Patent Application
20230417939
Embodiments of the disclosure will now be described by way of example with reference to the accompanying drawings, in which:
Many marine activities involve underwater sound emissions. These may be produced as a by-product of the activity (e.g. piling or explosives), or intentionally for sensing (e.g. air guns used for seismic surveys in oil and gas exploration, or military/commercial sonar). Marine mammals can be sensitive to sound underwater and this leads to concerns that they might be physically affected or their hearing might be affected if they are exposed to high levels of sound. One strategy for mitigating against these risks is to monitor for animals within a zone of influence and either delay or shut down noise producing operations if sensitive animals are detected within this zone.
One method for detecting marine mammals at sea is visual observations. However, marine mammals can be difficult to spot on the sea surface, especially when weather and light conditions are poor, and these techniques may only be viable during daylight hours.
Many marine mammals can produce loud and distinctive vocalizations, which can be used for detection in so-called Passive Acoustic Monitoring techniques. Compared with visual techniques, acoustic methods may have the advantages of: greater range, that the animal does not need to be at the surface, that the method may be less affected by weather and sighting conditions, or that animals can be detected acoustically equally well day and night, or any combination thereof.
PAM methods may rely on human operators monitoring the audio feed and/or using computerized tools to help them analyze the audio feed. However, marine mammal may vocalize over a wide range of frequencies and can extend beyond human hearing ranges. For instance, blue whales may produce infrasonic vocalizations below the lower bound of human sensitivity while harbor porpoise may produce narrow band pulses in the high ultrasonic above the higher bound of human hearing. Relying on human operators may also introduce an element of subjectivity leading to a lack of consistency and accuracy, as well as possibly being expensive in terms of the human resources required.
The current disclosure relates to a computerized system and method for marine mammal detection and to a user interface and to a method of training a model for marine mammal detection. In some embodiments, the marine mammals can be cetaceans (e.g., a species of aquatic mammals that can include whales, dolphins and porpoises).
Embodiments disclosed herein can detect marine mammal vocalizations via underwater acoustic signals. The system can include components including:
The system can include: appropriate hardware for underwater acoustic signal acquisition (e.g. detected by hydrophones on board a vessel or positioned on the seabed or fixed at any other desired position), and computer hardware to execute the machine learning software and host the user interface.
According to an aspect of the disclosure, a computer implemented method of detecting marine mammals can be provided, the method comprising any combination of the following:
The hydrophones may be onboard vessels (e.g., vessels conducting seismic surveying or participating in other acoustic noise generating activities off shore). The hydrophone array can be different from any acoustic sensors involved with the noise generating activities and can be tailored to picking up the frequencies associated with cetacean sounds.
Some prior art schemes can concentrate on narrow species of mammals, e.g., right whales, and particular vocalizations. Such schemes may only consider a very narrow frequency domain, e.g., as little as 10 Hz to 2 kHz. Some prior art schemes also may concentrate on data captured by relatively stationary offshore objects.
Prior art examples can include, e.g., “Auto-buoy”, a project undertaken by Woods Hole Oceanographic Institution (WHOI) and the Bioacoustics Research Program (BRP) at the Cornell Lab of Ornithology. In some prior art schemes a series of buoys can comprise on-board real-time systems that can listen for right whale upcalls and/or can transmit calls for expert analysis via satellite. In this environment, the background may be very calm, with the occasional anthropogenic noise. Embodiments may perform marine mammal detection over a wide range of species of cetaceans, encompassing whales, dolphins and/or porpoises, and may use data for training the models which can be incredibly noisy, with powerful seismic airgun explosions, distribution noises and vibrations from the vessel (e.g., damaged propellors), unexpected electrical interferences, etc. In many instances, biological vocalisations may be overlapped with these noises and may appear as large acoustic artefacts on the spectrograms. While embodiments may apply tonal noise reduction and may smother the model with difficult examples, the feature maps within the network can learn to identify and adequately weigh the biological artefacts.
The time-frequency image data may be a visual representation of the spectrum of frequencies of the acoustic data varying with time.
The detection may be in real time or near real time, which for example, may enable alerts to be generated and sent to an operator of onboard activities to suspend those activities, or via an automatic control signal to suspend those activities due to the detected presence of marine mammals in the vicinity of the vessel.
The method may comprise inputting the prepared input data to each of two different models, respectively arranged to detect marine mammal sounds or vocalizations in different frequency ranges corresponding respectively to different mammal sounds or vocalizations. First image data may be produced for mid frequency ranges and second image data may be produced for low frequency ranges by down sampling the acoustic data for the first image data at a mid-frequency rate and down sampling the acoustic data for the second image data at a low-frequency rate. Different Fourier transform properties may be used to transform the sample acoustic data to image data for the first image and the second image comprising at least the number of samples of acoustic data used in each Fourier transformation.
In some aspects of the disclosure, different preprocessing steps may be used to generate different image data that are appropriate for the different models/frequency ranges to give more accurate results. For example, tonal noise reduction can be applied to the image data. Thus, some embodiments may be better suited to deal with “real world” noisy data, e.g., derived from hydrophones on vessels where engine noise, acoustic gun noise, etc. is likely to be present alongside any marine mammal vocalisations. The tonal noise filter can reduce the effects of these sources of noise so that the models are better able to “see” the biological sounds in the image data to learn/predict. In contrast, in some prior art, an academic approach may only considers relatively “clean” data, e.g. from buoys, where less background noise is expected. This approach may be difficult to apply to a real-world application to lead to useful predictions.
