No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Complex data labeling with deep learning methods: Lessons from fisheries acoustics
Abstract
Quantitative and qualitative analysis of acoustic backscattered signals from the seabed bottom to the sea surface is used worldwide for fish stocks assessment and marine ecosystem monitoring. Huge amounts of raw data are collected yet require tedious expert labeling. This paper focuses on a case study where the ground truth labels are non-obvious: echograms labeling, which is time-consuming and critical for the quality of fisheries and ecological analysis. We investigate how these tasks can benefit from supervised learning algorithms and demonstrate that convolutional neural networks trained with non-stationary datasets can be used to stress parts of a new dataset needing human expert correction. Further development of this approach paves the way toward a standardization of the labeling process in fisheries acoustics and is a good case study for non-obvious data labeling processes.
... The challenges encountered in the selected studies can be classified into the following categories: data quality, data variability, data annotation, unbalanced datasets, and interpretation of models (Table 4). Unfortunately, the standard workflow of fishery acoustic data in species identification generally follows a pipeline consisting of data preprocessing, labeling, and annotation operations (Sarr et al., 2020). This process is challenging and requires a very large amount of work and time to be conducted. ...
... In addition, it requires a high level of knowledge and experience regarding the fish species behavior. (Rezvanifar et al., 2019;Marques et al., 2020Marques et al., , 2021Sarr et al., 2020) Unbalanced datasets ...
... These datasets are characterized by a most frequent background class, which limits and decreases the performance of models and their consistencies. (Villar et al., 2021;Choi et al., 2021;Rezvanifar et al., 2019;Marques et al., 2020Marques et al., , 2021Sarr et al., 2020;Aronica et al., 2019;Ordoñez et al., 2022) Models' interpretation ...
In fishery acoustics, surveys using sensor systems such as sonars and echosounders have been widely considered to be accurate tools for acquiring fish species data, fish species biomass, and abundance estimations. During acoustic surveys, research vessels are equipped with echosounders that produce sound waves and then record all echoes coming from objects and targets in the water column. The preprocessing and scrutinizing of acoustic fish species data have always been manually conducted and have been considered time-consuming. Meanwhile, deep learning and machine learning-based approaches have also been adopted to automate or partially automate the acoustic echo scrutinizing process and build an objective process with which the species echo classification uncertainty is expected to be lower than the uncertainty of scrutinizing experts. A review of the state-of-the-art of different deep learning and machine learning applications in acoustic fish species echo classification has been highly requested. Therefore, the present paper is conceived to identify and scan the studies conducted on acoustic fish echo identification using deep learning and machine learning approaches. This document can be extended to include other marine organisms rather than just fish species. To search for related papers, we used a systematic approach to search the most known electronic databases over the last five years. We were able to identify 13 related works, which have been processed to give a summary of multiple deep and machine learning approaches used in acoustic fish species identification, and then compare their architectures, performances, and the challenges encountered in their applications.
... A brief review of the literature reveals that in recent years, the development of machine learning methods to classify active acoustic data has converged toward random forests and deep neural networks (Fallon et al., 2016;Brautaset et al., 2020;Proud et al., 2020;Sarr et al., 2020;Marques et al., 2021;Blanluet et al., 2022). In their study, Sarr et al. (2020) obtained the best and most stable performance with a random forest model when the training dataset was small, while a deep neural network gave the best performance with a large training dataset. ...
... A brief review of the literature reveals that in recent years, the development of machine learning methods to classify active acoustic data has converged toward random forests and deep neural networks (Fallon et al., 2016;Brautaset et al., 2020;Proud et al., 2020;Sarr et al., 2020;Marques et al., 2021;Blanluet et al., 2022). In their study, Sarr et al. (2020) obtained the best and most stable performance with a random forest model when the training dataset was small, while a deep neural network gave the best performance with a large training dataset. Woodd-Walker et al. (2003) also evaluated a simple artificial neural network technique with a small training data set, but reported poor results for the least dominant class of their imbalanced dataset. ...
Acoustic surveys are the standard approach for evaluating many fish stocks around the world. The analysis of such survey data requires the accurate echo-classification of target species. This classification is often challenging as many organisms exhibit overlapping characteristics in terms of shape, acoustic amplitude, and behavior. In this study, a random forest approach was used to distinguish juvenile Pacific salmon (Oncorhynchus spp) from Pacific herring (Clupea pallasii) aggregations using the acoustic and morphological characteristics of their echo traces. The acoustic data was collected with an autonomous, multi-frequency echosounder deployed on the seafloor in the Discovery Islands, British Columbia from May to September 2015. The model was able to differentiate juvenile Pacific salmon from Pacific herring with a 98% accuracy. School depth and school mean volume backscattering strength were the most important predictors in determining the school classification. This study supports other publications suggesting that random forests represent a promising approach to acoustic target classification in fisheries science.
... Medical image annotation, for example, is mainly done by medical professionals with pertinent knowledge about human anatomy [4]. Other examples include the annotation of legal documents [5] or fisheries acoustics echograms [6]. Manual annotation is errorprone [1], as well as time-intensive, monotonous, and exhausting [6,7]. ...
... Other examples include the annotation of legal documents [5] or fisheries acoustics echograms [6]. Manual annotation is errorprone [1], as well as time-intensive, monotonous, and exhausting [6,7]. Thus, it is often difficult for annotators to stay motivated and engaged with the task, which is however an important determinant of annotation quality [8,9]. ...
Poorly annotated data is a common problem for data-intensive applications like supervised machine learning. In domains like healthcare, annotation tasks require specific domain knowledge and are thus often done manually by experts, which is error-prone, time-intensive, and tedious. In this study, we investigate gamification as a means to foster annotation quality through annota-tors' increased motivation and engagement. To this end, we conducted a literature review of 70 studies as well as a series of 16 workshops with a team of six experts in medical image annotation. We derive a set of seven meta-requirements (MRs) that represent the desired instrumental and experiential outcomes of gamified expert annotation systems (e.g., high-quality annotations, a sense of challenge) as well as a tentative design that can address the derived MRs. Our results help to understand the inner workings of gamification in the context of expert annotation and lay important groundwork for designing gamified expert annotation systems that can successfully motivate annotators and increase annotation quality.
... The quality of the labelled data is also a big concern where it brings lots of incertitude and biased decisions. Domain expertise is essential in labelling, but still different experts will generate different labels for the same elements [8]. Data labelling also introduces unnecessary privacy considerations. ...
... There are different classification algorithms such as Decision Tree [39], Artificial Neural Networks and Deep Learning models [8,40], Support Vector Machines [41], Naïve Bays etc. Supervised learning algorithms are applied in variety of fields such as object recognition [42,43], object detection [40,44,45], image and colour analysis [46][47][48] and natural language processing which includes a variety of tasks such as language detection, question answering, language understanding and translations [49][50][51]. For spam detection many of the current OSNs use supervised learning algorithms [6,7,52]. ...
Online Social Networks (OSNs) are a popular platform for communication and collaboration. Spammers are highly active in OSNs. Uncovering spammers has become one of the most challenging problems in OSNs. Classification-based supervised approaches are the most commonly used method for detecting spammers. Classification-based systems suffer from limitations of “data labelling”, “spam drift”, “imbalanced datasets” and “data fabrication”. These limitations effect the accuracy of a classifier’s detection. An unsupervised approach does not require labelled datasets. We aim to address the limitation of data labelling and spam drifting through an unsupervised approach.We present a pure unsupervised approach for spammer detection based on the peer acceptance of a user in a social network to distinguish spammers from genuine users. The peer acceptance of a user to another user is calculated based on common shared interests over multiple shared topics between the two users. The main contribution of this paper is the introduction of a pure unsupervised spammer detection approach based on users’ peer acceptance. Our approach does not require labelled training datasets. While it does not better the accuracy of supervised classification-based approaches, our approach has become a successful alternative for traditional classifiers for spam detection by achieving an accuracy of 96.9%.
... More recently, ML methods were applied to acoustic data gathered by a commercial echosounder buoy to identify tropical tuna aggregations (Baidai et al. 2020). Meanwhile, hydroacoustic data analysis has been streamlined through the development of CNN to aid in the task of labeling data (Sarr et al. 2020). Underwater in situ species identification can be carried out-in a labor-intensive and expensive manner-by divers, with minimal impact on sensitive communities. ...
