Ground Truth.

Medical image segmentation is essential for doctors to diagnose diseases and manage patient status. While deep learning has demonstrated potential in addressing segmentation challenges within the medical domain, obtaining a substantial amount of data with accurate ground truth for training high-performance segmentation models is both time-consuming and demands careful attention. While interactive segmentation methods can reduce the costs of acquiring segmentation labels for training supervised models, they often still necessitate considerable amounts of ground truth data. Moreover, achieving precise segmentation during the refinement phase results in increased interactions. In this work, we propose an interactive medical segmentation method called PixelDiffuser that requires no medical segmentation ground truth data and only a few clicks to obtain high-quality segmentation using a VGG19-based autoencoder. As the name suggests, PixelDiffuser starts with a small area upon the initial click and gradually detects the target segmentation region. Specifically, we segment the image by creating a distortion in the image and repeating it during the process of encoding and decoding the image through an autoencoder. Consequently, PixelDiffuser enables the user to click a part of the organ they wish to segment, allowing the segmented region to expand to nearby areas with pixel values similar to the chosen organ. To evaluate the performance of PixelDiffuser, we employed the dice score, based on the number of clicks, to compare the ground truth image with the inferred segment. For validation of our method's performance, we leveraged the BTCV dataset, containing CT images of various organs, and the CHAOS dataset, which encompasses both CT and MRI images of the liver, kidneys and spleen. Our proposed model is an efficient and effective tool for medical image segmentation, achieving competitive performance compared to previous work in less than five clicks and with very low memory consumption without additional training.

Physics-informed neural networks for modeling physiological time series for cuffless blood pressure estimation

The bold vision of AI-driven pervasive physiological monitoring, through the proliferation of off-the-shelf wearables that began a decade ago, has created immense opportunities to extract actionable information for precision medicine. These AI algorithms model input-output relationships of a system that, in many cases, exhibits complex nature and personalization requirements. A particular example is cuffless blood pressure estimation using wearable bioimpedance. However, these algorithms need training over significant amount of ground truth data. In the context of biomedical applications, collecting ground truth data, particularly at the personalized level is challenging, burdensome, and in some cases infeasible. Our objective is to establish physics-informed neural network (PINN) models for physiological time series data that would use minimal ground truth information to extract complex cardiovascular information. We achieve this by building Taylor’s approximation for gradually changing known cardiovascular relationships between input and output (e.g., sensor measurements to blood pressure) and incorporating this approximation into our proposed neural network training. The effectiveness of the framework is demonstrated through a case study: continuous cuffless BP estimation from time series bioimpedance data. We show that by using PINNs over the state-of-the-art time series models tested on the same datasets, we retain high correlations (systolic: 0.90, diastolic: 0.89) and low error (systolic: 1.3 ± 7.6 mmHg, diastolic: 0.6 ± 6.4 mmHg) while reducing the amount of ground truth training data on average by a factor of 15. This could be helpful in developing future AI algorithms to help interpret pervasive physiologic data using minimal amount of training data.

Background-foreground segmentation for interior sensing in automotive industry

To ensure safety in automated driving, the correct perception of the situation inside the car is as important as its environment. Thus, seat occupancy detection and classification of detected instances play an important role in interior sensing. By the knowledge of the seat occupancy status, it is possible to, e.g., automate the airbag deployment control. Furthermore, the presence of a driver, which is necessary for partially automated driving cars at the automation levels two to four can be verified. In this work, we compare different statistical methods from the field of image segmentation to approach the problem of background-foreground segmentation in camera based interior sensing. In the recent years, several methods based on different techniques have been developed and applied to images or videos from different applications. The peculiarity of the given scenarios of interior sensing is, that the foreground instances and the background both contain static as well as dynamic elements. In data considered in this work, even the camera position is not completely fixed. We review and benchmark three different methods ranging, i.e., Gaussian Mixture Models (GMM), Morphological Snakes and a deep neural network, namely a Mask R-CNN. In particular, the limitations of the classical methods, GMM and Morphological Snakes, for interior sensing are shown. Furthermore, it turns, that it is possible to overcome these limitations by deep learning, e.g. using a Mask R-CNN. Although only a small amount of ground truth data was available for training, we enabled the Mask R-CNN to produce high quality background-foreground masks via transfer learning. Moreover, we demonstrate that certain augmentation as well as pre- and post-processing methods further enhance the performance of the investigated methods.