In embodiments, the transformed first image data can be provided to the first model in successive first windows of a first time duration and the transformed second image data can be provided to the second model in successive second windows of a second time duration, wherein the first time duration is different from the second time duration.
In embodiments, the downsampled mid-frequency image data can be 0 to 24 kHz+−25% and the downsampled low-frequency image data can be 0 to 1.5 kHz+−25%.
In some embodiments, at least a first model can be a neural network iteratively trained to classify the mid frequency acoustic data on training set data comprising acoustic samples and label data indicating whether or not the sound or vocalization of a marine mammal is present in the sample. The training data can comprises positive examples of marine mammal vocalisations and negative examples of other noise, including boat noise and/or onboard operation noise, such as cavitation, boat engine noise, propeller wash bubbles, echo-sounding operations, and construction noise and any combination thereof. This can help the models learn to distinguish between the various sources of noise to which they are exposed in a typically noisy working on board environment.
In embodiments, the neural network can use one or more machine learning algorithms to learn and converge on a solution.
In an embodiment, processing the image data can include one or more of:
In an embodiment, the method can comprise splitting the input data into at least training data and test data, which can comprise training the model on the training data and/or testing the data on the test data to determine acceptable performance of the model.
In an embodiment, the neural network can comprise plural layers including at least one pooling layer before at least one convolution layer.
In an embodiment, the method can comprise further splitting the input data into validation data, wherein the validation data can be used to tune hyperparameters of the model.
In an embodiment, at least a second model can be a rule based approach operating on features extracted from the image data applied to low frequency acoustic data.
In an embodiment, the first model can be arranged to detect at least dolphin sounds and the second model can be arranged to detect at least whale sounds.
In an embodiment, processing the image data can include one or more of:
In an embodiment, the method comprising:
The rules based classifier may be a tree based classifier.
In an embodiment, the method can comprise drawing a bounding box around the acoustic artifact, wherein the plurality of features can include one or more of:
In an embodiment, the method can comprise standardizing the features using a pre-trained scaler. In an embodiment, the image data can be a power spectrogram obtained using a Short-Time Fourier Transform. In an embodiment, the output classification can distinguish between biological sources and non-biological sources. In an embodiment, the output classification can distinguish between groups of marine mammals, including whales and delphinids and/or between species.
In an embodiment, a high frequency model may be trained and used to predict the presence of vocalizations, e.g. echolocation clicks, in the audio data. Features may be extracted from the audio data suitable for predicting echolocation clicks from the audio domain or the frequency (e.g. spectrogram) domain or a combination of these. Optionally unsupervised learning may be used to cluster the training audio data into clusters representing biological and non-biological sounds, which may then be validated by an expert. The training data once validated can then be used to train a machine learning model to predict echolocation clicks in audio data based on the extracted features. The high frequency model may be in addition to or as an alternative to either or both of the low and mid frequency models.
In an embodiment, the method can comprise using a sliding window of plural time slices of audio data as input to the models. The audio samples may be fed to the pipeline in real time for real time detection. The method may comprise performing prediction pooling on plural successive windows such that plural positive detections output from the model are required for an overall positive detection. For instance, a 2 second window may be advanced in 0.5 second intervals, meaning that a noise signature possibly indicative of a biological sound present in the sample can be contained in more than one window input to the model, e.g. the signature moves across the window in successive inputs. The model may have more uncertainty in its predictions where the signature is close to the edge of the window, as context which may help the model determine whether or not is of a biological source is lost at the edges. Thus the model may predict a sound at the edge of the window is biological, but as the sample progresses in the window, the model can predict with increased confidence that the sound is not biological. By pooling multiple model outputs, e.g. aggregating the predictions, in order to arrive at an overall positive determination, false positives may be reduced For instance, where the model outputs a confidence score which is compared with a threshold, the model output may be required to be maintained above the threshold for a predetermined number of successive windows. Alternatively, an average may be taken, optionally with more weight being given to the central samples. The prediction pooling step can be thought of as an additional step in the model pipeline.
In an embodiment, the method can comprise standardization of labeling as part of the data preparation, wherein source data comprises audio files and metadata files, and the method comprising identifying samples within the audio files, associating each sample with a metadata file, and recursively extracting standardized metadata from the file to associate with the audio sample.
In an embodiment, the method can comprise recursively parsing each metadata file and matching metadata with candidate data using machine learning according to predefined rules, developing a score of how successful the current rules are in matching data, and recursively altering the rules and repeating the matching process until a predetermined threshold level of success has been met.
In an embodiment, the method can comprise automatically or in response to accepting user input, ceasing at least one on board marine activity if the prediction indicates the presence of a marine mammal. In an embodiment, the acoustic data and/or image data and prediction can be displayed to a user for validation. In an embodiment, the method can comprise receiving user input indicating validation of the prediction, wherein the user validation can override the decision to cease the marine seismic activity.
In an embodiment, the model can be implemented by a computing device on board a vessel from which the hydrophone measurements are taken and the output to a user for validation can comprise communicating the data to a remote user over a communication network. The method can further comprise receiving the validation back at the computing device for display to a user.
In an embodiment, the method can comprise adding the validated data to training data for refining the model.
