Precise Feeding For The Swimming Fish In Recirculating Aquaculture System

How to realize precise feeding for the cultured fish in the recirculating aquaculture system(RAS),is not only the challenge in production management,but the key scientific problem in welfare farming.In this study,based on tilapia,the relevant basic theories of welfare feeding were explored for the swimming fish in RAS,from the perspective of real production.By means of computer vision,image processing and deep learning,the passive feeding,active feeding and the effect light welfare during non-feeding stage which were specific for the swimming fish in RAS,were respectively studied.The main research contents and results are detailed below:Establishing two lossless,economic and effective quantification methods for the feeding activity level of fish school during the feeding in the conditions of different aquaculture environment:a)quantifying the feeding activity level of the whole school based on the fish behavior using Lucas-Kanade optical flow and entropy.b)quantifying the feeding activity level of the whole school based on the change characteristics of the water flow field caused by the fish behavior using the modified kinetic model.Results reveal that methods mentioned above both show the good quantification in the feeding activity level of tilapia under the conditions of four different digesta index of stomach and bowel(20.35 ± 10,117.28 ± 10,179.72 ± 10 and 286.33 ± 10),and the performances of these two methods are better than other methods.After that,a decision equation was proposed to determine the feeding quantity based on the feeding strategy of repeated feeding in an event.Analysis indicates that decision equation above shows strong theoretical guidance to the efficient feeding for the swimming fish in RAS.Developing a characterization method for the school appetite based on the spontaneous collective behaviors.In this method,the school appetite was quantified and characterized from the perspectives of dispersion degree,interaction force and the changing magnitude of the water flow field,respectively.Results indicate that this method has the good performance in quantifying and characterizing five typical appetites of 0.01,0.52,1.28,2.26,2.92,respectively.And the reliability and feasibility of this method were proved with a low non-match rate of 2.19 ± 0.81%.Aiming at the montoring of the emergent school behaviors representing the hunger level of fish school in RAS,two methods were proposed for the monitoring of the global and local emergent school behaviors,respectively.To the monitoring of the global emergent school behaviors,by means of the particle advection scheme and modified kinetic model,the corresponding method has the capacity in detecting the emergent gathering and scattering behaviors using the spatial distribution and behavior characteristics,without the segmentation of the foreground of fish school.For the monitoring of the local emergent school behaviors,by means of the particle advection scheme,modified motion influence map and recurrent neural networks,the corresponding method shows the capacity in monitoring three typical emergent school behaviors using the spatio-temporal motion characteristics,without the segmentation of the foreground of fish school.Results show that,compared to other algorithms,the proposed methods both perform better in the monitoring of the corresponding emergent behaviors of fish school(monitoring method for the global emergent behaviors:97.20 ± 1.23%average accuracy rate and 0.61 ± 0.08%miss-report rate of the;monitoring method for the local emergent behaviors:98.91%,91.67%and 89.89%average accuracy rates of the detection,localization and recognition,respectively.After that,the property and feasibility of the proposed methods were validated.Constructing a semi-supervised learning-based model for the live fish identification using the modified deep convolutional generative adversarial networks.This model was designed for the identity quantification of live fish during the happening of the emergent behaviors in RAS.From the actual production point of view,the model can achieve relatively good recognition using few labeled samples,with the full consideration of the limited resolution,low labeled property and the large poses of the input samples in aquaculture.Results indicate that,in contrast to the other convolutional neural networks-based live fish identification methods and the alternative generative adversarial networks-based methods,the presented method in this section shows better performance in the recognition accuracy and training speed with 80.52%,81.66%and 83.07%accuracy rates in Fish Recognition Ground-Truth dataset using 5%,10%and 15%labeled samples,respectively,and 65.13%,78.72%and 82.95%accuracy rates in Croatian Fish dataset using 25%,50%and 75%labeled samples,respectively.By means of data visualization techniques t-SNE and PCA,the high-dimensional data representing the spectrum preference of cultured fish were dimension-reduced and analyzed.Results in this section not only indicate that tilapia shows different preference of the light spectrum in the conditions of different hunger level,but provide the theoretical foundation for the light welfare of cultured fish during the non-feeding stage in RAS.