Deep-layers-assisted machine learning for accurate image segmentation of complex materials

In materials science, X-ray micro-computed tomography (μCT) imagery provides a unique tool for quantitative and qualitative characterization of structure and performance. The tomographic analysis relies on the segmentation of large data sets to produce quantifiable volumes, a highly time-intensive process. Great progress has been observed in developing robust and efficient segmentation techniques, and machine learning currently stands out as a powerful segmentation tool. These methods have demonstrated excellent performance when applied to various materials. However, the performance of such methods heavily relies on the amount of ground truth data. This presents a great challenge in image segmentation since the amount of a priori raw images can be limited by sample cost and size. Additionally, manual labeling of data sets is often very time-intensive and can be subjected to human error. Therefore, a machine learning method is needed which achieves high segmentation accuracies for reduced training data sets. To this end, we have offered a stochastic method for producing variations of the training images that will retain the important class-wide features and thereby enrich the machine learning's “understanding” of the variabilities. The proposed method can significantly increase the final overall accuracy. We found that by enlarging the initial training set by additional realizations, we were able to improve the average accuracy of segmentation from 81.1% to 90.0% for a very complex Mg-based alloy. The results of this study show that it is possible to increase the accuracy of predictions in imaging from X-ray microscopy using machine learning methods when enough data are not available. • Rapid and smart segmentation of multiphase complex materials is studied. • We presented a machine learning method that is combined with a stochastic method. • Our method shows a significant improvement over the original machine learning methods. • The proposed method can be used when enough data is not available for machine learning.

U-Net-Based Segmentation of Microscopic Images of Colorants and Simplification of Labeling in the Learning Process.

Colored product textures correspond to particle size distributions. The microscopic images of colorants must be divided into regions to determine the particle size distribution. The conventional method used for this process involves manually dividing images into areas, which may be inefficient. In this paper, we have overcome this issue by developing two different modified architectures of U-Net convolution neural networks to automatically determine the particle sizes. To develop these modified architectures, a significant amount of ground truth data must be prepared to train the U-Net, which is difficult for big data as the labeling is performed manually. Therefore, we also aim to reduce this process by using incomplete labeling data. The first objective of this study is to determine the accuracy of our modified U-Net architectures for this type of image. The second objective is to reduce the difficulty of preparing the ground truth data by testing the accuracy of training on incomplete labeling data. The results indicate that efficient segmentation can be realized using our modified U-Net architectures, and the generation of ground truth data can be simplified. This paper presents a preliminary study to improve the efficiency of determining particle size distributions with incomplete labeling data.

Fast location and segmentation of high-throughput damaged soybean seeds with invertible neural networks.

Fast identification of damaged soybean seeds has undeniable importance in seed sorting and food quality. Mechanical vibration is generally used in soybean seed sorting, but this can seriously damage soybean seeds. The convolutional neural network (CNN) is considered an effective method for location and segmentation tasks. However, a CNN requires a large amount of ground truth data and has high computational cost. First, we propose a self-supervision manner to automatically generate ground truths, which can theoretically create an almost unlimited number of labeled images. Second, instead of using popular CNNs, a novel invertible convolution (involution)-enabled scheme is proposed by using the bottleneck block of the residual networks. Third, a feature selection feature pyramid network (FS-FPN) based on involution is designed, which selects features more flexibly and adaptively. We further merge involution-based backbones and FS-FPN into a unified network, achieving an end-to-end seed location and segmentation model; the best mean average precision of location and segmentation achieved was 85.1% and 81% respectively. The experimental results demonstrate that the proposed method greatly improves the performance of the baseline network with faster speed and fewer parameters, enabling it to detect soybean seeds more effectively. © 2022 Society of Chemical Industry.