In an embodiment, the method can comprise training the model by one or more of:
According to another aspect of the disclosure, a system can be provided for detection of marine mammals, the system can comprise one or more of:
According to yet another aspect of the disclosure, a user interface can be provided for a computerized system for use with the method described above in detecting marine mammals. The user interface can comprise one or more of:
According to yet another aspect of the disclosure, a method of training a model for detecting marine mammals can be provided, the method can comprise one or more of:
Embodiments of the disclosure provide a marine mammal detection system that uses advanced technologies including machine learning in particular to detect the presence of a marine mammal through acoustic events. The resulting system may be used in operations to support PAM operators and improve rate and accuracy of detection, thus improving mitigation and ensuring offshore activities do not impact the marine mammal population.
Objects of embodiments of the disclosure can include: —
Develop a system than can accurately and reliably detect the presence of a marine mammal through vocalization;
The system may be highly capable and statistically robust in differentiating between marine mammal vocals/calls, non-biological sources of noise and ambient noise;
Model(s) may learn high-level presentations that can generalize to additional species; and
Handling of recordings with multiple vocalizing mammals (e.g. Dolphin pods, mixed whales and dolphins etc.) may be done.
It will be appreciated that any features expressed herein as being provided “in one example” or “in an embodiment” or as being “preferable” may be provided in combination with any one or more other such features together with any one or more of the aspects of the present invention.
The vessel may also be engaged in onboard activities that generate acoustic noise, such as seismic surveying as described above. In such activities, the vessel may also tow a sound wave source and acoustic receivers (sometimes known as streamers) (not specifically shown). These acoustic receivers may be separate from and/or of a different type from the hydrophone array 25 used to detect marine mammals, e.g. tailored to the different sounds required to be detected by each application. In other activities that generate noise, such as piling, there may be no further acoustic receivers other than the hydrophone array 25.
The computer program may comprise a frontend where the user can opt to create a project. This project may be accompanied by a database stored in the memory where subsequently received marine mammal acoustic detections may be stored automatically. A front-end component allows the user to validate the samples and add additional, valuable metadata.
It will be appreciated however that in other embodiments the model may be executed by computing devices not local to the vessel 20, such as in the cloud 120 or a remote computer 130, e.g. using a satellite communication link to transmit data to and from the vessel. Similarly, other sensors than a hydrophone or hydrophone array may be used to detect vocalizations.
If there is a sound 360, this sound file may be transferred 365 over a suitable communication network, e.g. cloud network 120, to a remote computer 130 or cloud computer, where the on-shore PAM operator can assess the issue and make a decision.
This may be done in real time or near real time. Individual operators may receive feeds from individual vessels. Alternatively, the operator may receive feeds from plural vessels and so a single operator may validate decisions made in respect of plural vessels, leading to a significant reduction in operator-hours needed to run the system.
In the case of the model making a positive detection, this may be communicated to a remote computing device 130 or a cloud server via a communication link and validated by a remote operator 135. For instance, a two second packet of audio may be received on shore by the operator. Their on-shore system 135 may create the spectrogram from the 2 sec audio packet they have received. The PAM operator may make their own decision based on both the sound and visual data presented to them of whether there a marine mammal present, e.g. validate the decision made by the model.
Their decision may be returned 380 to computing device on the vessel via the communication network. The operation of the ship is suspended or not, e.g. ship signal/noise turned off 385 or kept on 390 depending on the decision. This might be automatic or the decision displayed to the local operator 5 who implements the necessary actions. The PAM operator decision may be given greater weight than the model decision and allowed to override it.
The remote operator 135 may also labels the audio sample, e.g. with presence or not of marine mammal and optionally type of marine mammal. Having been labelled, these assessed samples may be returned to the project database 395 to become additional data for updating the model(s) via the continuous improvement life-cycle. High-quality labeled data may be helpful to machine learning, and so iteratively improving the data may allow the model performance to quickly improve (without necessarily changing the model hyperparameters), e.g. using a data-centric approach to continually improving the model.
In the case of the model making a negative detection (False positive), the PAM operator may return their response of no marine mammal detected. The false positive sample may be maintained in the database records. The ship's operation may continue to run as normal. Optionally, the operator may assess the next n number of audio samples, to help ensure they have made an accurate decision.
In the case of the model missing a detection (False negative), then the operator may have the facility via front-end tools to capture the audio snippet, create a sample in the database and complete the necessary metadata. Again, this may be logged appropriately in the database as a false negative (missed sample) to help ensure subsequent model calibrations are focused on improving detection.
Optionally, a persistence may be set on each positive detection before the next packet can be assessed based on assessing the average length of call. For example, if the average Dolphin sound is 3 seconds, keep the ship off and send no further assessment samples until N seconds after the last call.
If there is no response from an operator, the model decision may be acted on as the Marine Mammal model has high levels of accuracy. After N seconds of no response, the model may flag the action to the ship.
Other validation checks may be implemented. For instance, the software may track: when was the last validation sample sent and received; what is the average length of time currently, between pods of Dolphins or Whales; or has there been any validation samples sent to/between the ship and PAM operator in this length of time, and if not, investigate possible reasons (for instance, is there an issue with signals and/or equipment?; or any combination thereof. Other validation checks may also be used.
If there is no signal/audio from the hydrophones in N amount of seconds, this may be immediately flagged by the operator as they could be missing marine mammal sounds.
