Ground Truth Dataset Research Articles

Intelligent advice system for human drivers to prevent overtaking accidents in roads

Many road accidents happen due to the misjudgments and miscalculations of human drivers when overtaking their lead vehicles. Using unaided sight, the human drivers cannot calculate the accurate velocity of lead vehicles that traveling in similar and opposite directions with respect to their vehicles and can work only by approximation. This incapability of human drivers ends in miscalculation of Time-to-Collision (TTC) and causes accidents between ego and lead vehicles, which kill people in road accidents. A novel Intelligent Overtaking Advice System (IOAS) is proposed in this paper to provide advice to human drivers during attempts to overtake. IOAS is developed to predict both accurate velocity of lead vehicles and accurate TTC. The IOAS is powered by a Velocity Network (VNet) and a Time-to-Collision-Network (TTC-Net). The proposed VNet and TTC-Net are well trained with a huge volume of ground truth datasets to provide better accuracy, robustness, and quick response time. The performance of the proposed IOAS is analyzed and it is observed that the proposed VNet and TTC-Net provide 97% and 98% accuracy, respectively. The proposed IOAS can be integrated into real vehicles as an add-on to the Advanced Driver Assistance System (ADAS) to prevent accidents during the overtaking process.

PixelPrint: Three-dimensional printing of realistic patient-specific lung phantoms for CT imaging.

Phantoms are essential tools for assessing and verifying performance in computed tomography (CT). Realistic patient-based lung phantoms that accurately represent textures and densities are essential in developing and evaluating novel CT hardware and software. This study introduces PixelPrint, a 3D-printing solution to create patient-specific lung phantoms with accurate contrast and textures. PixelPrint converts patient images directly into printer instructions, where density is modeled as the ratio of filament to voxel volume to emulate local attenuation values. For evaluation of PixelPrint, phantoms based on four COVID-19 pneumonia patients were manufactured and scanned with the original (clinical) CT scanners and protocols. Density and geometrical accuracies between phantom and patient images were evaluated for various anatomical features in the lung, and a radiomic feature comparison was performed for mild, moderate, and severe COVID-19 pneumonia patient-based phantoms. Qualitatively, CT images of the patient-based phantoms closely resemble the original CT images, both in texture and contrast levels, with clearly visible vascular and parenchymal structures. Regions-of-interest (ROIs) comparing attenuation demonstrated differences below 15 HU. Manual size measurements performed by an experienced thoracic radiologist revealed a high degree of geometrical correlation between identical patient and phantom features, with differences smaller than the intrinsic spatial resolution of the images. Radiomic feature analysis revealed high correspondence, with correlations of 0.95-0.99 between patient and phantom images. Our study demonstrates the feasibility of 3D-printed patient-based lung phantoms with accurate geometry, texture, and contrast that will enable protocol optimization, CT research and development advancements, and generation of ground-truth datasets for radiomic evaluations.

A Travel Mode Identification Framework Based on Cellular Signaling Data

The rapid development of telecommunication network has produced a large amount of spatial-temporal information of mobile phone users. GPS data are typically collected by smartphones apps, which are restricted to small samples of the population. Cellular signaling data (CSD) are usually collected by mobile network operators, which enables researchers to conduct travel behavior analysis of the entire population at a relatively low cost compared to GPS. However, extracting travel mode information from CSD is particularly challenging due to the noise data and low positioning accuracy. This paper proposes a travel mode identification framework based on CSD, which includes data cleansing and travel mode identification. In terms of data cleansing, oscillation sequence and drift data are mainly cleansed. For the oscillation sequence, this paper proposes a detection algorithm based on time window. For the drift data, this paper proposes a detection algorithm based on distance, velocity and frequency. In terms of travel mode identification, the task is divided into two dichotomous problems: motor and non-motor transport identification and public and private transport identification. Each dichotomous problem proposes an algorithm that does not rely on the ground truth dataset for model training. Finally, a ground truth dataset is constructed to verify all algorithms. The result shows that, in terms of data cleansing, the similarity between CSD after cleansing and the actual trajectory according to DTW improved by 101.13% on average. In terms of travel mode identification, the proposed method can achieve similar or even better accuracy than traditional supervised-learning algorithms (94% in motor and non-motor transport identification, 83.5% in public and private transport identification), which can be directly applied to large-scale population analysis scenarios.

