Back to search

BIA-Brukerstyrt innovasjonsarena

AI4FISH - AI for sustainability in marine recreational fishing participation: Scalable spatio-temporally resolved ML models

Alternative title: AI4FISH - Kunstig intelligens for bærekraft i marin fritidsfiske deltakelse: Skalerbar romlig-temporal oppløst ML modeller

Awarded: NOK 4.7 mill.

Project Manager:

Project Number:

321620

Project Period:

2021 - 2023

Funding received from:

Organisation:

Our goal is to generate dynamic fishing locations to enhance the probability of users successfully catching fish. This is based on the hypothesis that unsuccessful fishing experiences at indicated spots are influenced by adverse conditions, including temperature, salinity, currents, food accessibility, predator avoidance strategies, and so on. This digital service showcases illustrative instances created through deductive reasoning, utilizing a form of knowledge referred to as tacit knowledge or know-how. This tacit knowledge, acquired through inductive reasoning based on limited personal field experiences, is then generalized through inference to predict probabilities at locations not previously explored. As a result, the digital service contains subjectivity, bias, and limitations owing to the limited scope of personal experiences and lacking of dynamics. Therefore, the project aimed to leverage machine learning (ML) for modeling tacit knowledge using positive examples and to exploit the ML model's generalization capabilities for global scalability and for speed. The primary research challenges encompass four aspects: 1) The dataset lacks non-positive examples, comprising only positive instances. 2) The manual generation of positive examples introduces contamination through subjectivity, errors, and inconsistency. Their correctness is not validated. 3) Significant data gaps exist, particularly in proximity to the shoreline, which is an area of particular interest. 4) An initial guess or prior probability regarding a location being positive (as a fishing spot) might be 1-5% or much less. This conflicts with the assumption of the classifiers most instances being positive. We provided geospatial packages comprising polygons generated through the AI pipeline. These polygons represent both static and dynamic modes. These are provided not for 1 species, but for 23 marine fish species, across marine territories, encompassing Norway, Denmark, Sweden, Finland, Germany, and the Netherlands. These outcomes are based on innovations on AI pipelines. We established AI pipelines which can generate fishing spots globally within a couple of days. Performing this task manually, as done previously, would require several years of costly labor. We developed an AI pipeline consists of algorithms for data pre-processing (Merging, harmonizing and interpolating), feature generation employing first and second-order partial derivatives, extraction of spatial patterns through multi-scale convolution operations using novel convolution kernel shapes like annulus and wedge at various spatial scales. Additionally, we introduced the idea of associating outlier fraction with subjectivity, errors, and inconsistency arising from the human annotation process to formulate the positive-only learning problem as a one-class classification problem. Ensemble based isolation forest and histogram-based classifiers outperformed other algorithms. We obtained accuracy measured by ROC-AUC ranges from 52% to 79%. From the experimental results on the three scales, we found that the generalization capabilities suffer from fundamental challenges of generalization in topographical data and the adaptiveness or global inconsistency of annotations. These AI pipelines were developed through learning from many unsuccessful investigations, which, however, provided us with a deeper understanding of the characteristics of the learning problem. Our approach involved categorizing the problem into fundamental learning paradigms. Treating spatially discretized pixels as independent samples allowed us to employ traditional ML algorithms for binary classification in tabular data (1D). Alternatively, considering the spatial dependence of neighboring pixels enabled the use of convolutional neural networks for end-to-end feature extraction and learning for either semantic segmentation or multi-object detection. To address the irregularity in ocean boundaries and prevent dilation of missing data, we employed graphs neural networks instead of convolutional neural networks (CNNs), solving the dilation issue while still leveraging spatial correlations. Stacking spatial grids in the time domain allowed us to approach the problem as a 3D or 4D learning task in both space and time. What we learned from these experiments are that to blindly treating non-annotated samples as negative examples lead to 1) difficulty to converge in optimization or 2) random predictions. We propose to continue to pursue the development of new algorithms to 1) develop novel one-class algorithms to relax the assumption that the majority being positive, i.e., addressing the conflict between train and test data: most positive in train data as opposed to most negative in test data. 2) solve generalization issue of ML models in topographical data. 3) create algorithms to generate representative negative examples enabling more powerful models such as CNNs, GNNs, and visual transformers to be built.