Data Understanding and Gathering for Modelling
A challenge with using artificial intelligence to detect marine mammal vocalizations may be the disparate sources and structures of data of previous detections from which the AI is expected to learn. Accordingly, initial steps may be: understanding and preprocessing the data to standardize its form; and putting it in a suitable format to support data science and machine learning experiments, as shown by
In particular, the audio data made available for the development of the models may originate from historical seismic surveys, potentially together with metadata files generated by human operators which attempt to classify the vocalization and give additional detail.
Shared files may include:
Due to the variation of file storage techniques (between exploration projects, operators, and data storage protocols), a relatively complex task of data exploration and understanding may be conducted. This activity may aim to decipher as much data as necessary from the folder structures, and/or help ensure audio detection samples are matched with the correct metadata file, and support initial data analyses.
The relationship between detection documents and audio file may be one to many (e.g., one document can be considered metadata for many audio files).
Audio files may be variable in length, source vessel, operator, etc. amongst other mainly categorical variables. The vocalization may be present in only a small part of the audio file, or multiple vocalizations being present of the same or different types at different times in the audio file. The input files may also contain QAQC (quality assurance quality control) audio, such as other sounds that can be expected to be picked up in the operating field, such as marine diesel engines, etc. which may also be presented to models for training purposes to help it learn to distinguish between it and marine mammal vocalization.
The “metadata” associated with the audio file may also vary in that some operators may classify as whale or dolphin, and others may attempt to identify individual species of mammal. Some may use different nomenclature, e.g. dolphin, delphinid, etc. for the same thing, and many files may include human-produced error in terminology and spelling.
To address these technical problems, the data exploration and understanding stages, combined with the volume of data may show a need to automate the extraction of audio data along with respective metadata documents. Custom code constructs may be developed to recursively dive through the directories, select audio files, use file naming conventions and timestamps to link metadata, etc., as shown by
Furthermore, the metadata files (typically word processing files of various types such as .doc, .docx, .odf, .rtf) may require dynamic approaches and use of machine learning techniques to review and extract key information into a structured format. The machine learning component may be adopted to help account for inconsistencies in the document data and may rather use similarity distance measuring to match specific values to a pre-defined schema.
Where the similarity distance measure is below an acceptable threshold, then the associated value may not be accepted (since it has likely matched the incorrect schema key). Matches at or above the similarity threshold may be accepted.
The outputs of metadata may then be further reviewed by data scientists to help ensure the values are within a defined range of possible, appropriate values. At times, it may be necessary to review the documents manually and potentially remove the record of metadata and all audio samples due to high uncertainty.
This data extraction and collection process may lead to the development of a preliminary audio sample and metadata database for review and analysis as shown by steps 405,410 in
The results may take the form of a database of semi-structured data comprising entries which may include: 1) the sample identifier, 2) whether or not a marine mammal is present, 3) the category of the sound, e.g., group of marine mammal (dolphin/whale) or QAQC, 4) species of marine mammal (if known); 5) details of the state of the sea and hydrophone set up, or 6) audio sample properties; or any combination thereof.
Analysis of the audio detection samples, and respective metadata may reveal high variability in the quality of audio data, length of recordings, number of channels etc., at step 415. Additionally, there may be no indication or detail relating to the exact timestamp of a marine mammal vocalization within any audio detection sample. Some of the audio samples may be very long and may or may not contain any marine mammal vocalization.
Consequently, there may be an obvious need to investigate data, and more importantly, metadata quality.
Quality Assessment
Optionally, at this stage, the audio data can be reviewed by a bio-acoustician to provide validation labels. This might provide additional data concerning the order/species, the type of vocalization (e.g. click/whistles/moans, or combinations thereof). Additional data may be included as to a confidence rating of the classification (which can be used to dispense with marginal identifications), and precise start and end times of the vocalization within the sample.
The existing metadata and reviewed metadata can be compared at this stage to determine the confidence level in the accuracy of the data.
Data Annotation
This analysis may result in a conclusion that the data needs strong annotations (e.g., rather than an audio file; of arbitrary length, being regarded as a positive sample, the ability to calibrate models relies on samples that are focused on the vocal signal. Hence, any audio sample could contain zero or more vocalizations and thus, the start and end time (as well as frequency and assumed mammal group) may be annotated appropriately).
Higher quality annotations and representative data may provide the best opportunity to calibrate machine learning models. The annotations may form a representation of what class of objects the data belongs to and may help a machine learning model learn to identify that particular class of objects when encountered in unseen data.
In one example case, batches of 1000 audio files may be prepared and shared with a small team of experienced passive acoustic monitoring operatives.
Use of open source software (Audacity https://www.audacityteam.org/) may be used to annotate the audio files from scratch and generate annotation files.
Each annotation record may provides the start, end, min and max frequency as well as D for “Dolphin”, “W” for “Whale” and “NO” for non-biological.
Training Data Preparation
As shown by
The output of this process are thousands of small audio .wav files (both positive and negative samples) with unique global identification as the filename, further suffixed with 1 or 0 to indicate a positive or negative sample respectively. A link may be maintained between the sample name and its parent audio file, allowing for traceability and further cohort analysis based on factors such as vessel name, operator name, sea state, etc.
Training Pipeline
The pipeline is designed to closely reflect deployment scenario to minimize inconsistencies between the training code constructs and deployment code constructs. The pipeline comprises the following steps:
Read raw audio from source file
As shown by
At step 630, the spectrograms are transformed in order to help the model distinguish. Examples include cropping the image, tonal noise reduction, smoothing, signal enhancements, resizing, standardization, normalization, etc.