Domain-specific trained model to auto grade AV Nicking severity level

ArterioVenous Nicking (AVN) is a prominent vascular biomarker in predicting cardiovascular risk factors such as atherosclerosis, hypertension, coronary artery diseases, and heart stroke. The morphological change in the retinal junction calibre presents various clues to diagnose the aforesaid risk factors. In this paper, we introduced a novel hierarchical training strategy to improve the grading accuracy in assessing the abnormality. In the first level, a domain-specific weights are generated from the pretrained weights. At the second level, domain-specific weights are transferred to the target architecture called Scale Attention Region-Based Deep Architecture, which is again trained on the AV Nicked dataset. In the target architecture, the scale attention feature extractor generates adaptive weights based on the semantic information at different levels of resolution. The advantage of this module is that it extracts minute features of severity so that the grading can be done precisely. The hierarchically trained classifier’s performance for AVN grading is compared with manual grading and also with a generalised pretrained(PT) model. The ResNet101 + DST metrics are evaluated and compared with a ground truth dataset.

IDC: quantitative evaluation benchmark of interpretation methods for deep text classification models

Recent advances in deep neural networks have achieved outstanding success in natural language processing tasks. Interpretation methods that provide insight into the decision-making process of these models have received an influx of research attention because of the success and the black-box nature of the deep text classification models. Evaluation of these methods has been based on changes in classification accuracy or prediction confidence when removing important words identified by these methods. There are no measurements of the actual difference between the predicted important words and humans’ interpretation of ground truth because of the lack of interpretation ground truth. A large publicly available interpretation ground truth has the potential to advance the development of interpretation methods. Manual labeling important words for each document to create a large interpretation ground truth is very time-consuming. This paper presents (1) IDC, a new benchmark for quantitative evaluation of interpretation methods for deep text classification models, and (2) evaluation of six interpretation methods using the benchmark. The IDC benchmark consists of: (1) Three methods that generate three pseudo-interpretation ground truth datasets. (2) Three performance metrics: interpretation recall, interpretation precision, and Cohen’s kappa inter-agreement. Findings: IDC-generated interpretation ground truth agrees with human annotators on sampled movie reviews. IDC identifies Layer-wise Relevance Propagation and the gradient-by-input methods as the winning interpretation methods in this study.

VIQoR: a web service for visually supervised protein inference and protein quantification.

In quantitative bottom-up mass spectrometry (MS)-based proteomics, the reliable estimation of protein concentration changes from peptide quantifications between different biological samples is essential. This estimation is not a single task but comprises the two processes of protein inference and protein abundance summarization. Furthermore, due to the high complexity of proteomics data and associated uncertainty about the performance of these processes, there is a demand for comprehensive visualization methods able to integrate protein with peptide quantitative data including their post-translational modifications. Hence, there is a lack of a suitable tool that provides post-identification quantitative analysis of proteins with simultaneous interactive visualization. In this article, we present VIQoR, a user-friendly web service that accepts peptide quantitative data of both labeled and label-free experiments and accomplishes the crucial components protein inference and summarization and interactive visualization modules, including the novel VIQoR plot. We implemented two different parsimonious algorithms to solve the protein inference problem, while protein summarization is facilitated by a well-established factor analysis algorithm called fast-FARMS followed by a weighted average summarization function that minimizes the effect of missing values. In addition, summarization is optimized by the so-called Global Correlation Indicator (GCI). We test the tool on three publicly available ground truth datasets and demonstrate the ability of the protein inference algorithms to handle shared peptides. We furthermore show that GCI increases the accuracy of the quantitative analysis in datasets with replicated design. VIQoR is accessible at: http://computproteomics.bmb.sdu.dk/Apps/VIQoR/. The source code is available at: https://bitbucket.org/veitveit/viqor/. Supplementary data are available at Bioinformatics online.

Open-source analysis and visualization of segmented vasculature datasets with VesselVio

SummaryVascular networks are fundamental components of biological systems. Quantitative analysis and observation of the features of these networks can improve our understanding of their roles in health and disease. Recent advancements in imaging technologies have enabled the generation of large-scale vasculature datasets, but barriers to analyzing these datasets remain. Modern analysis options are mainly limited to paid applications or open-source terminal-based software that requires programming knowledge with high learning curves. Here, we describe VesselVio, an open-source application developed to analyze and visualize pre-binarized vasculature datasets and pre-constructed vascular graphs. Vasculature datasets and graphs can be loaded with annotations and processed with custom parameters. Here, the program is tested on ground-truth datasets and is compared with current pipelines. The utility of VesselVio is demonstrated by the analysis of multiple formats of 2D and 3D datasets acquired with several imaging modalities, including annotated mouse whole-brain vasculature volumes.