Fisheries science aims to understand and manage marine natural resources. It relies onresource-intensive sampling and data analysis. Within this context, the emergence of machinelearning (ML) systems holds significant promise for understanding disparate components ofthese marine ecosystems and gaining a greater understanding of their dynamics. The goal ofthis paper is to present a review of ML applications in fisheries science. It highlights boththeir advantages over conventional approaches and their drawbacks, particularly in terms ofoperationality and possible robustness issues. This review is organized from small to largescales. It begins with genomics and subsequently expands to individuals (catch items),aggregations of different species in situ , on-board processing, stock/populations assessmentand dynamics, spatial mapping, fishing-related organizational units, and finally ecosystemdynamics. Each field has its own set of challenges, such as pre-processing steps, the quantityand quality of training data, the necessity of appropriate model validation, and knowingwhere ML algorithms are more limited, and we discuss some of these discipline-specificchallenges. The scope of discussion of applied methods ranges from conventional statisticalmethods to data-specific approaches that use a higher level of semantics. The paper concludeswith the potential implications of ML applications on management decisions and a summaryof the benefits and challenges of using these techniques in fisheries.
... Sarr et al. utilized deep learning methods for data labeling. It helped human experts refine the essence of echogram data [31]. ...
The main problem with supervised learning is data labeling, an activity that seems trivial when the data is small, but not if the data is very large, such as LC-MS (Liquid Chromatography-Mass Spectrometry) data. This task requires high concentration and accuracy if done by humans and impacts processing time. This paper discusses a method to automate labeling of LC-MS data to speed up processing time. In this case, webscraping technique is utilized to retrieve the labels because they are stored in an online database. It has been done in previous studies, but the results are not satisfactory because it still takes a long time to get the required label which is the name of the chemical compound. This is due to frequent disconnections. To solve this problem, a local mirror database is built so that it can be accessed locally. We built two system architectures. The first utilizes two separate computers as a server and client. They are connected to the access point. The second is to utilize a single computer, acting as both server and client at the same time. Theoretically, this will reduce the distance and save labeling time. The system architecture has succeeded in labeling the required data and has a time efficiency of 96.4% and 96.67%, respectively, compared to previous studies. This is a massive time saver.
... In addition, frequency differencing has to assume that scattering at each location is dominated by a single type of animal, and therefore is ill-suited to mixed aggregations of scatterers. More recently, a variety of machine learning and artificial intelligence (ML/AI) approaches have been applied to the problem (Roberts et al., 2011 ;Brautaset et al., 2020 ;Sarr et al., 2021 ). The promise of these techniques is that with big-enough data, advanced algorithms will be able to detect subtle patterns that humans and simple models cannot. ...
Identifying sound-scattering organisms is a perennial challenge in fisheries acoustics. Most practitioners classify backscatter based on direct sampling, frequency-difference thresholds, and expert judgement, then echo-integrate at a single frequency. However, this approach struggles with species mixtures, and discards multi-frequency information when integrating. In version methods do not ha v e these limitations, but are not widely used because species identifications are often ambiguous and the algorithms are complicated to implement. We address these shortcomings using a probabilistic, Bayesian in version method. Like other inversion methods, it handles species mixtures, uses all available frequencies, and extends naturally to broadband signals. Unlike previous approaches, it leverages Bayesian priors to rigorously incorporate information from direct sampling and biological knowledge, constraining the inversion and reducing ambiguity in species identification. Because it is probabilistic, a well-specified model should not produce solutions that are both wrong and confident. The model is based on physical scattering processes, so its output is fully interpretable, unlike some machine learning methods. Finally, the approach can be implemented using existing Bayesian libraries and is easily parallelized f or large datasets. We present examples using simulations and field data from the Gulf of Alaska, and discuss possible applications and extensions of the method.
... In addition, a great effort is needed when organizing and labeling data; this task can be time consuming and critical in the present field of application. A standardization of the expert labelling process of complex data, exploiting innovative approaches, is desirable and should be investigated in the future 86 . ...
Although the Mediterranean Sea is a crucial hotspot in marine biodiversity, it has been threatened by numerous anthropogenic pressures. As flagship species, Cetaceans are exposed to those anthropogenic impacts and global changes. Assessing their conservation status becomes strategic to set effective management plans. The aim of this paper is to understand the habitat requirements of cetaceans, exploiting the advantages of a machine-learning framework. To this end, 28 physical and biogeochemical variables were identified as environmental predictors related to the abundance of three odontocete species in the Northern Ionian Sea (Central-eastern Mediterranean Sea). In fact, habitat models were built using sighting data collected for striped dolphins Stenella coeruleoalba, common bottlenose dolphins Tursiops truncatus, and Risso’s dolphins Grampus griseus between July 2009 and October 2021. Random Forest was a suitable machine learning algorithm for the cetacean abundance estimation. Nitrate, phytoplankton carbon biomass, temperature, and salinity were the most common influential predictors, followed by latitude, 3D-chlorophyll and density. The habitat models proposed here were validated using sighting data acquired during 2022 in the study area, confirming the good performance of the strategy. This study provides valuable information to support management decisions and conservation measures in the EU marine spatial planning context.
... Due to the heavy workload of manual data labeling, some automatic methods were tried in practice, effectively with audio signal [23], still relying on data preprocessing and correction by expert intervention. Utilizing the BIT record to label the flight data might be a feasible approach for UAV compound fault diagnosis in the flight test stage, and would be beneficial. ...
In the process of Unmanned Aerial Vehicle (UAV) flight testing, plenty of compound faults exist, which could be composed of concurrent single faults or over-limit states alarmed by Built-In-Test (BIT) equipment. At present, there still lacks a suitable automatic labeling approach for UAV flight data, effectively utilizing the information of the BIT record. The performance of the originally employed flight data-driven fault diagnosis models based on machine learning needs to be improved as well. A compound fault labeling and diagnosis method based on actual flight data and the BIT record of the UAV during flight test phase is proposed, through labeling the flight data with compound fault modes corresponding to concurrent single faults recorded by the BIT system, and upgrading the original diagnosis model based on Gradient Boosting Decision Tree (GBDT) and Fully Convolutional Network (FCNN), to eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and modified Convolutional Neural Network (CNN). The experimental results based on actual test flight data show that the proposed method could effectively label the flight data and obtain a significant improvement in diagnostic performance, appearing to be practical in the UAV test flight process.
A study examining animal use in research, testing and education in Africa, as well as alternative, non-animal, approaches across the continent. The report also looks at legislation of animal research, animal welfare across the continent and other related issues.
Nature provides strength to people, and the ocean is one that is important. In the ocean, various resources are available to both nature and people. One such priceless resource is fish. Many varieties of fish are available in the sea. Seafood (fish and crab) is a healthy food for enhancing people’s health. But, nowadays so much offshore drilling for natural oil and gas production creates a huge destructive impact on marine resources, not only offshore due to other human mistakes spoiling the marine resources, especially fish. For this reason, we can save our marine resources with the help of new computer technology (artificial intelligence, deep learning, image processing). Aquaculture is one of the industries that currently frequently uses deep learning-based evaluation of images. Integrating an undersea time-lapse image camera with image analysis based on deep learning, we proposed this abstract for marine resources especially fish resources, first and foremost identify the fish, which means a variety of fish classified, then restoring the particular species, and eventually monitoring the movements under the sea. Optical cameras have been employed as a quick and affordable aquaculture technique for monitoring fish populations and detecting marine species. The photographs utilized in this research came from a variety of cameras; therefore, the pixels and size ratios varied. Additionally, it would be computationally expensive and ineffective to process the entire image. To eliminate the additional computational load, it was decided to crop the photographs to a uniform size. For millions of years, eating aquatic animals—especially fish—has been a key element of human nutrition. It is widely acknowledged that fish is an excellent provider of vital elements, such as proteins and vitamins. If the above technologies are utilized effectively, we can save our marine resources forever.