At step 640, the image classification algorithm is trained on the transformed visual data. For instance, a convoluted neural network model may be trained on the data to detect and/or classify sounds.
Many epochs may be set to run until the model converges. The dataset may be shuffled for every epoch to help ensure there is no undesired ordering learned, and to promote model generalization.
Model performance may be monitored both on the training data and the hold-out test set at the end of every epoch via performance metrics (loss and accuracy) which may further expose any overfitting or underfitting. Overfitting is where the model performance on the training set exceeds performance on the unseen test-set. Underfitting is where the model performance on the unseen test-set exceed that on the training set. The aim is generally to maintain a difference between the metrics of training and testing sets within an acceptable tolerance. Model checkpoints and early stopping may allow for the best model parameterization to be maintained, even if further training proceeds (which could result in a model with worse predictive performance).
As described above, the training data in this case may comprise annotated data. Supervised learning, is a type of machine learning algorithm that requires data and corresponding annotated labels to train. The typical training procedure may comprise: feeding annotated data to the machine to help the model learn, and testing the learned model on unannotated data. To find the accuracy of such a method, annotated data with hidden labels may be used in the testing stage of the algorithm. Thus, annotated data may be a necessity for training machine learning models in a supervised manner. Validation data may also be used to help tune the hyperparameters of the model. Thus, the input data set may be split into training data, validation data and test data, e.g. in the proportions 70%, 15%, 15%. Note that many other proportions may also be used.
The test dataset can change throughout training models in a cross-validation approach. It may be used for determining which model architecture performs best on average given different splits of train/test data. Validation data may be held-out entirely, at all times and it may be the final validation point.
The pipeline for using the model for detecting marine mammal vocalizations may follow a similar pipeline to that shown in
The model may be configured to just detect the presence or not of a mammal through its vocalization, e.g. a binary decision to inform the vessel operator whether to continue activities or stop. Thus, the training data set, e.g. inputs to the model used for training and testing, are simply the transformed audio samples and a binary Y/N indicating whether or not a marine mammal is present. However, in other examples, it may be useful to also classify the type of mammal detected, e.g. whales or delphinids, species, etc. by allowing the model to learn to distinguish different groups or species of mammal by training the model on data that includes the mammal group or species.
In more detail, the model may use two different pipelines for identifying marine mammal vocalizations concentrating respectively on mid frequency and low frequency vocalizations.
Mid Frequency Model
In the mid frequency model the process may proceeds as follows: —
1. 2 second audio is decimated to 24 kHz using a sample rate of 48 kHz (multichannel input)
Many AI approaches will reduce the sample rate of their input signals (called downsampling) to reduce the computational load during training time. Downsampling removes high frequency information from a signal, which may be seen as a necessary tradeoff when developing models.
2. Audio is converted from stereo to mono (single channel)
Similarly, it may be that the additional information contained in a stereo audio file can be dispensed with without significantly impacting the models ability to learn and predict. It may not be suitable for species classification.
3. Audio is Fourier transformed to produce the two-dimensional power (dB) spectrogram with the following typical properties:
Number Fourier transforms: 512
Hop length: 256 (number of audio samples between adjacent Fourier transform columns)
Window length: 512
Window type: cosine window (‘hann’)
(https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.windows.hann.html)
4. Spectrogram is resized to satisfy model input requirements without distorting the image
The two-dimensional convolutional neural network model input may require for example images with dimensions 256×256×1 (H, W, C), where the last dimension is the channel.
5. Tonal noise reduction is applied to the spectrogram. This may help the neural network distinguish the important features from the background noise.
Tonal noises may be evident in many of the audio samples and may be continuous throughout the spectrogram image. Tonal noise may be reduced or removed as a pre-processing step to avoid the model learning features corresponding to tonal noise.
6. Spectrogram is standardised to zero mean and unit variance.
The purpose of standardizing the spectrogram values is to help treat each one and their respective features fairly. Reducing the value ranges of the images may also help to increase calculation times, thus potentially leading to faster model convergence.
7. Spectrogram is issued to a deep convolutional neural network (e.g., Inception V3 architecture as the backbone model with global average pooling, dropout layer and final dense prediction layer.)
Many model architectures now exist for image classification with varying rates of success across many applications. One such model (ResNET50) was found to result in excellent performance metrics on detecting humpback calls. Azure Machine Learning Studio was used to trial various model architectures, which led to a high performing model being the Inception V3 architecture.
In production, the software issues 2 seconds of audio every 500 ms to the mid-frequency pipeline for generating a prediction. Between approx. 500 ms and 700 ms may give a good balance between speed and computational overhead.
The output of the pipeline is a floating-point value representing the prediction on a scale of 0-1 where a 0 represents “no detection” and 1 represents “detection”. The threshold may be defaulted to, e.g., 0.75, but can be altered by the user.
Low Frequency Model
In the low frequency model the process may proceed as follows:—
1. Audio is decimated to 1.5 kHz using a sample rate of 3 kHz (multichannel input)
This is effectively zooming into the lower frequency domain where ultra-low frequency whale calls are typically observed. It may also be the noisiest part of the spectrogram.
2. Audio is converted from stereo to mono (single channel)
This may be the same process as in the mid-frequency model—e.g., a mean through all samples in the audio.
3. Audio is Fourier transformed to produce the two-dimensional spectrogram. This Fourier transform algorithm parameters may differ from that of the mid-frequency and may be specific to this low frequency domain, for example.