Motion correction for three-dimensional chemical exchange saturation transfer imaging without direct water saturation artifacts.

In chemical exchange saturation transfer (CEST) MRI, motion correction is compromised by the drastically changing image contrast at different frequency offsets, particularly at the direct water saturation. In this study, a simple extension for conventional image registration algorithms is proposed, enabling robust and accurate motion correction of CEST-MRI data. The proposed method uses weighted averaging of motion parameters from a conventional rigid image registration to identify and mitigate erroneously misaligned images. Functionality of the proposed method was verified by ground truth datasets generated from 10 three-dimensional in vivo measurements at 3T with simulated realistic random rigid motion patterns and noise. Performance was assessed using two different criteria: the maximum image misalignment as a measure for the robustness against direct water saturation artifacts, and the spectral error as a measure of the overall accuracy. For both criteria, the proposed method achieved the best scores compared with two motion-correction algorithms specifically developed to handle the varying contrasts in CEST-MRI. Compared with a straightforward linear interpolation of the motion parameters at frequency offsets close to the direct water saturation, the proposed method offers better performance in the absence of artifacts. The proposed method for motion correction in CEST-MRI allows identification and mitigation of direct water saturation artifacts that occur with conventional image registration algorithms. The resulting improved robustness and accuracy enable reliable motion correction, which is particularly crucial for an automated and carefree evaluation of spectral CEST-MRI data, e.g., for large patient cohorts or in clinical routines.

Database search engines and target database features impinge upon the identification of post-translationally cis-spliced peptides in HLA class I immunopeptidomes.

Unconventional epitopes presented by HLA class I complexes are emerging targets for T cell targeted immunotherapies. Their identification by mass spectrometry (MS) required development of novel methods to cope with the large number of theoretical candidates. Methods to identify post‐translationally spliced peptides led to a broad range of outcomes. We here investigated the impact of three common database search engines – that is, Mascot, Mascot+Percolator, and PEAKS DB – as final identification step, as well as the features of target database on the ability to correctly identify non‐spliced and cis‐spliced peptides. We used ground truth datasets measured by MS to benchmark methods’ performance and extended the analysis to HLA class I immunopeptidomes. PEAKS DB showed better precision and recall of cis‐spliced peptides and larger number of identified peptides in HLA class I immunopeptidomes than the other search engine strategies. The better performance of PEAKS DB appears to result from better discrimination between target and decoy hits and hence a more robust FDR estimation, and seems independent to peptide and spectrum features here investigated.

A scalable artificial intelligence platform that automatically finds copy number variations (CNVs) in journal articles and transforms them into a database: CNV extraction, transformation, and loading AI (CNV-ETLAI)

BackgroundAlthough copy number variations (CNVs) are infrequent, each anomaly is unique, and multiple CNVs can appear simultaneously. Growing evidence suggests that CNVs contribute to a wide range of diseases. When CNVs are detected, assessment of their clinical significance requires a thorough literature review. This process can be extremely time-consuming and may delay disease diagnosis. Therefore, we have developed CNV Extraction, Transformation, and Loading Artificial Intelligence (CNV-ETLAI), an innovative tool that allows experts to classify and interpret CNVs accurately and efficiently. MethodsWe combined text, table, and image processing algorithms to develop an artificial intelligence platform that automatically extracts, transforms, and organizes CNV information into a database. To validate CNV-ETLAI, we compared its performance to ground truth datasets labeled by a human expert. In addition, we analyzed the CNV data, which was collected using CNV-ETLAI via a crowdsourcing approach. ResultsIn comparison to a human expert, CNV-ETLAI improved CNV detection accuracy by 4% and performed the analysis 60 times faster. This performance can improve even further with upscaling of the CNV-ETLAI database as usage increases. 5,800 CNVs from 2,313 journal articles were collected. Total CNV frequency for the whole chromosome was highest for chromosome X, whereas CNV frequency per 1 Mb of genomic length was highest for chromosome 22. ConclusionsWe have developed, tested, and shared CNV-ETLAI for research and clinical purposes (https://lmic.mgh.harvard.edu/CNV-ETLAI). Use of CNV-ETLAI is expected to ease and accelerate diagnostic classification and interpretation of CNVs.