Generating biomass-at-age indices for fisheries stock assessments with acoustic data collected by uncrewed surface vessels (USVs) has been hampered by the need to resolve acoustic backscatter with contemporaneous biological (e.g., age) composition data. To address this limitation, Pacific hake (Merluccius productus; “hake”) acoustic data were gathered from a USV survey (in 2019) and acoustic-trawl survey (ATS; 2019 and eight previous years), and biological data were gathered from fishery-dependent and non-target (i.e., not specifically targeting hake) fishery-independent sources (2019 and eight previous years). To overcome the lack of contemporaneous biological sampling in the USV survey, age class compositions were estimated from a generalized linear mixed spatio-temporal model (STM) fit to the fishery-dependent and non-target fishery-independent data. The validity of the STM age composition estimation procedure was assessed by comparing estimates to age compositions from the ATS in each year. Hake biomass-at-age was estimated from all combinations of acoustics (USV or ATS in 2019, ATS only in other years) and age composition information (STM or ATS in all years). Across the survey area, proportional age class compositions derived from the best STM differed from ATS observations by 0.09 on average in 2019 (median relative error (MRE): 19.45%) and 0.14 across all years (MRE: 79.03%). In data-rich areas (i.e., areas with regular fishery operations), proportional age class compositions from the STM differed from ATS observations by 0.03 on average in 2019 (MRE: 11.46%) and 0.09 across years (MRE: 54.96%). On average, total biomass estimates derived using STM age compositions differed from ATS age composition-based estimates by approximately 7% across the study period (~ 3% in 2019) given the same source of acoustic data. When biomass estimates from different sources of acoustic data (USV or ATS) were compared given the same source of age composition data, differences were nearly ten-fold greater (22% or 27%, depending on if ATS or STM age compositions were used). STMs fit to non-contemporaneous data may provide suitable information for assigning population structure to acoustic backscatter in data-rich areas, but advancements in acoustic data processing (e.g., automated echo classification) may be needed to generate viable USV-based estimates of biomass-at-age.
The knowledge of fish assemblage in submarine environments is a keystone for fishery management but remain poorly monitored due to heavy methodological requirements for scientific fish sampling. Passive acoustic monitoring appeared as a promising non intrusive alternative to fish sampling, potentially providing more objective data at a reduced cost, but still rely on heavy expert analysis of the acoustic signal. We propose in this paper to improve passive acoustic monitoring efficiency with deep learning. Using convolutional recurrent neural networks, we built and tested models able to detect two types of fish vocalizations as well as motor engine sounds. The detector had a high F1 score and an error rate . These methods are promising tools for the fish communities management as they allow automatic fish sounds detection and classification.
Introduction
With a higher degree of automation, fishing vessels have gradually begun adopting a fishing monitoring method that combines human and electronic observers. However, the objective data of electronic monitoring systems (EMS) has not yet been fully applied in various fishing boat scenarios such as ship behavior recognition.
Methods
In order to make full use of EMS data and improve the accuracy of behaviors recognition of fishing vessels, the present study proposes applying popular deep learning technologies such as convolutional neural network, long short-term memory, and attention mechanism to Chub mackerel (Scomber japonicus) fishing vessel behaviors recognition. The operation process of Chub mackerel fishing vessels was divided into nine kinds of behaviors, such as “pulling nets”, “putting nets”, “fish pick”, “reprint”, etc. According to the characteristics of their fishing work, four networks with different convolutional layers were designed in the pre-experiment. And the feasibility of each network in behavior recognition of the fishing vessels was observed. The pre-experiment is optimized from the perspective of the data set and the network. From the standpoint of the data set, the size of the optimized data set is significantly reduced, and the original data characteristics are preserved as much as possible. From the perspective of the network, different combinations of pooling, long short-term memory(LSTM) network, and attention(including CBAM and SE) are added to the network, and their effects on training time and recognition effect are compared.
Results
The experimental results reveal that the deep learning methods have outstanding performance in behaviors recognition of fishing vessels. The LSTM and SE module combination produced the most apparent optimization effect on the network, and the optimized model can achieve an F1 score of 97.12% in the test set, surpassing the classic ResNet, VGGNet, and AlexNet.
Discussion
This research is of great significance to the management of intelligent fishery vessels and can promote the development of electronic monitoring systems for ships.
This study reviews current artificial intelligence (AI) systems legislation, the AI techniques definition proposed by the AI Act and main applications of AI methods in the fisheries sector with special focus on applications to enhance traceability of fishery products, fishing gear selectivity, good practices, and potential to help young people finding jobs. Finally, this study offers policy recommendations relevant to EU decision-making to achieve a better use of AI systems in the fisheries sector. Full report and translated executive summary available at: https://www.europarl.europa.eu/thinktank/es/document/IPOL_STU(2022)699643
Understanding the relationship between sound-scattering layers (SSLs) and pelagic habitat characteristics is a substantial step to apprehend ecosystem dynamics. SSLs are detected on echo sounders representing aggregated marine pelagic organisms. In this study, SSL characteristics of zooplankton and micronekton were identified during an upwelling event in two contrasting areas of the Senegalese continental shelf. Here a cold upwelling-influenced inshore area was sharply separated by a strong thermal boundary from a deeper, warmer, stratified offshore area. Mean SSL thickness and SSL vertical depth increased with the shelf depth. The thickest and deepest SSLs were observed in the offshore part of the shelf. Hence, zooplankton and micronekton seem to occur more frequently in stratified water conditions rather than in fresh upwelled water. Diel vertical and horizontal migrations of SSLs were observed in the study area. Diel period and physicochemical water characteristics influenced SSL depth and SSL thickness. Although chlorophyll-a concentration insignificantly affected SSL characteristics, the peak of chlorophyll a was always located above or in the middle of the SSLs, regularly matching with the peak of SSL biomass. Such observations indicate trophic relationships, suggesting SSLs to be mainly composed of phytoplanktivorous zooplankton and micronekton. Despite local hypoxia, below 30 m depth, distribution patterns of SSLs indicate no vertical migration boundary. The results increase the understanding of the spatial organization of mid-trophic species and migration patterns of zooplankton and micronekton, and they will also improve dispersal models for organisms in upwelling regions.
Acoustic data provide scientific and engineering insights in fields ranging from biology and communications to ocean and Earth science. We survey the recent advances and transformative potential of machine learning (ML), including deep learning, in the field of acoustics. ML is a broad family of techniques, which are often based in statistics, for automatically detecting and utilizing patterns in data. Relative to conventional acoustics and signal processing, ML is data-driven. Given sufficient training data, ML can discover complex relationships between features and desired labels or actions, or between features themselves. With large volumes of training data, ML can discover models describing complex acoustic phenomena such as human speech and reverberation. ML in acoustics is rapidly developing with compelling results and significant future promise. We first introduce ML, then highlight ML developments in four acoustics research areas: source localization in speech processing, source localization in ocean acoustics, bioacoustics, and environmental sounds in everyday scenes.
When a feed-forward neural network (FNN) is trained for acoustic source ranging in an ocean waveguide, it is difficult evaluating the FNN ranging accuracy of unlabeled test data. The label is the distance between source and receiver array. A fitting-based early stopping (FEAST) method is introduced to evaluate the FNN ranging error on test data where the distance to the source is unknown. Based on FEAST, when the evaluated ranging error is minimum on test data, training is stopped. This will improve the FNN ranging accuracy on the test data. The FEAST is demonstrated on simulated and experimental data.
This work presents a methodological synthesis for the in situ monitoring of fish aggregating devices (FADs) using a combination of optical, echosounder and SCUBA observations conducted in the vicinity of drifting FADs. The acoustic methods allowed, according to the devices used, the description of the spatial organisation and dynamics of biotic scattering layers, individual fishes, schools, shoals and mammals, while visual, photographic and video observations permitted species identification within a range of 0 to ~ 25 m. Based on these results, we elaborate on the interest to combine acoustic and visual methods, and present an autonomous instrumented drifting buoy for remotely monitoring fish diversity and abundance in the pelagic ecosystems. The perspective of autonomously collecting large amounts of basic information useful for ecological and fisheries studies in an ecosystem approach for open sea, as well as coastal pelagic environment, is also emphasized. As perspective we present “Seaorbiter” a futuristic large drifting platform which will allow performing innovative ecosystemic studies taking into account simultaneously all macro components of the pelagic ecosystem.
