Final Report Summary - AFISA (Automated fish ageing)
Most fish stock assessment studies depend on age-based models, which, in case they use otoliths, imply a significant cost. The AFISA project aimed to improve the quality of growth studies and to reduce the cost of data acquisition through the provision of means to standardise ageing among laboratories and the creation of interpreted image databases which ensured the information conservation.
The specific project objectives were to:
1. develop algorithms for fish ageing automation from otolith features;
2. implement the automated modules in a software platform for otolith imaging;
3. perform a cost-benefit analysis of the proposed systems.
AFISA focused on three case studies and developed two approaches for automated estimations. The first included the development of a classifier that estimated individual ages based on intensity profiles. The second was based on models that estimated age proportion based on the otolith and fish features. A series of image processing methods specifically designed for the examined cases was utilised, combined with a nearest neighbour classification method. The final estimation of each fish age was either the result of a probabilistic evaluation, depending on its morphologic and otolith characteristics, or the outcome of a discriminant analysis which assessed the age distribution in an unknown production sample.
The project was structured in four interconnected work packages (WPs) with the following targets:
1. to collate otolith material and create databases of annotated otolith images. Initial problems related to image quality were overcome during the project elaboration and age distributions for the selected cases were successfully generated. Moreover, cost estimates for the undertaken activities were provided as part of this AFISA component.
2. to develop efficient algorithms for automated fish age assessment. The proposal included algorithms for the acquisition of image series, the extraction of image structures and the automation of ageing issues. Even though the final programmes were not as qualitative as initially planned, they provided great advancement in the available technology.
3. to demonstrate the degree of automation of the developed systems and to integrate them within existing software. The modules, firstly developed in MATLAB, were therefore translated in C++ codes. The mixture model could not be incorporated in the software; hence the conditional model was used for both age structure estimation and automated evaluation of individual ages.
4. to demonstrate the consistency of the proposed solutions, examine their accuracy and precision in comparison to experts' feedback and elaborate a cost-benefit analysis. The cost-effectiveness of alternative techniques was evaluated and an optimal setup was defined, depending on sample sizes, for the production and calibration of information. The cost-benefit analysis was undertaken in comparison to the age length key (ALK) method, which served as a reference, and utilised the mean squared error (MSE) and the relative bias (RB) as performance indicators.
The results' verification illustrated significant performance variations between the case studies, either due to difficult discrimination of the growth rings for some species or due to poor representation of specific age groups. The proposed methodology could thus not entirely substitute experts' involvement, which remained crucial for evaluation of the quality and exploitability of the results. Nevertheless, AFISA advanced significantly the potential of experts to extract automated information on fish ageing and could result in the development of an operational tool in the future.
The specific project objectives were to:
1. develop algorithms for fish ageing automation from otolith features;
2. implement the automated modules in a software platform for otolith imaging;
3. perform a cost-benefit analysis of the proposed systems.
AFISA focused on three case studies and developed two approaches for automated estimations. The first included the development of a classifier that estimated individual ages based on intensity profiles. The second was based on models that estimated age proportion based on the otolith and fish features. A series of image processing methods specifically designed for the examined cases was utilised, combined with a nearest neighbour classification method. The final estimation of each fish age was either the result of a probabilistic evaluation, depending on its morphologic and otolith characteristics, or the outcome of a discriminant analysis which assessed the age distribution in an unknown production sample.
The project was structured in four interconnected work packages (WPs) with the following targets:
1. to collate otolith material and create databases of annotated otolith images. Initial problems related to image quality were overcome during the project elaboration and age distributions for the selected cases were successfully generated. Moreover, cost estimates for the undertaken activities were provided as part of this AFISA component.
2. to develop efficient algorithms for automated fish age assessment. The proposal included algorithms for the acquisition of image series, the extraction of image structures and the automation of ageing issues. Even though the final programmes were not as qualitative as initially planned, they provided great advancement in the available technology.
3. to demonstrate the degree of automation of the developed systems and to integrate them within existing software. The modules, firstly developed in MATLAB, were therefore translated in C++ codes. The mixture model could not be incorporated in the software; hence the conditional model was used for both age structure estimation and automated evaluation of individual ages.
4. to demonstrate the consistency of the proposed solutions, examine their accuracy and precision in comparison to experts' feedback and elaborate a cost-benefit analysis. The cost-effectiveness of alternative techniques was evaluated and an optimal setup was defined, depending on sample sizes, for the production and calibration of information. The cost-benefit analysis was undertaken in comparison to the age length key (ALK) method, which served as a reference, and utilised the mean squared error (MSE) and the relative bias (RB) as performance indicators.
The results' verification illustrated significant performance variations between the case studies, either due to difficult discrimination of the growth rings for some species or due to poor representation of specific age groups. The proposed methodology could thus not entirely substitute experts' involvement, which remained crucial for evaluation of the quality and exploitability of the results. Nevertheless, AFISA advanced significantly the potential of experts to extract automated information on fish ageing and could result in the development of an operational tool in the future.