User Identity Linkage via Co-Attentive Neural Network From Heterogeneous Mobility Data

Online services are playing critical roles in almost all aspects of users’ life. Users usually have multiple online identities (IDs) in different online services. In order to fuse the separated user data in multiple services for better business intelligence, it is critical for service providers to link online IDs belonging to the same user. On the other hand, the popularity of mobile networks and GPS-equipped smart devices have provided a generic way to link IDs, i.e., utilizing the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mobility traces</i> of IDs. However, linking IDs based on their mobility traces has been a challenging problem due to the highly heterogeneous, incomplete and noisy mobility data across services. In this paper, we propose <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DPLink</i> , an end-to-end deep learning based framework, to complete the user identity linkage task for heterogeneous mobility data collected from different services with different properties. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DPLink</i> is made up by a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">feature extractor</i> including a location encoder and a trajectory encoder to extract representative features from trajectory and a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">comparator</i> to compare and decide whether to link two trajectories as the same user. Particularly, we propose a pre-training strategy with a simple task to train the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DPLink</i> model to overcome the training difficulties introduced by the highly heterogeneous nature of different source mobility data. Besides, we introduce a multi-modal embedding network and a co-attention mechanism in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DPLink</i> to deal with the low-quality problem of mobility data. By conducting extensive experiments on two real-life ground-truth mobility datasets with eight baselines, we demonstrate that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DPLink</i> outperforms the state-of-the-art solutions by more than 15 percent in terms of hit-precision. Moreover, it is expandable to add external geographical context data and works stably with heterogeneous noisy mobility traces.

Traffic speed forecasting for urban roads: A deep ensemble neural network model

Real-time and accurate traffic state forecasting of urban roads is of great significance to improve traffic efficiency and optimize travel routes. However, future traffic state forecasting is still a challenging issue as it is influenced by several complicated factors including the dynamic spatio-temporal dependencies. Existing models usually consider the dependencies from the road sections with physical connections and ignore the road sections without physical connections. To this end, this paper proposes a deep ensemble neural network (DENN) model to improve the accuracy of urban traffic state forecasting by forming the road sections with high relevance into a virtual graph. To capture the spatio-temporal characteristics efficiently and simultaneously, the DENN integrates the graph convolutional neural network, bidirectional gated recurrent unit network, and a dense layer with attention mechanism into an end-to-end fashion. Validated on two ground-truth urban traffic speed datasets, the DENN model can well fit the nonlinear fluctuation of urban speed and indicate superior performance than the state-of-the-art benchmark methods in terms of prediction precision and robustness.

Tree stem extraction from TLS point-cloud data of natural forests based on geometric features and DBSCAN

Conventional methods for tree stem extraction from point-cloud data of natural forests suffer from problems of low accuracy and poor universality. In this article, an enhanced method is proposed for tree stem extraction from point-cloud data collected by terrestrial laser scanning (TLS). First, principal component analysis is used to calculate the eigenvalues and eigenvectors of the point-cloud data, and an information entropy criterion is minimized in order to achieve the best neighbourhood scale selection. Then, three-dimensional spatial geometric forest features are combined with the Z-axis component of the normal vector. These geometric features are used for rough extraction of tree stem points, while a large number of non-stem points is filtered out by thresholding. Finally, the DBSCAN algorithm is used to achieve accurate extraction of tree stem points. The proposed method for tree stem detection and extraction is experimentally evaluated in the case of two representative natural forest plots of Pinus densata Mast. and Picea asperata Mast. in the Shangri-La City in China. All stem points in these two plots were detected and extracted with a reference method to create a ground-truth dataset. Correlation analysis was carried out between the stem points extracted by the proposed and reference methods for the two plots. This analysis resulted in an R2 value of 0.990 for the Pinus densata Mast. sample plot, and an R2 value of 0.982 for the Picea asperata Mast. sample plot which has a more complex growth environment.