Number Fourier transforms: 256
Window length: 256
Hop length: 8 (number of audio samples between adjacent Fourier transform columns)
Window type: cosine window (‘hann’)
These parameters (e.g., the number of Fourier transforms, hop length and window length) may be manually adjusted on samples with a marine mammal vocalization present. Adjusting these parameters affects the temporal and/or frequency resolution. The final set of parameters were found to best expose the vocalization on the low frequency spectrogram.
4. Spectrogram is resized
This is to standardise the sizes of the images without distorting them and/or to benefit from matrix/array operations which are computationally efficient.
5. Tonal noise reduction is applied to the spectrogram.
This may be the same process as in the mid-frequency model
6. Filters are applied to the spectrogram to expose acoustic artifacts
A gaussian filter is applied to help remove noise, followed by a Frangi filter (e.g., specifically for detection of continuous ridges)
7. Artifact blobs are isolated and labelled
An isodata threshold is applied to create a binary/Boolean image, followed by a labelling process algorithm to label each individual artifact in the image and calculate its area.
8. The largest artifact is extracted
Manual review of outputs at this stage lead to the conclusion that the largest artifact in the image has a much higher chance of being a marine mammal vocalisation. However, large artifacts generated by non-biological sources will may also be output. A further stage of classification may therefore be required.
9. Any combination of the following features may be generated from this artifact based, for example:
10. The features are standardised using a pre-trained scaler.
Once all features are generated for the feature, these values of the features are standardised to give all features (initially) the same weight and to help ensure features with larger magnitude do not affect model learning.
11. The scaled features are issued to a rules-based classifier for inference.
Rules-based classification models, e.g. tree based classifiers, are a type of supervised machine learning algorithm that uses a series of conditional statements to partition training data into subsets. Each successive split adds some complexity to the model, which can be used to make predictions. The end result model can be visualized as a roadmap of logical tests that describes the data set.
The entire process from raw sound input to classification may be wrapped into a data processing pipeline which serves to allow further training and inference with new data.
The classification model can be automatically derived from training data by iteratively splitting the data into separate cohorts based on its features and then measuring the purity of the leaf. If it is pure, there may be no need to keep splitting it, if it is not, then the splitting may continue until convergence is reached and there are no more features to split to achieve a better performance metric.
In production, the software may issue 1 second of audio every 500 ms to the low frequency pipeline for generating a prediction. Between approx. 500 ms and 700 ms may give a good balance between speed and computational overhead.
The output of the pipeline is a floating-point value representing the prediction on a scale of 0-1 where a 0 represents “no detection” and 1 represents “detection”. The threshold may be defaulted to e.g. 0.75.
A decision tree based approach may work well with low frequency sounds and whale vocalizations and sounds in particular, and may exhibit signs of good generalization across large test sets which have been validated by marine mammal acoustic operators as the input data may contain quite a lot of noise in the low frequencies and this technique may be resistant to noise. Low frequency images may have such low resolution features that the model could not develop weights to properly converge if a neural network technique, such as used for the mid frequency model, were used. Also, ultra low frequency whale sounds may not be so complex in shape, whereas the mid frequency sounds can be very complex, undulating sounds, which is why the low frequency approach may not work in the mid range.
High-Frequency Model
In some examples, a high frequency model may be employed as an alternative or in addition to the mid- and/or low-frequency models. Such a model may be targeted to detecting marine mammal echolocation clicks using high-frequency audio data and machine learning techniques.
The process may proceed using one or more of:
The output of the pipeline can be a floating-point value representing the prediction on a scale of 0-1, where a 0 represents “no detection” and 1 represents “detection”. The threshold may be defaulted to e.g. 0.75.
The above techniques in combination may have a high (e.g., at least an 80%) success rate in detecting marine mammal vocalizations and this may increase to 90% or more over time with sufficient high quality input data to train the models.
User Interface
It can be seen in this example, that the output probability for the low frequency detector (shown by the line in plot 502) peaks 510 above 0.75 (or whatever threshold has been set) at points coinciding (in some embodiments, the positive detection may slightly trail the vocalization) with the vocalizations 515 shown in the spectrogram 501, which may indicate that the low frequency model has made a positive detection. The mid frequency model output 503 may maintain a low probability output throughout.
The software may automatically submit audio recordings to the project database. When this happens, these samples can be accessible via the sample interface and displayed to the user as a spectrogram, as shown in
A level of persistence may be required in order for an audio recording to be submitted to the database.
As shown in the example in
It may be difficult to see the artifact, but the software may offer the user the option of interactive zooming and/or panning on the live spectrogram and/or the sample interface spectrogram using for instance the mouse scroll.
Any detections made by the models may be submitted to the database autonomously. Validation of these samples can be undertaken by the user while the system is running, or it can be completed later. The sample interface can be the software component which can allow the user to navigate the samples in the database and work through validation. The sample interface can offer the user a simple form to complete for each sample, some basic metrics to understand how many samples reside in the database and how many have been validated.
After inspecting the sample the user can click ‘submit’ so that the decisions are tagged to the sample. Alternatively, the user can choose to export the sample to a .wav file for any further analyses that may be required.
In the right hand panel, the operator can add metadata such as whether the sample is identified as a whale or dolphin and the call type. This can be used in analysis in determining the accuracy of the model for different types of call.