Object-Based Visual Camera Pose Estimation From Ellipsoidal Model and 3D-Aware Ellipse Prediction

In this paper, we propose a method for initial camera pose estimation from just a single image which is robust to viewing conditions and does not require a detailed model of the scene. This method meets the growing need of easy deployment of robotics or augmented reality applications in any environments, especially those for which no accurate 3D model nor huge amount of ground truth data are available. It exploits the ability of deep learning techniques to reliably detect objects regardless of viewing conditions. Previous works have also shown that abstracting the geometry of a scene of objects by an ellipsoid cloud allows to compute the camera pose accurately enough for various application needs. Though promising, these approaches use the ellipses fitted to the detection bounding boxes as an approximation of the imaged objects. In this paper, we go one step further and propose a learning-based method which detects improved elliptic approximations of objects which are coherent with the 3D ellipsoids in terms of perspective projection. Experiments prove that the accuracy of the computed pose significantly increases thanks to our method. This is achieved with very little effort in terms of training data acquisition - a few hundred calibrated images of which only three need manual object annotation. Code and models are released at https://gitlab.inria.fr/tangram/3d-aware-ellipses-for-visual-localization

Geolocation Prediction in Twitter Using Social Networks: A Critical Analysis and Review of Current Practice

Geolocated social media data provides a powerful source of information about place and regional human behavior. Because little social media data is geolocation-annotated, inference techniques serve an essential role for increasing the volume of annotated data. One major class of inference approaches has relied on the social network of Twitter, where the locations of a user's friends serve as evidence for that user's location. While many such inference techniques have been recently proposed, we actually know little about their relative performance, with the amount of ground truth data varying between 5% and 100% of the network, the size of the social network varying by four orders of magnitude, and little standardization in evaluation metrics. We conduct a systematic comparative analysis of nine state-of-the-art network-based methods for performing geolocation inference at the global scale, controlling for the source of ground truth data, dataset size, and temporal recency in test data. Furthermore, we identify a comprehensive set of evaluation metrics that clarify performance differences. Our analysis identifies a large performance disparity between that reported in the literature and that seen in real-world conditions. To aid reproducibility and future comparison, all implementations have been released in an open source geoinference package.

High-throughput soybean seeds phenotyping with convolutional neural networks and transfer learning

BackgroundEffective soybean seed phenotyping demands large-scale accurate quantities of morphological parameters. The traditional manual acquisition of soybean seed morphological phenotype information is error-prone, and time-consuming, which is not feasible for large-scale collection. The segmentation of individual soybean seed is the prerequisite step for obtaining phenotypic traits such as seed length and seed width. Nevertheless, traditional image-based methods for obtaining high-throughput soybean seed phenotype are not robust and practical. Although deep learning-based algorithms can achieve accurate training and strong generalization capabilities, it requires a large amount of ground truth data which is often the limitation step.ResultsWe showed a novel synthetic image generation and augmentation method based on domain randomization. We synthesized a plenty of labeled image dataset automatedly by our method to train instance segmentation network for high throughput soybean seeds segmentation. It can pronouncedly decrease the cost of manual annotation and facilitate the preparation of training dataset. And the convolutional neural network can be purely trained by our synthetic image dataset to achieve a good performance. In the process of training Mask R-CNN, we proposed a transfer learning method which can reduce the computing costs significantly by finetuning the pre-trained model weights. We demonstrated the robustness and generalization ability of our method by analyzing the result of synthetic test datasets with different resolution and the real-world soybean seeds test dataset.ConclusionThe experimental results show that the proposed method realized the effective segmentation of individual soybean seed and the efficient calculation of the morphological parameters of each seed and it is practical to use this approach for high-throughput objects instance segmentation and high-throughput seeds phenotyping.