Deep-Learning-Based Cerebral Artery Semantic Segmentation in Neurosurgical Operating Microscope Vision Using Indocyanine Green Fluorescence Videoangiography

There have been few anatomical structure segmentation studies using deep learning. Numbers of training and ground truth images applied were small and the accuracies of which were low or inconsistent. For a surgical video anatomy analysis, various obstacles, including a variable fast-changing view, large deformations, occlusions, low illumination, and inadequate focus occur. In addition, it is difficult and costly to obtain a large and accurate dataset on operational video anatomical structures, including arteries. In this study, we investigated cerebral artery segmentation using an automatic ground-truth generation method. Indocyanine green (ICG) fluorescence intraoperative cerebral videoangiography was used to create a ground-truth dataset mainly for cerebral arteries and partly for cerebral blood vessels, including veins. Four different neural network models were trained using the dataset and compared. Before augmentation, 35,975 training images and 11,266 validation images were used. After augmentation, 260,499 training and 90,129 validation images were used. A Dice score of 79% for cerebral artery segmentation was achieved using the DeepLabv3+ model trained using an automatically generated dataset. Strict validation in different patient groups was conducted. Arteries were also discerned from the veins using the ICG videoangiography phase. We achieved fair accuracy, which demonstrated the appropriateness of the methodology. This study proved the feasibility of operating field view of the cerebral artery segmentation using deep learning, and the effectiveness of the automatic blood vessel ground truth generation method using ICG fluorescence videoangiography. Using this method, computer vision can discern blood vessels and arteries from veins in a neurosurgical microscope field of view. Thus, this technique is essential for neurosurgical field vessel anatomy-based navigation. In addition, surgical assistance, safety, and autonomous surgery neurorobotics that can detect or manipulate cerebral vessels would require computer vision to identify blood vessels and arteries.

CellSium: versatile cell simulator for microcolony ground truth generation.

To train deep learning-based segmentation models, large ground truth datasets are needed. To address this need in microfluidic live-cell imaging, we present CellSium, a flexibly configurable cell simulator built to synthesize realistic image sequences of bacterial microcolonies growing in monolayers. We illustrate that the simulated images are suitable for training neural networks. Synthetic time-lapse videos with and without fluorescence, using programmable cell growth models, and simulation-ready 3D colony geometries for computational fluid dynamics are also supported. CellSium is free and open source software under the BSD license, implemented in Python, available at github.com/modsim/cellsium (DOI: 10.5281/zenodo.6193033), along with documentation, usage examples and Docker images. Supplementary data are available at Bioinformatics Advances online.

Behavioral Intrusion Prediction Model on Bayesian Network over Healthcare Infrastructure

Due to polymorphic nature of malware attack, a signature-based analysis is no longer sufficient to solve polymorphic and stealth nature of malware attacks. On the other hand, state-of-the-art methods like deep learning require labelled dataset as a target to train a supervised model. This is unlikely to be the case in production network as the dataset is unstructured and has no label. Hence an unsupervised learning is recommended. Behavioral study is one of the techniques to elicit traffic pattern. However, studies have shown that existing behavioral intrusion detection model had a few issues which had been parameterized into its common characteristics, namely lack of prior information (p (θ)), and reduced parameters (θ). Therefore, this study aims to utilize the previously built Feature Selection Model subsequently to design a Predictive Analytics Model based on Bayesian Network used to improve the analysis prediction. Feature Selection Model is used to learn significant label as a target and Bayesian Network is a sophisticated probabilistic approach to predict intrusion. Finally, the results are extended to evaluate detection, accuracy and false alarm rate of the model against the subject matter expert model, Support Vector Machine (SVM), k nearest neighbor (k-NN) using simulated and ground-truth dataset. The ground-truth dataset from the production traffic of one of the largest healthcare provider in Malaysia is used to promote realism on the real use case scenario. Results have shown that the proposed model consistently outperformed other models.