Matecho is an automated processing method to extract information and perform echo-integration and fish shoal extraction from various scientific echo-sounder sources providing digital acoustic data on fisheries and aquatic ecosystem. The open-source initiative helps foster collaboration and technological transfer. Matecho supports various formats, such as the international standard format for the exchange of fisheries acoustics raw and edited data. The procedure allows the semiautomatic cleaning of echogram data and the application of automatic data filters, i.e., transient noise, attenuated signal and impulsive noise removal and background noise reduction. Echo-integration processing is executed for each depth layer and integrates their characteristics per elementary sampling unit. Sound scattered layers are automatically detected by segmentation from the echo-integrated echogram, and shoals are extracted according to an iterative process of aggregation of filtered echogram echoes that allows, in both cases, the calculation of the ad hoc parameters describing morphological, spatial location and acoustic characteristics of sound scattered layers and shoals. Matecho is open-source software for researchers and provides end users with a user-friendly, free executable program.
Deep neural networks (DNNs) are advantageous for representing complex nonlinear relationships. This paper applies DNNs to source localization in a shallow water environment. Two methods are proposed to estimate the range and depth of a broadband source through different neural network architectures. The first adopts the classical two-stage scheme, in which feature extraction and DNN analysis are independent steps. The eigenvectors associated with the modal signal space are extracted as the input feature. Then, the time delay neural network is exploited to model the long term feature representation and constructs the regression model. The second concerns a convolutional neural network−feed-forward neural network (CNN–FNN) architecture, which trains the network directly by taking the raw multi-channel waveforms as input. The CNNs are expected to perform spatial filtering for multi-channel signals, in an operation analogous to time domain filters. The outputs of CNNs are summed as the input to FNN. Several experiments are conducted on the simulated and experimental data to evaluate the performance of the proposed methods. The results demonstrate that DNNs are effective for source localization in complex and varied water environments, especially when there is little precise environmental information.
The classification and recognition technology of underwater acoustic signal were always an important research content in the field of underwater acoustic signal processing. Currently, wavelet transform, Hilbert-Huang transform, and Mel frequency cepstral coefficients are used as a method of underwater acoustic signal feature extraction. In this paper, a method for feature extraction and identification of underwater noise data based on CNN and ELM is proposed. An automatic feature extraction method of underwater acoustic signals is proposed using depth convolution network. An underwater target recognition classifier is based on extreme learning machine. Although convolution neural networks can execute both feature extraction and classification, their function mainly relies on a full connection layer, which is trained by gradient descent-based; the generalization ability is limited and suboptimal, so an extreme learning machine (ELM) was used in classification stage. Firstly, CNN learns deep and robust features, followed by the removing of the fully connected layers. Then ELM fed with the CNN features is used as the classifier to conduct an excellent classification. Experiments on the actual data set of civil ships obtained 93.04% recognition rate; compared to the traditional Mel frequency cepstral coefficients and Hilbert-Huang feature, recognition rate greatly improved.
This work compiles hydroacoustic recordings and catch data over Senegambia (Senegal and The Gambia) from assessment surveys on the major clupeid species to identify sources of bias in abundance estimates caused by their horizontal distribution. The latitudinal distribution of small pelagic fish is often well known, while their "across shelf" distribution on the continental shelf is less understood. The southern part of the Senegambian shelf has a wide shallow water (<10 m) area that makes up 20% (1500 NM²) of the total shelf surface, while the northern part accounts for 3% (200 NM²). These areas are not assessed by conventional fisheries acoustics surveys and therefore increase the uncertainty of the assessment of these species. Our findings show that this likely introduces a bias in the assessment of Sardinella maderensis, while for S. aurita no major estimation-error is caused by their horizontal distribution. The data confirm that Ethmalosa fimbriata and Ilisha africana are challenging to assess by conventional surveys, due to their mostly inshore distribution. We emphasise the usefulness of assessing S. aurita through fisheries independent hydroacoustic surveys, and propose alternative methods to survey shallow water areas to reduce biases in biomass estimates and distribution mapping.
Deep Learning has revolutionized vision via convolutional neural networks (CNNs) and natural language processing via recurrent neural networks (RNNs). However, success stories of Deep Learning with standard feed-forward neural networks (FNNs) are rare. FNNs that perform well are typically shallow and, therefore cannot exploit many levels of abstract representations. We introduce self-normalizing neural networks (SNNs) to enable high-level abstract representations. While batch normalization requires explicit normalization, neuron activations of SNNs automatically converge towards zero mean and unit variance. The activation function of SNNs are "scaled exponential linear units" (SELUs), which induce self-normalizing properties. Using the Banach fixed-point theorem, we prove that activations close to zero mean and unit variance that are propagated through many network layers will converge towards zero mean and unit variance -- even under the presence of noise and perturbations. This convergence property of SNNs allows to (1) train deep networks with many layers, (2) employ strong regularization, and (3) to make learning highly robust. Furthermore, for activations not close to unit variance, we prove an upper and lower bound on the variance, thus, vanishing and exploding gradients are impossible. We compared SNNs on (a) 121 tasks from the UCI machine learning repository, on (b) drug discovery benchmarks, and on (c) astronomy tasks with standard FNNs and other machine learning methods such as random forests and support vector machines. SNNs significantly outperformed all competing FNN methods at 121 UCI tasks, outperformed all competing methods at the Tox21 dataset, and set a new record at an astronomy data set. The winning SNN architectures are often very deep. Implementations are available at: github.com/bioinf-jku/SNNs.
Source localization is solved as a classification problem by training a feed-forward neural network (FNN) on ocean acoustic data. The pressure received by a vertical linear array is preprocessed by constructing a normalized sample covariance matrix (SCM), which is used as input for the FNN. Each neuron of the output layer represents a discrete source range. FNN is a data-driven method that learns features directly from observed acoustic data, unlike model-based localization methods such as matched-field processing that require accurate sound propagation modeling. The FNN achieves a good performance (the mean absolute percentage error below 10\%) for predicting source ranges for vertical array data from the Noise09 experiment. The effects of varying the parameters of the method, such as number of hidden neurons and layers, number of output neurons and number of snapshots in each input sample are discussed.
Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction. These methods have dramatically improved the state-of-the-art in speech recognition, visual object recognition, object detection and many other domains such as drug discovery and genomics. Deep learning discovers intricate structure in large data sets by using the backpropagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representation in each layer from the representation in the previous layer. Deep convolutional nets have brought about breakthroughs in processing images, video, speech and audio, whereas recurrent nets have shone light on sequential data such as text and speech.
This book reviews an extensive range of material from a wide range of established scientists in the fisheries acoustics community. It is a timely and indispensable book for fisheries science which needs efficient and standardized procedures for fish sampling. Today, fisheries acoustics is a central discipline for in situ observations of aquatic organisms extending from plankton to whales. Acoustic methods are needed for stock assessment exercises and for behavioural studies, starting from freshwater rivers and extending to the open ocean, including inland lakes and estuaries. Present uses of fisheries acoustics are not only directed at assessment methods, but also at ecological and management studies. The main advantage of acoustics is that it provides for the possibility of collecting information either in an instant or over an extended period, with observations being at all scales from mm to km; for example, from fish to schools at the ‘micro‐scale’, school to a cluster of schools at the ‘meso‐scale’ and clusters to populations at the ‘macro‐scale’. Observations can be made independently of intrusive fishing operations and are not constrained by the limits of visual observation methods. In fisheries science, whether the approach is at the ecosystem level or is just dealing with individual populations using classical models, fisheries acoustics methods are crucial for an accurate validation of some key parameters for management models. It should also be born in mind that the observation of marine organisms remains particularly difficult in comparison to aerial or terrestrial animals.
This new edition of Fisheries acoustics is thus very welcome; the techniques and methods of measurement have quickly evolved during the 12 years that have passed since the first edition. The authors have integrated into the text the new developments that have occurred since then, following the main new trends in the field. Formerly, analyses of acoustics data were limited to specialists and required a long time for processing. The developments in personal computers have increased capabilities in all domains such as in central unit processors, virtual memory, as well as in signal and image analysis. These developments now allow data collection, treatment and analysis using adapted software as a real possibility even to the non‐specialists.