Learning to Calibrate Battery Models in Real-Time with Deep Reinforcement Learning

Lithium-ion (Li-I) batteries have recently become pervasive and are used in many physical assets. For the effective management of the batteries, reliable predictions of the end-of-discharge (EOD) and end-of-life (EOL) are essential. Many detailed electrochemical models have been developed for the batteries. Their parameters are calibrated before they are taken into operation and are typically not re-calibrated during operation. However, the degradation of batteries increases the reality gap between the computational models and the physical systems and leads to inaccurate predictions of EOD/EOL. The current calibration approaches are either computationally expensive (model-based calibration) or require large amounts of ground truth data for degradation parameters (supervised data-driven calibration). This is often infeasible for many practical applications. In this paper, we introduce a reinforcement learning-based framework for reliably inferring calibration parameters of battery models in real time. Most importantly, the proposed methodology does not need any labeled data samples of observations and the ground truth parameters. The experimental results demonstrate that our framework is capable of inferring the model parameters in real time with better accuracy compared to approaches based on unscented Kalman filters. Furthermore, our results show better generalizability than supervised learning approaches even though our methodology does not rely on ground truth information during training.

Estimation of Crop Yield From Combined Optical and SAR Imagery Using Gaussian Kernel Regression

The synthetic aperture radar (SAR) interferometric coherence can complement optical data for the estimation of crop growth parameters, but it has not been yet investigated for predicting crop yield. Many studies have used machine-learning methods, such as neural networks, random forest, and Gaussian process regression, to estimate crop yield from remotely sensed data. However, their performance depends on the amount of available ground truth data. This study proposed Gaussian kernel regression for rice yield estimation from optical and SAR imagery using a limited amount of ground truth data. The main objective was to investigate the synergetic use of Sentinel-2 vegetation indices and Sentinel-1 interferometric coherence data through Gaussian kernel regression for estimating rice grain yield. The prediction accuracy was assessed using in situ measured yield data collected in 2019 and 2020 over Xinghua county in Jiangsu Province, China. In all cases, Gaussian kernel regression outperformed the probabilistic Gaussian regression and Bayesian linear inference. With the independently used optical and SAR data, a better prediction accuracy was achieved with the optical red edge difference vegetation index (RDVI1) ( r2 = 0.65, RMSE = 0.61 t/ha) than with the interferometric coherence ( r2 = 0.52 and RMSE = 0.79 t/ha).The highest prediction accuracy can be achieved by combining RDVI1 with interferometric coherence at the heading stage ( r2 = 0.81 and RMSE = 0.55 t/ha). The results suggest the value of synergy between satellite interferometric coherence and optical indices for crop yield mapping with Gaussian kernel regression.

Mapping sub-field maize yields in Nebraska, USA by combining remote sensing imagery, crop simulation models, and machine learning

Crop yield maps are valuable for many applications in precision agriculture, but are often inaccessible to growers and researchers wishing to better understand yield determinants and improve site-specific management strategies. A method for mapping sub-field crop yields from remote sensing imagery could increase the availability of crop yield maps. A variation of the scalable crop yield mapping approach (SCYM, Lobell et al. in Remote Sensing of Environment 164:324–333, 2015) was developed and tested for estimating sub-field maize (Zea mays L.) yields at 10–30 m without the use of site-specific input data. The method was validated using harvester yield monitor records for 21 site-years for irrigated and rainfed fields in eastern Nebraska, USA. Prediction error ranged greatly across site-years, with relative RMSE scores of 10.8 to 38.5%, and R2 values of 0.003 to 0.37. Significant proportional bias was detected in the predictions, but could be corrected with a small amount of ground truth data. Crop yield prediction accuracies without calibration were suitable for some precision applications such as mapping relative yields and delineating management zones, but model improvements or calibration datasets are needed for applications requiring absolute yield estimates.