A Deep Learning Approach to Mapping Irrigation Using Landsat: IrrMapper U-Net

Accurate maps of irrigation are essential for understanding and managing water resources. We present a new method of mapping irrigation based on an ensemble of convolutional neural networks that use reflectance information from Landsat imagery to classify irrigated pixels. The methodology does not rely on extensive feature engineering and does not condition the classification with land use information from existing geospatial datasets. The ensemble does not need exhaustive hyperparameter tuning and the analysis pipeline is lightweight enough to be implemented on a personal computer. Furthermore, the proposed methodology provides an estimate of the uncertainty associated with classification. We evaluated our methodology and the resulting irrigation maps using a highly accurate novel spatially-explicit ground truth data set, using county-scale USDA surveys of irrigation extent, and using cadastral surveys. We demonstrate the accuracy of the method by mapping irrigation over the state of Montana from years 2000-2019. We found that our method outperforms other methods that use satellite remote sensing information in terms of overall accuracy and precision. We found that our method agrees better statewide with the USDA National Agricultural Statistics Survey estimates of irrigated area compared to other methods, and has far fewer errors of commission in rainfed agriculture areas. This methodology has the potential to be applied across the entire United States and for the complete Landsat record.

Automatic Annotation Performance of TextBlob and VADER on Covid Vaccination Dataset

With the recent boom in the corpus size of sentiment analysis tasks, automatic annotation is poised to be a necessary alternative to manual annotation for generating ground truth dataset labels. This article aims to investigate and validate the performance of two widely used lexicon-based automatic annotation approaches, TextBlob and Valence Aware Dictionary and Sentiment Reasoner (VADER), by comparing them with manual annotation. The dataset of 5402 Arabic tweets was annotated manually, containing 3124 positive tweets, 1463 negative tweets, and 815 neutral tweets. The tweets were translated into English so that TextBlob and VADER could be used for their annotation. TextBlob and VADER automatically classified the tweets to positive, negative, and neutral sentiments and compared them with manual annotation. This study shows that automatic annotation cannot be trusted as the gold standard for annotation. In addition, the study discussed many drawbacks and limitations of automatic annotation using lexicon-based algorithms. The highest level of accuracies of 75% and 70% were achieved by TextBlob and VADER, respectively.

A 3-D Storm Motion Estimation Method Based on Point Cloud Learning and Doppler Weather Radar Data

In recent years, deep learning techniques have been developed in the field of storm nowcasting, and they primarily focus on 2D radar product processing. Deep learning’s application to obtaining motion fields is limited by the lack of a motion field ground truth dataset for training. In this paper, we propose a method for storm motion estimation based on a point cloud deep learning network, which can perform 3D motion field estimation on two consecutive frames of weather radar volumetric data. To address the lack of ground truth data, we propose a synthetic data generation method based on Perlin noise and also generate a dataset to train the network. Our model’s architecture is based on FlowNet 3D. In this sense, we import an attention architecture to improve its ability to embed motion features. We also redesign the loss function to make it suitable for our task. In the experiment section, we first evaluate the performance of the trained model on our synthetic dataset. The experimental results demonstrate the following: first, that the trained network has the ability to estimate 3D motion from radar point cloud data; second, that it is a robust solution for radar data with different resolutions and observation ranges; and finally, that for the examined task, our proposed architecture has a 5-15% lower end point error than the original. Then, we directly apply the model trained on synthetic data to real cases, demonstrating the feasibility of our trained model on real data. Although motion field is a low-level product when considering a storm’s evolutionary process, it has the potential to serve a higher-level purpose, such as lightning or hail nowcasting. The point cloud approach provides a new perspective on weather radar data processing.

Trade-off between deep learning for species identification and inference about predator-prey co-occurrence: Reproducible R workflow integrating models in computer vision and ecological statistics

Deep learning is used in computer vision problems with important applications in several scientific fields. In ecology for example, there is a growing interest in deep learning for automatizing repetitive analyses on large amounts of images, such as animal species identification. However, there are challenging issues toward the wide adoption of deep learning by the community of ecologists. First, there is a programming barrier as most algorithms are written in Python while most ecologists are versed in R. Second, recent applications of deep learning in ecology have focused on computational aspects and simple tasks without addressing the underlying ecological questions or carrying out the statistical data analysis to answer these questions. Here, we showcase a reproducible R workflow integrating both deep learning and statistical models using predator-prey relationships as a case study. We illustrate deep learning for the identification of animal species on images collected with camera traps, and quantify spatial co-occurrence using multispecies occupancy models. Despite average model classification performances, ecological inference was similar whether we analysed the ground truth dataset or the classified dataset. This result calls for further work on the trade-offs between time and resources allocated to train models with deep learning and our ability to properly address key ecological questions with biodiversity monitoring. We hope that our reproducible workflow will be useful to ecologists and applied statisticians. All material (source of the Rmarkdown notebook and auxiliary files) is available from this https URL.