Fisheries acoustics manages to be both a book that can be used to learn about underwater acoustics in fisheries and to be also a good overview of the methodologies already used widely by national and international organizations. The book has been redrafted with well‐organized chapters each written by one of the two authors. It reviews everything from underwater sound introduction to the state of the art in the analysis of acoustic data. Electronic tagging and tracking techniques are not included, reflecting the regrettable ‘separation’ between the two scientific communities. Some biological aspects are perhaps dealt with rather briefly, such as the avoidance reaction of fish in front of a vessel, but the pragmatism of both authors avoids scientific polemics and in most areas focuses directly on the essential points, which allows specialists and non‐specialists alike to understand the science of fisheries acoustics. There are also some very well‐organized colour figures. It would have been perhaps useful to have a reference list at the end of each chapter, or maybe a classification of references by topic.
From a retrospective point of view a book captures a moment in time. As with all scientific fields, fisheries acoustics techniques and methods continue to progress. We can already mention, as an example, that just after the book appeared, results were published on the use of omnidirectional sonar in fisheries science, and on the use of ocean‐acoustic waveguide remote sensing in continental shelf environments. Considering the prolific literature produced by scientists around the world, the authors have judiciously selected 770 references as the key introductory texts, in order to offer an essential book of theory and practice to the reader.
It is necessary to refine the acoustic estimation method applied in the Gulf of California, especially regarding the acoustic energy allocation to the different species coexisting in the same area, in order to produce a more precise and reliable fishery independent biomass index for the sustainable management of the fishery. The series of acoustic surveys carried out in this work are the first systematic attempts to estimate Pacific sardine biomass in the Gulf of California. Due to similar acoustic reflectivity of other pelagic species, we need to refine the allocation of acoustic energy to species. From 2012, besides the 38 kHz transducer another frequency has been incorporated (120 kHz) in the surveys. With both frequencies we will apply acoustic signal filtering procedures in order to (1). Discriminate zooplankton echoes from fish echoes and (2). Investigate the possibility to better discriminate the Pacific sardine from other species.
Vocal communication is a primary communication method of killer and pilot whales, and is used for transmitting a broad range of messages and information for short and long distance. The large variation in call types of these species makes it challenging to categorize them. In this study, sounds recorded by audio sensors carried by ten killer whales and eight pilot whales close to the coasts of Norway, Iceland, and the Bahamas were analyzed using computer methods and citizen scientists as part of the Whale FM project. Results show that the computer analysis automatically separated the killer whales into Icelandic and Norwegian whales, and the pilot whales were separated into Norwegian long-finned and Bahamas short-finned pilot whales, showing that at least some whales from these two locations have different acoustic repertoires that can be sensed by the computer analysis. The citizen science analysis was also able to separate the whales to locations by their sounds, but the separation was somewhat less accurate compared to the computer method.
Echo sounding is a powerful and widely used technique for remote sensing of the marine environment. In order to enhance the power of the echo sounder, a postprocessing system has been designed and realized in standard software that is essentially machine independent. This has been done by adhering to the following international standards: UNIX operating system, C programming language, X Window Systems, Structured-Query Language (SQL) for communication with a relational database, and Transport Control Protocol/Internet Protocol (TCP?IP). Preprocessed data are transferred from the echo sounder to the postprocessing system by means of a local-area network (LAN), namely Ethernet. Development of the postprocessing system, for analysis of such diverse scatterers as plankton, pelagic, and bottom fish, and the bottom itself, is documented in the following way. The history of echo integration is summarized. User requirements for the new system are listed. Reasons are given for the choice of the particular computing environment, including both hardware, software, and external communications. The system design, consisting of data flow and graphical user interfaces, is described. Implementation of the system is defined through integration techniques and a discussion of performance issues. Operating procedures and the first field trials of the system are described. Several features characteristic of and perhaps unique to the postprocessing system are, for example: (1) user definition of arbitrarily shaped integration regions, including non-constant-depth intervals, by means of interactive graphics; (2) preprocessor error correction, e.g., adjustment of the noise threshold or redefinition of the detected bottom; (3) use of several color map techniques in order to extract such information as signal strength and shape; and (4) the scheme of interconnections of graphical user interfaces, database, and data files. This work does not introduce a set of computer instructions. It does describe a design philosophy and method of realization that may have broader applications in acoustics than that ostensibly concerned only with the quantitative estimation of fish abundance.
A class of predictive densities is derived by weighting the observed samples in maximizing the log-likelihood function. This approach is effective in cases such as sample surveys or design of experiments, where the observed covariate follows a different distribution than that in the whole population. Under misspecification of the parametric model, the optimal choice of the weight function is asymptotically shown to be the ratio of the density function of the covariate in the population to that in the observations. This is the pseudo-maximum likelihood estimation of sample surveys. The optimality is defined by the expected Kullback–Leibler loss, and the optimal weight is obtained by considering the importance sampling identity. Under correct specification of the model, however, the ordinary maximum likelihood estimate (i.e. the uniform weight) is shown to be optimal asymptotically. For moderate sample size, the situation is in between the two extreme cases, and the weight function is selected by minimizing a variant of the information criterion derived as an estimate of the expected loss. The method is also applied to a weighted version of the Bayesian predictive density. Numerical examples as well as Monte-Carlo simulations are shown for polynomial regression. A connection with the robust parametric estimation is discussed.
To investigate the hydrodynamic activity of the seabed in the German Bight, underwater remote sensing was carried out over an area of 32km2 located 20km northeast of Helgoland island in the southern North Sea in January, May and August 2001. On the basis of acoustic seabed classification, six seabed types have been identified by the combined evaluation of side-scan sonar records, wave-shape analysis of echo-sounder data, and 100 grab samples. In five seabed types, the acoustic classes can be distinguished on the basis of sediment characteristics, comprising size components ranging from coarse pebbles to fine sand. The sixth seabed type corresponds to large pebbles and cobbles which are completely overgrown with brown algae. Statistically, the complex spatial patchiness of the six classes varied significantly in the course of the study period. During the winter period (January 2001), the study site was dominated by coarse material, except for a small area of finer sediment in the centre. With the onset of more moderate weather conditions in spring (May 2001), a general fining trend in sediment composition was observed, especially in the deeper western parts of the study area. In summer (August 2001), finer sediments still dominated but a slight increase in signal roughness suggests an overprint by coarser lag deposits and/or denser coverage by benthic organisms (e.g. Lanice conchilega) which then were found more frequently in grab samples, in association with finer sand. These findings demonstrate that the distribution of seafloor sediments and their benthic fauna in the deeper part of the German Bight region are controlled largely by seasonal changes in hydrodynamic conditions. These changes are reflected in correspondingly high variability in the complex patchiness of sediment distribution patterns, which would not have been adequately resolved by any standard sampling procedure.
One of the most troublesome problems for fisheries acoustic surveys is the acquisition of ground truth data. A major incentive for the use of acoustics is the deficiencies of the same alternate techniques that must be used to collect ground truth data. However, such data are often required for biological information such as species composition, body size and spawning condition. In addition, since acoustic techniques are not universally accepted, ground truth information was sometimes needed to establish confidence in the results. Many different approaches are used to obtain ground truth data. Examination of these results provides some interesting insights into the characteristics of both the acoustic and the ground truth techniques and suggests some principles for ground truth efforts used in conjunction with acoustics surveys. The appropriate role of ground truth efforts is not only verification, but also to aid interpretation of the acoustic data. We conclude that the use of a single method to obtain ground truth data is generally likely to be inadequate. We recommend multiple methods for ground truth, so that inadequacy of individual techniques is to some extent compensated by supplementary methods. This requires a reorientation of planning for most fisheries acoustic surveys.
We consider the problem of learning accurate models from multiple sources of “nearby” data. Given distinct samples from multiple data sources and estimates of the dissimilarities between these sources, we provide a general theory of which samples should be used to learn models for each source. This theory is applicable in a broad decision-theoretic learning framework, and yields general results for classification and regression. A key component of our approach is the development of approximate triangle inequalities for expected loss, which may be of independent interest. We discuss the related problem of learning parameters of a distribution from multiple data sources. Finally, we illustrate our theory through a series of synthetic simulations.