Identifying malicious social media contents using multi-view Context-Aware active learning

Abstract This paper presents a semi-supervised, multi-view, active learning method, which uses an optimized set of most informative samples and utilizes domain specific context information to efficiently and effectively identify malicious forum content in web-based social media platforms. As research shows, the task of automated identification of malicious forum posts, which also helps in detecting their associated key suspects in web forums, faces numerous challenges: (1) Online data, particularly social media data originate from diverse and heterogeneous sources and are largely unstructured; (2) Online data characteristics evolve quickly; and, (3) There are limited amounts of ground truth data to support the development of effective classification technologies in a strictly supervised scenario. In order to address the above challenges, the proposed human–machine collaborative, semi-supervised learning method is designed to efficiently and effectively identify harmful, provocative, or fabricated forum content by observing only a small number of annotated samples. Our learning framework is initiated by modeling initial view-dependent classifiers from a limited labeled data collection and allows each, in an interactive manner, to evolve dynamically into a sophisticated model by observing data patterns from a shared shortlist of most informative samples, identified via a graph-based optimization method and solved by a maximum flow algorithm. By designing a context rich metric definition in a data-driven manner, the proposed framework is able to learn a sufficiently robust classification model, that utilizes only a small number of human annotated samples, typically 1–2 orders of magnitude fewer as compared to a fully supervised solution. We validate our method using a large collection of flagged words with a wide range of origins, words frequently appearing in web-based forums and manually verified by multiple experienced, independent domain experts.

Remotely sensed vegetation index and LAI for parameter determination of the CSM-CROPGRO-Soybean model when in situ data are not available

An agricultural system is a complex combination of many different components that require different types of data for analysis and modeling. Remote sensing information is an alternative source of data for areas that only have a small amount of ground truth data. The goal of this study was to evaluate whether remotely sensed data can be used for calibration of genetic specific parameters (GSPs) with the ultimate goal of yield estimation. This study used the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) from Moderate Resolution Imaging Spectroradiometer (MODIS) with measured Leaf Area Index (LAI) for soybean fields in Paraná, Brazil and Iowa, USA, to calibrate the cultivar parameters of the CSM-CROPGRO-Soybean model. Three calibration methods were performed including field-measured LAI, remotely sensed derived LAI, and remotely sensed derived Light Interception. The cultivar parameters sensitive to LAI and LI were calibrated for yield with a mean error of -4.5 kg/ha (0.1%) and with a R2 of 0.89 for Parana. The availability of crop growth measurements for Iowa resulted in an average RMSE of 895 kg/ha (average nRMSE of 6%), and Willmott agreement index of 0.98 for time-series biomass, and an average RMSE of 941 kg/ha (average nRMSE of 15%) for pod weight. This study showed that remotely sensed LAI and LI from NDVI data can be used for calibration of GSPs with the ultimate goal of improving yield predictions based on local dynamic temporal and spatial variability.

SynSys: A Synthetic Data Generation System for Healthcare Applications.

Creation of realistic synthetic behavior-based sensor data is an important aspect of testing machine learning techniques for healthcare applications. Many of the existing approaches for generating synthetic data are often limited in terms of complexity and realism. We introduce SynSys, a machine learning-based synthetic data generation method, to improve upon these limitations. We use this method to generate synthetic time series data that is composed of nested sequences using hidden Markov models and regression models which are initially trained on real datasets. We test our synthetic data generation technique on a real annotated smart home dataset. We use time series distance measures as a baseline to determine how realistic the generated data is compared to real data and demonstrate that SynSys produces more realistic data in terms of distance compared to random data generation, data from another home, and data from another time period. Finally, we apply our synthetic data generation technique to the problem of generating data when only a small amount of ground truth data is available. Using semi-supervised learning we demonstrate that SynSys is able to improve activity recognition accuracy compared to using the small amount of real data alone.