Results Pooling
Within a spectrogram sample, the appearance of both biological and non-biological acoustic artefacts may appear off-centre, causing ambiguity in classifying a feature as either biological or non-biological, especially when only a small section of it is visible. The model predictive performance may drop off where the sound of interest is off-centre, for example because of the loss of context of the sound of interest where it is close to an edge. This can cause false positives where a recent acoustic event appears on the extremity of the spectrogram and can be classified as a positive detection but later identified as non-biological when fully in frame. For example, boat propeller noise may initially resemble a biological sound and may be predicted as such by the model, but on further context to the sample, i.e. as the sound advances in the sliding window, its resemblance drops and the model can predict with more confidence that it is a non-biological sound. To mitigate this issue, the technique of aggregating multiple predictions on the same audio segment may be employed to generate an overall value for the given audio. This approach can leverage the fact that a given acoustic artefact will appear within multiple frames and is predicted on multiple times, thereby reducing the model's vulnerability to false positives.
Furthermore, in the event of true positives, the model can consistently predict positive on all audio samples containing biological sounds, which, when pooled together, can yield a net positive detection, thereby preserving the model's efficacy in detecting biological sounds.
The practice of pooling model results in this application can be a useful technique to ensure more accurate and reliable predictions and may be applied to some or all of the models, i.e. the low-, mid- and high-frequency models. By aggregating multiple predictions of the same audio snippet, the process can generate a probability for the period of audio that is contained in all the samples being predicted on, thus providing a more representative result. The prediction pooling function can then output an aggregated score. For instance, for each model, a moving average of the output of the model may be used, and then compared with the threshold for that model. Alternatively, the output of model may be required to stay above the threshold for a predetermined number of iterations (e.g. the persistence illustrated in
In the example shown, the model outputs predictions of 0.812 (positive), 0.314 (negative), 0.213 (negative) and 0.115 (negative) for the successive 0.5 second windows from t=0 s to t=2 s. The prediction pooling function can output a value of 0.298 based on an average with a higher weight to the middle samples, which is below the threshold and so gives a negative outcome for the period of t=2 s to t=2.5 s. Thus, the false negative prediction output by the model for the t=0 to t=0.5 s input can be suppressed by the prediction pooling technique.
Various embodiments of the present disclosure are described in terms of the example computer system of
Processor device may be a special purpose, or a general-purpose processor device specifically configured to perform the functions discussed herein. The processor device may be connected to a communications infrastructure, such as a bus, message queue, network, multi-core message-passing scheme, etc. The network may be any network suitable for performing the functions as disclosed herein and may include a local area network (LAN), a wide area network ONAN), a wireless network (e.g., WiFi), a mobile communication network, a satellite network, the Internet, fiber optic, coaxial cable, infrared, radio frequency (RF), or any combination thereof. Other suitable network types and configurations will be apparent to persons having skill in the relevant art. The computer system may also include a main memory (e. g., random access memory, read—only memory, etc.), and may also include a secondary memory. The secondary memory may include the hard disk drive and a removable storage drive, such as a floppy disk drive, a magnetic tape drive, an optical disk drive, a flash memory, etc.
The removable storage drive may read from and/or write to the removable storage unit in a well-known manner. The removable storage unit may include a removable storage media that may be read by and written to by the removable storage drive. For example, if the removable storage drive is a floppy disk drive or universal serial bus port, the removable storage unit may be a floppy disk or portable flash drive, respectively. In one embodiment, the removable storage unit may be non-transitory computer readable recording media In some embodiments, the secondary memory may include alternative means for allowing computer programs or other instructions to be loaded into the computer system, for example, the removable storage unit and an interface. Examples of such means may include a program cartridge and cartridge interface (e. g., as found in video game systems), a removable memory chip (e.g., EEPROM, PROM, etc.) and associated socket, and other removable storage units and interfaces as will be apparent to persons having skill in the relevant art. Data stored in the computer system (e. g., in the main memory and/or the secondary memory) may be stored on any type of suitable computer readable media, such as optical storage (e.g., a compact disc, digital versatile disc, Blu-ray disc, etc.) or magnetic tape storage (e. g., a hard disk drive). The data may be configured in any type of suitable database configuration, such as a relational database, a structured query language (SQL) database, a distributed database, an object database, etc. Suitable configurations and storage types will be apparent to persons having skill in the relevant art.
The computer system may also include a communications interface. The communications interface may be configured to allow software and data to be transferred between the computer system and external devices. Exemplary communications interfaces may include a modem, a network interface (e.g., an Ethernet card), a communications port, a PCMCIA slot and card, etc. Software and data transferred via the communications interface may be in the form of signals, which may be electronic, electromagnetic, optical, or other signals as will be apparent to persons having skill in the relevant art. The signals may travel via a communications path, which may be configured to carry the signals and may be implemented using wire, cable, fiber optics, a phone line, a cellular phone link, a radio frequency link, etc. The computer system may further include a display interface. The display interface may be configured to allow data to be transferred between the computer system and external display. Exemplary display interfaces may include high-definition multimedia interface (HDMI), digital visual interface (DVI), video graphics array (VGA), etc. The display may be any suitable type of display for displaying data transmitted via the display interface of the computer system, including a cathode ray tube (CRT) display, liquid crystal display (LCD), light-emitting diode (LED) display, capacitive touch display, thin-film transistor (TFT) display, etc.