The explosion of image data on the Internet has the potential to foster more sophisticated and robust models and algorithms to index, retrieve, organize and interact with images and multimedia data. But exactly how such data can be harnessed and organized remains a critical problem. We introduce here a new database called ldquoImageNetrdquo, a large-scale ontology of images built upon the backbone of the WordNet structure. ImageNet aims to populate the majority of the 80,000 synsets of WordNet with an average of 500-1000 clean and full resolution images. This will result in tens of millions of annotated images organized by the semantic hierarchy of WordNet. This paper offers a detailed analysis of ImageNet in its current state: 12 subtrees with 5247 synsets and 3.2 million images in total. We show that ImageNet is much larger in scale and diversity and much more accurate than the current image datasets. Constructing such a large-scale database is a challenging task. We describe the data collection scheme with Amazon Mechanical Turk. Lastly, we illustrate the usefulness of ImageNet through three simple applications in object recognition, image classification and automatic object clustering. We hope that the scale, accuracy, diversity and hierarchical structure of ImageNet can offer unparalleled opportunities to researchers in the computer vision community and beyond.
This work aims at defining and testing a set of techniques that enables agents to use information from several sources during learning. In Multiagent Systems (MAS) it is frequent that several agents need to learn similar concepts in paral- lel. In this type of environment there are more possibilities for learning than in classical Machine Learning. Exchange of information between teams of agents that are attempting to solve similar problems may increase accuracy and learn- ing speed at the expense of communication. One of the most interesting possibilities to explore is when the agents may have different structures and learning algorithms, thus pro- viding different perspectives on the problems they are fac- ing. In this paper the authors report the results of experi- ments made in a traffic control simulation with and without exchanging information between learning agents.
For many supervised learning tasks it may be infeasible (or very expensive) to obtain objective and reliable labels. Instead, we can collect subjective (possi bly noisy) labels from multiple experts or annotators. In practice, there is a substantial amount of di sagreement among the annotators, and hence it is of great practical interest to address conventio nal supervised learning problems in this scenario. In this paper we describe a probabilistic approac h for supervised learning when we have multiple annotators providing (possibly noisy) labels but no absolute gold standard. The proposed algorithm evaluates the different experts and also gives an estimate of the actual hidden labels. Experimental results indicate that the proposed method is superior to the commonly used majority voting baseline.
INTRODUCTIONThe ability of multilayer back-propagation networks to learn complex, high-dimensional, nonlinearmappings from large collections of examples makes them obvious candidates for imagerecognition or speech recognition tasks (see PATTERN RECOGNITION AND NEURALNETWORKS). In the traditional model of pattern recognition, a hand-designed featureextractor gathers relevant information from the input and eliminates irrelevant variabilities.A trainable classifier then categorizes the...
Scikit-learn is a Python module integrating a wide range of state-of-the-art
machine learning algorithms for medium-scale supervised and unsupervised
problems. This package focuses on bringing machine learning to non-specialists
using a general-purpose high-level language. Emphasis is put on ease of use,
performance, documentation, and API consistency. It has minimal dependencies
and is distributed under the simplified BSD license, encouraging its use in
both academic and commercial settings. Source code, binaries, and documentation
can be downloaded from http://scikit-learn.sourceforge.net.
We introduce a challenging set of 256 object categories containing a total of 30607 images. The original Caltech-101 [1] was collected by choosing a set of object categories, downloading examples from Google Images and then manually screening out all images that did not fit the category. Caltech-256 is collected in a similar manner with several improvements: a) the number of categories is more than doubled, b) the minimum number of images in any category is increased from 31 to 80, c) artifacts due to image rotation are avoided and d) a new and larger clutter category is introduced for testing background rejection. We suggest several testing paradigms to measure classification performance, then benchmark the dataset using two simple metrics as well as a state-of-the-art spatial pyramid matching [2] algorithm. Finally we use the clutter category to train an interest detector which rejects uninformative background regions.
Acoustic target classification is the process of assigning observed acoustic backscattering intensity to an acoustic category. A deep learning strategy for acoustic target classification using a convolutional network is developed, consisting of an encoder and a decoder, which allow the network to use pixel information and more abstract features. The network can learn features directly from data, and the learned feature space may include both frequency response and school morphology. We tested the method on multifrequency data collected between 2007 and 2018 during the Norwegian sandeel survey. The network was able to distinguish between sandeel schools, schools of other species, and background pixels (including seabed) in new survey data with an F1 score of 0.87 when tested against manually labelled schools. The network separated schools of sandeel and schools of other species with an F1 score of 0.94. A traditional school classification algorithm obtained substantially lower F1 scores (0.77 and 0.82) when tested against the manually labelled schools. To train the network, it was necessary to develop sampling and preprocessing strategies to account for unbalanced classes, inaccurate annotations, and biases in the training data. This is a step towards a method to be applied across a range of acoustic trawl surveys.
In this paper, machine learning is introduced to source localization in underwater ocean waveguides. Source localization is regarded as a supervised learning regression problem and is solved by generalized regression neural network (GRNN). As a feed-forward network, GRNN is built using training data with fixed structure and configuration. The normalized sample covariance matrix (SCM) formed over a number of snapshots, and the corresponding source position are used as the input and output for GRNN. The source position can be estimated directly from the normalized SCM with GRNN; the proposed approach is thus in theory data driven. In addition, there is only one parameter, the spread factor, to be learned for GRNN. The optimal spread factor is determined using cross-validation. The regression method of GRNN is compared with the classification method of feed-forward neural network (FNN), as well as the classical method of matched field processing (MFP) for vertical array data from the SWellEx-96 experiment. The results show that GRNN achieves a satisfactory localization performance that outperforms both FNN and MFP. The proposed approach provides an alternative way for underwater source localization, especially in the absence of a priori environmental information or an appropriate propagation model.
The propagation of sound in a shallow water environment is characterized by boundary reflections from the sea surface and sea floor. These reflections result in multiple (indirect) sound propagation paths, which can degrade the performance of passive sound source localization methods. This paper proposes the use of convolutional neural networks (CNNs) for the localization of sources of broadband acoustic radiated noise (such as motor vessels) in shallow water multipath environments. It is shown that CNNs operating on cepstrogram and generalized cross-correlogram inputs are able to more reliably estimate the instantaneous range and bearing of transiting motor vessels when the source localization performance of conventional passive ranging methods is degraded. The ensuing improvement in source localization performance is demonstrated using real data collected during an at-sea experiment.
TensorFlow [1] is an interface for expressing machine learning algorithms, and an implementation for executing such algorithms. A computation expressed using TensorFlow can be executed with little or no change on a wide variety of heterogeneous systems, ranging from mobile devices such as phones and tablets up to large-scale distributed systems of hundreds of machines and thousands of computational devices such as GPU cards. The system is flexible and can be used to express a wide variety of algorithms, including training and inference algorithms for deep neural network models, and it has been used for conducting research and for deploying machine learning systems into production across more than a dozen areas of computer science and other fields, including speech recognition, computer vision, robotics, information retrieval, natural language processing, geographic information extraction, and computational drug discovery. This paper describes the TensorFlow interface and an implementation of that interface that we have built at Google. The TensorFlow API and a reference implementation were released as an open-source package under the Apache 2.0 license in November, 2015 and are available at www.tensorflow.org.
In recent years, Deep Learning (DL), a re-branding of neural networks (NNs), has risen to the top in numerous areas, namely Computer Vision (CV), speech recognition, natural language processing, etc. Whereas remote sensing (RS) possesses a number of unique challenges, primarily related to sensors and applications, inevitably RS draws from many of the same theories as CV; e.g., statistics, fusion, and machine learning, to name a few. This means that the RS community should not only be aware of advancements like DL, but also be leading researchers in this area. Herein, we provide the most comprehensive survey of state-of-the-art RS DL research. We also review recent new developments in the DL field that can be used in DL for RS. Namely, we focus on theories, tools and challenges for the RS community. Specifically, we focus on unsolved challenges and opportunities as it relates to (i) inadequate data sets, (ii) human-understandable solutions for modelling physical phenomena, (iii) Big Data, (iv) non-traditional heterogeneous data sources, (v) DL architectures and learning algorithms for spectral, spatial and temporal data, (vi) transfer learning, (vii) an improved theoretical understanding of DL systems, (viii) high barriers to entry, and (ix) training and optimizing the DL.