Development of a Classification Method for Forest Vegetation on the Stand Level, Using KOMPSAT-3A Imagery and Land Coverage Map

Due to the advance in remote sensing technology, it has become easier to more frequently obtain high resolution imagery to detect delicate changes in an extensive area, particularly including forest which is not readily sub-classified. Time-series analysis on high resolution images requires to collect extensive amount of ground truth data. In this study, the potential of land coverage mapas ground truth data was tested in classifying high-resolution imagery. The study site was Wonju-si at Gangwon-do, South Korea, having a mix of urban and natural areas. KOMPSAT-3A imagery taken on March 2015 and land coverage map published in 2017 were used as source data. Two pixel-based classification algorithms, Support Vector Machine (SVM) and Random Forest (RF), were selected for the analysis. Forest only classification was compared with that of the whole study area except wetland. Confusion matrixes from the classification presented that overall accuracies for both the targets were higher in RF algorithm than in SVM. While the overall accuracy in the forest only analysis by RF algorithm was higher by 18.3% than SVM, in the case of the whole region analysis, the difference was relatively smaller by 5.5%. For the SVM algorithm, adding the Majority analysis process indicated a marginal improvement of about 1% than the normal SVM analysis. It was found that the RF algorithm was more effective to identify the broad-leaved forest within the forest, but for the other classes the SVM algorithm was more effective. As the two pixel-based classification algorithms were tested here, it is expected that future classification will improve the overall accuracy and the reliability by introducing a time-series analysis and an object-based algorithm. It is considered that this approach will contribute to improving a large-scale land planning by providing an effective land classification method on higher spatial and temporal scales.

The Hyper-Angular Cube Concept for Improving the Spatial and Acoustic Resolution of MBES Backscatter Angular Response Analysis

This study presents a novel approach, based on high-dimensionality hydro-acoustic data, for improving the performance of angular response analysis (ARA) on multibeam backscatter data in terms of acoustic class separation and spatial resolution. This approach is based on the hyper-angular cube (HAC) data structure which offers the possibility to extract one angular response from each cell of the cube. The HAC consists of a finite number of backscatter layers, each representing backscatter values corresponding to single-incidence angle ensonifications. The construction of the HAC layers can be achieved either by interpolating dense soundings from highly overlapping multibeam echo-sounder (MBES) surveys (interpolated HAC, iHAC) or by producing several backscatter mosaics, each being normalized at a different incidence angle (synthetic HAC, sHAC). The latter approach can be applied to multibeam data with standard overlap, thus minimizing the cost for data acquisition. The sHAC is as efficient as the iHAC produced by actual soundings, providing distinct angular responses for each seafloor type. The HAC data structure increases acoustic class separability between different acoustic features. Moreover, the results of angular response analysis are applied on a fine spatial scale (cell dimensions) offering more detailed acoustic maps of the seafloor. Considering that angular information is expressed through high-dimensional backscatter layers, we further applied three machine learning algorithms (random forest, support vector machine, and artificial neural network) and one pattern recognition method (sum of absolute differences) for supervised classification of the HAC, using a limited amount of ground truth data (one sample per seafloor type). Results from supervised classification were compared with results from an unsupervised method for inter-comparison of the supervised algorithms. It was found that all algorithms (regarding both the iHAC and the sHAC) produced very similar results with good agreement (>0.5 kappa) with the unsupervised classification. Only the artificial neural network required the total amount of ground truth data for producing comparable results with the remaining algorithms.

Citywide Traffic Volume Estimation Using Trajectory Data

Traffic volume estimation at the city scale is an important problem useful to many transportation operations and urban applications. This paper proposes a hybrid framework that integrates both state-of-art machine learning techniques and well-established traffic flow theory to estimate citywide traffic volume. In addition to typical urban context features extracted from multiple sources, we extract a special set of features from GPS trajectories based on the implications of traffic flow theory, which provide extra information on the speed-flow relationship. Using the network-wide speed information estimated from a travel speed estimation model, a volume related high level feature is first learned using an unsupervised graphical model. A volume re-interpretation model is then introduced to map the volume related high level feature to the predicted volume using a small amount of ground truth data for training. The framework is evaluated using a GPS trajectory dataset from 33,000 Beijing taxis and volume ground truth data obtained from 4,980 video clips. The results demonstrate effectiveness and potential of the proposed framework in citywide traffic volume estimation.