Fritidsfiske engasjerer minst 226 millioner deltakere globalt og genererer en årlig økonomisk verdi på 5,9 milliarder euro i EU og 29 milliarder euro i USA. Dersom 0,1% av deltakerne kjøper tjenesten, vil dette gi en årlig inntekter på over 50 millioner NOK. Skalering av tjenesten er den avgjørende faktoren, og her er AI avgjørende. Samtidig må tjenesten tilby klar verdi for brukerne, noe som betyr at brukerne opplever økt fangst ved bruk av appen sammenlignet med uten eller alternative metoder. Dette oppnås gjennom forbedret tjeneste, som inkluderer dynamiske plasser med dybdeinformasjon. I tillegg, er popularitet en viktig faktor, og dette kan ta tid å oppnå. Dette prosjektet benytter AI for å modellere den unike formen av kunnskap som kun skapes gjennom praktisk erfaring. Denne kunnskapen blir innebygd gjennom eksempler. I dette markedet, er det få seriøse aktør som konkurrerer. Fiskeridirektoratet er nokså mer interessert i sannsynligheten i større områder i stedet for å diskriminere punkter med avstand på noen hundre meter. Imidlertid kan AI-algoritmer brukes til å forutsi hvor det mest sannsynlig vil være flom eller skred, eller for å identifisere de beste stedene for å bygge vindfarm for bedrifter innen fornybar energisektor. I tillegg til økt skala på kundebasen og nye, forbedrede tjenester, er økt popularitet viktig. I dagens samfunn er popularitet viktigere enn tidligere. Å skape interessant innhold for å vise den mystiske verdenen av livet under vann var en idé som ble vurdert, men det viste seg å være utfordrende. Å skape interessant innhold for å vise den mystiske verdenen av livet under vann var en idé som ble vurdert, men det viste seg å være vanskelig. Prosjektet har levert en AI-basert metode som, ila få timer, genererer fiskplasser i stor skala, i motsetning til manuelt arbeid som ville ta flere år. Videre har prosjektet utviklet en metode for å generere dynamiske fiskeplasser og med egnet dybde. Disse funksjonene var ikke tilgjengelige i appen tidligere. Likevel står det igjen oppgaver knyttet til hvordan man kan ytterligere redusere antallet fiskeplasser eller forbedre diskrimineringskraften i algoritmen. Antallet fiskeplasser og størrelsen på disse områdene anses som en del av kvalitetsindikatoren. De bærekraftsmålene fra FN (SDG): SDG 2 (Sult) og SDG 14 (Liv under vann) er de to relevante bærekraftsmålene, som kan være i konflikt så vel som i harmoni. På samme måte kan målene for bevaring av marint mangfold og de samfunnsøkonomiske målene for rekreasjonsfiske være i konflikt, så vel som i harmoni. Dette prosjektet bidrar til etableringen av en universell plattform som kobler rekreasjonsfiskeentusiaster over hele verden. Som et resultat kan det effektivt håndtere de identifiserte utfordringene ved å tilby en løsning for dataregistrering, overvåking, kommunikasjon, påvirkning og tilkobling, og forener ulike interessenter. Dette markerer et betydelig fremskritt i dette ganske diffuse og intrikate landskapet.

Recreational fishing (RF) involves at least 226 million participants worldwide, it generates an annual economic value of 5.9 billion euros in EU and 29 billion euros in the US. Until recent years, studies show that RF has nearly invisibly contributed to approximately 12% of global fishing removal; and it has caused 27% to declining in stock of two endangered species. We have identified 5 main barriers in achieving sustainability in RF: - No or limited real-world georeferenced and time-tagged catch data. - No or limited tools for monitoring and control RF activities. - No or limited communication channels. - Limited knowledge about where to fish. - Post-reactive, not predictive or proactive. The FiskHer App could be an good tool in tackling the 5 barriers towards sustainable RF. We develop a new digital service – FiskHer.ai – leveraging state-of-the-art data science and machine learning technologies for promoting and facilitating of sustainable marine recreational fishing. The idea is to discover scalable machine learned models that accurately predict most probable RF spots with high spatial and temporal resolution in Norway and worldwide. We achieve this via synergized transformation of the domain expertise earned in field owned by FiskHer AS to machine learned models, which are incrementally learned continuously lifelong with user contributed new data. The expected results of the project are a new integrated large dataset with validated labels, new underwater 360 degree video and hyperspectral imaging data for a few fishing spots, software for data integration and automated machine learning, scalable machine learned models for accurate and spatio-temporal resolved predictions. This innovation is most important to us because: 1. it allows us to go from static to dynamic areas. 2. It gives us a wonderful tool that will save us years of manual registration. It will thus be cost-effective and innovative for us, and enable us to deliver many years ahead of schedule.

Funding scheme:

BIA-Brukerstyrt innovasjonsarena