Omnidirectional sonar surveys were conducted in close proximity to drifting fish aggregating devices (FADs) offshore Seychelles, western Indian Ocean, to investigate the number, size, and distribution of FAD-associated fish schools. Echotrace detection techniques applied on the raw multibeam data enabled the extraction of empirical statistics regarding inter-school distances, and allowed the visualization of the temporal evolution of the pelagic aggregation on a FAD-centered coordinate system. The sonar recordings revealed the concurrent existence of multiple fish schools that were spatially clustered and exhibited low permanence in size and structure. Schools were predominantly detected within a radius of 500 m from the FADs, although 15% of detections occurred between 500 to 1500 m from the floating devices. Fish school biomass detected with the sonar was aggregated into a few, large schools during daytime, and dispersed into a larger number of small schools during nighttime. Compared to daytime observations, nighttime schools maintained smaller inter-school distances and were located closer to the drifting FADs. The study demonstrates that horizontal sonars are powerful tools for studying the spatiotemproral distribution of large pelagic schools in the vicinity of drifting FADs.
Living resources of the sea and fresh water have long been an important source of food and economic activity. With fish stocks continuing to be over-exploited, there is a clear focus on fisheries management, to which acoustic methods can and do make an important contribution. The second edition of this widely used book covers the many technological developments which have occurred since the first edition; highly sophisticated sonar and computer processing equipment offer great new opportunities and Fisheries Acoustic, 2e provides the reader with a better understanding of how to interpret acoustic observations and put them to practical use.
Deep neural nets with a large number of parameters are very powerful machine learning systems. However, overfitting is a serious problem in such networks. Large networks are also slow to use, making it difficult to deal with overfitting by combining the predictions of many different large neural nets at test time. Dropout is a technique for addressing this problem. The key idea is to randomly drop units (along with their connections) from the neural network during training. This prevents units from co-adapting too much. During training, dropout samples from an exponential number of different "thinned" networks. At test time, it is easy to approximate the effect of averaging the predictions of all these thinned networks by simply using a single unthinned network that has smaller weights. This significantly reduces overfitting and gives major improvements over other regularization methods. We show that dropout improves the performance of neural networks on supervised learning tasks in vision, speech recognition, document classification and computational biology, obtaining state-of-the-art results on many benchmark data sets. © 2014 Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever and Ruslan Salakhutdinov.
During acoustic surveys for biomass estimation, fish aggregations near the seabed may not be correctly measured due to false detection of the bottom echo. The extent of this problem depends on the spatial and density features of the near-seabed aggregations. We conducted experiments using a tower structure deployed close to the bottom of a sea loch. The tower is 10 m high with a split-beam transducer at the top and a fish cage at the bottom. The effect of a bottom slope is simulated by tilting the transducer. In experiments with various densities and sizes of gadoids in the cage, echoes from the vicinity of the seabed were studied over hard and soft ground. In addition to the range, the split-beam echosounder gives two angular coordinates of the target direction. These may be combined into one measure of angular displacement, namely, the angle between the apparent target direction and the transducer axis. We call this the split-beam angle (SBA). We found that the SBA is not necessarily an accurate indication of the target direction. Echoes from fish aggregations and the seabed have different characteristics in this respect. When the seabed echo is detected with few interfering targets above, the SBA is an accurate indication of the seabed slope, and assuming the slope does not change over a short series of pings, the SBA is highly correlated. On the other hand, the SBA from multiple fish echoes is highly variable, as expected, and the ping-to-ping variation is essentially random. Furthermore, when the seabed echo is transmitted through a substantial density of fish, the interference can change the SBA, although the ping-to-ping correlation cf the seabed SBA remains superior to that of fish aggregations. We also studied records from acoustic surveys on various research vessels to provide comparable results at full scale. When there is a low density of near-seabed fish, the correlation between the fore-aft SBA and the seabed gradient is optimal at the start of the first seabed echo; it declines at sub-bottom ranges. When there are dense aggregations of fish near the seabed, the automatic bottom-detection algorithm may be located on top of the aggregation, so that the echo integration misses a substantial quantity of fish. Examples from acoustic surveys in the North Sea are presented to illustrate this problem. Crown Copyright
A commercial data-visualization package, AVS, and database are used as the basis for a powerful and highly flexible acoustic
data analysis system. The system is easy to use and can be modified by the user to incorporate novel visualization and analysis
capabilities as required. Multi-frequency ping-by-ping or integrated data from a variety of echo-sounders may be viewed and
manipulated within the system. Here, we describe the main features of the system and illustrate how it may be used to mark,
transform, analyse, and compare dual-frequency acoustic data.
There are particular difficulties in making acoustic estimates of the abundance of demersal and semi-demersal fish. One possibility which exists in any survey situation is that the fish may move from the direct path of the vessel because of the noise it is radiating. However, the problems addressed here are primarily due to the physical characteristics of the transmitted acoustic pulse from the echo-sounder and its interaction with fish close to the seabed. This paper looks at the factors controlling the detection of these fish in terms of the acoustic sampling volumes near the bottom, the discrimination theoretically possible between fish and seabed echoes and the ''depth anomaly''. The acoustic deadzone is defined and its volume is determined. Practical aspects of signal processing in this near seabed situation are then described, including seabed recognition and safeguarding fish signals from contamination by the bottom echo and from noise. Next, an echo-integrator deadzone comprising the acoustic deadzone, the backstep zone, and the partial integration zone (related to pulse length) is described and defined. Equations for calculating the effective volume or effective height of this deadzone are developed. Estimation errors due to the echo-integrator deadzone are investigated and equations derived for the necessary corrections. An example is shown of partial failure of the bottom recognition system and how the echo integrator result can be corrected to compensate. (C) 1996 International Council for the Exploration of the Sea
The time varied gain (TVG) of a sonar is intended to remove the range dependence of echo strength. The conventional “40 log R” and “20 log R” TVG functions, which apply to single and distributed targets respectively, provide exact compensation only at infinite range. At short range, the conventional functions are inexact due to bandwidth related delays and the change in receiver gain over a pulse length. The theory of echo formation is used to derive exact gain functions which make the echo energy integral independent of the target range. In the case of randomly distributed targets, the linear form of the exact function is shown to be (t)=ct exp (αct/2)√{(1−T1/t)2−(T2/t)2}, for sound speed c and absorption coefficient α. T1 and T2 are constants for a given sonar and target. The ct exp (αct/2) term is equivalent to “20 log R+2 αR”. The single target function is similarly the conventional function multiplied by a polynomial expression in 1/t. Analytic functions are derived for systems with simple transfer functions. As the pulse length bandwidth product increases, the exact function tends to that of the wideband ideal system for which T1=T/2 and T2 = T/√(12), T being the transmitter pulse length. Exact TVG functions are derived numerically for two echo sounders used in fishery research and are compared with the measured gain variation. The TVG function realized in sonars may depart considerably from the exact form. Delaying the start of the TVG ramp may reduce the error. The delay required for exact compensation depends upon the target range and is at least half the pulse length.
The key idea behind active learning is that a machine learning algorithm can achieve greater accuracy with fewer labeled training instances if it is allowed to choose the data from which is learns. An active learner may ask queries in the form of unlabeled instances to be labeled by an oracle (e.g., a human annotator). Active learning is well-motivated in many modern machine learning problems, where unlabeled data may be abundant but labels are difficult, time-consuming, or expensive to obtain. This report provides a general introduction to active learning and a survey of the literature. This includes a discussion of the scenarios in which queries can be formulated, and an overview of the query strategy frameworks proposed in the literature to date. An analysis of the empirical and theoretical evidence for active learning, a summary of several problem setting variants, and a discussion of related topics in machine learning research are also presented.
Guillard J, Balay P, Colon M, Brehmer P. Survey boat effect on YOY fish schools in a pre-alpine lake: evidence from multibeam sonar and split-beam echosounder data. Ecology of Freshwater Fish 2010: 19: 373–380. © 2010 John Wiley & Sons A/S
Abstract – Hydroacoustic methods are widely employed by fish scientists for assessing fish stocks. The method most often used is echosounding, beaming vertically. Nowadays the multibeam sonar, and therefore the 3-D presentation of fish schools, has yielded better knowledge of school morphology. Using the data collected simultaneously by both sonar and echosounding in a lake, we have identified boat-induced behavioural changes in small pelagic fish schools. Using high resolution sonar data, we showed that the fish schools detected under the boat have a significantly larger volume than those alongside the boat. This finding is explained according to behavioural response due to the theoretical characteristics of the boat diagram sound pressure, and the existence of a strong thermocline. Then we compared two descriptors, the height of the fish school and the backscatter energy. We found significant differences, which reveal vertical fish school compression occurring simultaneously with the horizontal and sidelong escape behaviour.