Computer program medium and computer usable medium may refer to memories, such as the main memory and secondary memory, which may be memory semiconductors (e.g., DRAMs, etc.). These computer program products may be means for providing software to the computer system. Computer programs (e.g., computer control logic) may be stored in the main memory and/or the secondary memory. Computer programs may also be received via the communications interface. Such computer programs, when executed, may enable computer system to implement the present methods as discussed herein. In particular, the computer programs, when executed, may enable processor device to implement the methods, as discussed herein. Accordingly, such computer programs may represent controllers of the computer system. Where the present disclosure is implemented using software, the software may be stored in a computer program product and loaded into the computer system using the removable storage drive, interface, and hard disk drive, or communications interface.
The processor device may comprise one or more modules or engines configured to perform the functions of the computer system. Each of the modules or engines may be implemented using hardware and, in some instances, may also utilize software, such as corresponding to program code and/or programs stored in the main memory or secondary memory. In such instances, program code may be compiled by the processor device (e.g., by a compiling module or engine) prior to execution by the hardware of the computer system.
For example, the program code may be source code written in a programming language that is translated into a lower level language, such as assembly language or machine code, for execution by the processor device and/or any additional hardware components of the computer system. The process of compiling may include the use of lexical analysis, preprocessing, parsing, semantic analysis, syntax-directed translation, code generation, code optimization, and any other techniques that may be suitable for translation of program code into a lower level language suitable for controlling the computer system to perform the functions disclosed herein. It will be apparent to persons having skill in the relevant art that such processes result in the computer system being a specially configured computer system uniquely programmed to perform the functions discussed above.
Aspects of the disclosure provide a computer implemented method of detecting marine mammals, the method comprising: receiving acoustic data from one or more hydrophones; sampling the acoustic data and transforming the sampled acoustic data to time-frequency image data; processing the image data to transform the data to be suitable for input to a model; input the transformed input data to at least one model trained to detect the presence or absence of marine mammal vocalizations in the acoustic data, wherein the model automatically outputs a prediction of whether or not a mammal is present; providing output to a user indicating the prediction. This may allow for automation of the process of detecting marine mammals based on an audio feed and so provide objective, accurate predictions that require much reduced or no operator input.
In other aspects of the disclosure, inputting the prepared input data to each of two or more different models, respectively arranged to detect marine mammal sounds or vocalizations in different frequency ranges corresponding respectively to different mammal sounds or vocalizations. It has been found that different models are particularly well adapted for detecting mammal vocalizations in different frequency ranges, where different vocalizations and different types of noise may be expected. Thus, an important part of this aspect can be using different models to detect in particular whale noises in the low frequencies (e.g. including frequency ranges in the order of at least 10 Hz to 1 kHz to pick up moans and/or other typical whale vocalizations), and dolphin noises in the mid frequencies (e.g. including frequency ranges of at least 5 kHz-10 kHz). Example frequencies to use are 0 Hz-3000 Hz for the low frequency model and 0 Hz to 48,000 Hz for the mid frequency model (which in the spectrogram is 0-1500 Hz and 0-24,000 Hz respectively). High frequency models may target vocalizations >100 kHz, e.g. in the 192 kHz to 250 kHz frequency range. Different preprocessing of the data may be used for each model, such as generating spectrograms from the audio data with different parameters, etc.
In an example, at least a first model is a neural network iteratively trained to classify the mid frequency acoustic data on training set data comprising acoustic samples and label data indicating whether or not the sound or vocalization of a marine mammal is present in the sample. This model is found to be particularly effective for detecting whistles and moans of dolphins. Mid frequency sounds can be very complex, undulating sounds, leading to advantages for a neural network based approach, and which is why the low frequency approach would not work in the mid range.
In another example, a second model is a rule based approach operating on features extracted from the image data applied to low frequency acoustic data. This model has been found to be particularly effective in detecting vocalizations, e.g. moans, of whales and in particular picking out low resolution features from the noisy low frequency audio ranges. Low frequency images are such low resolution features that the mid frequency neural network model could not develop weights to properly converge.
In another example, a third model uses unsupervised learning techniques to detect echolocation signals in high frequency ranges.
Embodiments of the present disclosure have been described with particular reference to the examples illustrated. However, it will be appreciated that variations and modifications may be made to the examples described within the scope of the present claims.
It should be noted that Applicant has, for consistency reasons, used the phrase “comprising” throughout the claims instead of “including, but not limited to.” However, it should be noted that “comprising” should be interpreted as meaning “including, but not limited to.”
In addition, it should be noted that, if not already set forth explicitly in the claims, the term “a” should be interpreted as “at least one” and “the”, “said”, etc. should be interpreted as “the at least one”, “said at least one”, etc.
It is the Applicant's intent that only claims that include the express language “means for” or “step for” be interpreted under 35 U.S.C. 112(f). Claims that do not expressly include the phrase “means for” or “step for” are not to be interpreted under 35 U.S.C. 112(f).
Although any prior art made of record and not relied upon may be considered pertinent to the disclosure, none of these references anticipates or makes obvious the recited aspects of the claims. In addition, the fact that Applicant may not have specifically traversed any particular assertion by the Patent Office should not be construed as indicating Applicant's agreement therewith.
| Number | Date | Country | Kind |
|---|---|---|---|
| 23180010.3 | Jun 2023 | EP | regional |
This application is a Continuation in Part of U.S. patent application Ser. No. 17/808,406, filed Jun. 23, 2022, and claims Priority to European Application No. 23180010.3, filed Jun. 19, 2023. All of these applications are incorporated herein by reference in their entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 17808406 | Jun 2022 | US |
| Child | 18342292 | US |