In this paper we demonstrate that finite linear combinations of compositions of a fixed, univariate function and a set of
affine functionals can uniformly approximate any continuous function ofn real variables with support in the unit hypercube; only mild conditions are imposed on the univariate function. Our results
settle an open question about representability in the class of single hidden layer neural networks. In particular, we show
that arbitrary decision regions can be arbitrarily well approximated by continuous feedforward neural networks with only a
single internal, hidden layer and any continuous sigmoidal nonlinearity. The paper discusses approximation properties of other
possible types of nonlinearities that might be implemented by artificial neural networks.
Machine learning algorithms frequently require careful tuning of model
hyperparameters, regularization terms, and optimization parameters.
Unfortunately, this tuning is often a "black art" that requires expert
experience, unwritten rules of thumb, or sometimes brute-force search. Much
more appealing is the idea of developing automatic approaches which can
optimize the performance of a given learning algorithm to the task at hand. In
this work, we consider the automatic tuning problem within the framework of
Bayesian optimization, in which a learning algorithm's generalization
performance is modeled as a sample from a Gaussian process (GP). The tractable
posterior distribution induced by the GP leads to efficient use of the
information gathered by previous experiments, enabling optimal choices about
what parameters to try next. Here we show how the effects of the Gaussian
process prior and the associated inference procedure can have a large impact on
the success or failure of Bayesian optimization. We show that thoughtful
choices can lead to results that exceed expert-level performance in tuning
machine learning algorithms. We also describe new algorithms that take into
account the variable cost (duration) of learning experiments and that can
leverage the presence of multiple cores for parallel experimentation. We show
that these proposed algorithms improve on previous automatic procedures and can
reach or surpass human expert-level optimization on a diverse set of
contemporary algorithms including latent Dirichlet allocation, structured SVMs
and convolutional neural networks.
The exploitation of offshore mussel farms is becoming important throughout the world, but monitoring this activity remains a difficult task. Here, we propose a specific method for this purpose. A total of 140 long-lines were monitored on a mussel culture ground in the French Mediterranean Sea during four experimental surveys deploying multibeam sonar devices mounted on poles (Reson Seabat 6012, 455 kHz) on small boats. This allowed geo-referenced observations to be made of the submerged mussel long-lines, as well as three-dimensional (3D) drawings of the long-line structures and the sea bed shapes, using long-line longitudinal sonar sampling. Three sonar data-analysis methods were applied: (i) direct two-dimensional (2D) visual interpretation of raw sonar video images; (ii) indirect 2D long-line drawings; and (iii) 3D digital long-line reconstructions. The development of these acoustic methods in shallow water provides scientists, managers and local authorities with a tool for observing the 3D position (geographical position and depth) of mussel cultures, for counting each structure by the ‘long-line echo-counting’ method, for monitoring their shape in situ, and for classifying the mussel rope segments into three growth categories (‘in growth’, ‘full’ and ‘empty’). The use of acoustic tools for monitoring underwater mussel culture grounds, for management purposes and for scientific studies, could be extended to other artificial structures in shallow water environments.
The Bergen echo integrator (BEI), which has evolved over many years, is a software system designed for convenient post-processing of echo-sounder data. BEI meets international standards and is essentially machine-independent. The development of BEI for the analysis of various scatterers such as zooplankton, pelagic fish, demersal fish and bottom is reviewed. The system design is described and the latest improvements are discussed. Among these improvements are the ability to quantify and remove noise, and a system that combines data at different acoustic frequencies for the generation and analysis of synthetic echograms. General operating procedures are described for the extraction of information while scrutinising and interpreting acoustical data.
Echo sounder data were used to investigate the spatio-temporal variability of shoal behaviour in the Bay of Biscay. Data collected from annual surveys were processed using MOVIES-B software in order to mcasure this variability. The software was designed to measure morphological, energetic and space-time distribution descriptors from the acoustic signal received from fish shoals. Two surveys, DAAG 90 and DAAG 91, provided the appropriate characteristics for such an analysis. The survey's objective was to obtain relative abundance indices for the anchovy (Engraulis encrasicolus) biomass in the Bay of Biscay. The surveys were carried out in the same area (southern Bay of Biscay), at the same period (April) within one year interval (1990 and 1991), using the same equipment (vessel, acoustic system, fishing gear) in a multispecies environment. Frequency distributions for every descriptor were obtained and used to describe the acoustic detection of fish shoals. The analysis of frequency distributions of space-time descriptors (year, day-hour and bottom depth) allowed the construction of derived discrete variables, which defined new subsets of detections. The subsets were then described by the continuous variables. A principal components analysis was used to describe the multidimensional data structure and to describe behaviour patterns. The size and external outline unevenness are correlated groups of shoal descriptors, but are independent of the water column shoal position and the degree of internal shoal structure. An important feature is shoal size variability between years. Significant differences in shoal characteristics were found between bathymetric zones of the same region and the pattern was similar between years. This spatial wariability is related to the distribution of different species between bathymetric zones. Although it was not possible to explain size variability between years, this will be necessary to improve shoal characterization. More knowledge about oceanographic conditions, the productivity level and availability of food, predator pressure and accurate identification of shoal speeies is required, in order to study the spatial or temporal variability in size and behaviour of shoals. Des données acoustiques ont été utilisées pour étudier la variabilité spatio-temporelle du comportement des bancs de poissons dans le golfe de Gascogne. Pour mesurer cette variabilité, les données collectées pendant deux campagnes annuelles ont été traitées à l'aide du logiciel MOVIES-B. Ce logiciel est capable de calculer des descripteurs morphologiques, énergétiques et de distribution spatio-temporelle des signaux acoustiques retro-diffusés par les bancs de poissons. Les deux campagnes, DAAG 90 et DAAG 91, présentent des caractéristiques appropriées pour ce type d'analyse. Leur objectif était d'obtenir des indices relatifs d'abondance pour la biomasse de l'anchois (Engraulis encrasicolus) dans le golfe de Gascogne. Le travail de prospection a été réalisé dans la même zone (sud du golfe de Gascogne), pendant la même période de l'année (avril) avec un an d'intervalle (1990 et 1991), en utilisant le même équipement (navire, système acoustique et engin de pêche) dans un environnement multispécifique. Les distributions de fréquence de chaque descripteur ont servi pour la description des détections acoustiques des bancs de poissons. L'analyse des distributions de fréquence des descripteurs spatio-temporels (année, heure de la journée et sonde) a permis la construction des variables discrètes qui définissent des nouveaux sous-ensembles de détections. Les sous-ensembles sont décrits par les variables continues. Une analyse en composantes principales, réalisée sur les données, a permis de décrire la structure multidimensionnelle sous-jacente et de dégager des modèles de comportement. Les descripteurs de taille et les descripteurs de l'irrégularité du contour des détections de bancs sont positivement corrélés entre eux, mais non-corrélés avec la position du banc dans la colonne d'eau, ni avec le degré de structure interne du banc. Une importante variabilité de la taille des bancs de poissons a été observée entre les deux années. Les détections acoustiques des bancs sont sensiblement différentes parmi les trois zones bathymétriques dans la même région, caractéristique qui se maintient d'une année sur l'autre. La variabilité spatiale est liée à la distribution des espèces en fonction de la bathymétrie. Malgré l'impossibilité d'expliquer la variabilité de la taille des bancs entre ces deux campagnes, identifier les causes de cette variabilité sera nécessaire pour améliorer la qualité de la caractérisation des bancs de poissons. Pour étudier la variabilité spatiale ou temporelle de la taille et du comportement des bancs de poissons, une meilleure connaissance des conditions océanographiques, du niveau de productivité et de la disponibilité de nourriture, de la pression des prédateurs et une complète identification de l'espèce de poissons qui forment le banc serait nécessaire.