Introduction
Rainbow trout (Oncorhynchus mykiss) is the main freshwater species farmed at European level, both in terms of volume and market value, where it contributed 17.7% of the overall aquaculture production in 2023, largely supported by Italy (19%), Denmark (17%), and France (14%). Monoculture is the most common rainbow trout farming, and intensive systems are considered necessary in most situations to make the operation economically attractive. In this context, the production cycle can be divided in different phases, each characterized by important critical aspects that must be considered to ensure high productivity rates and high-quality standards. The knowledge behind most of these phases has been deeply consolidated over time, allowing farmers to minimize losses within the supply chain. However, the early embryonic development, from fertilization to the eyed-egg stage, still remains a critical bottleneck, since embryos are extremely sensitive to mechanical stress and handling. For that reason, developing embryos are left undisturbed in vertical flow incubators, and are only subjected to visual inspections. In this phase, the primary challenge is the high mortality caused by infections caused by water molds (mainly Saprolegnia spp.). Since oomycetes can colonize the surface of dead embryos while suffocating the surrounding living ones, a quick pathogen detection is necessary. However, if the infection is extensively spread and most embryos result infected after the visual inspection, the entire batch contained in the vertical incubator (infected and alive) is generally discarded. Within this context, the present research aimed to develop and validate a technological automatic vertical incubator prototype able to gently lift the embryos and to promptly recognize (through multi-spectral cameras) and remove the non-developing or infected ones, in light to reduce possible food losses deriving from the whole batch discard.
Materials and Methods
In a preliminary stage, 50 batches of both viable (orange) and non-viable (white) rainbow trout embryos, obtained from an Italian supplier, were analyzed with XIMEA multi-spectral cameras. Particularly, embryos from each batch were disposed on a single layer, and cameras acquisitions were performed in the visible (VIS) and near infrared (NIR) spectrum. The same procedure was conducted covering the embryos' layer with water to simulate the state of immersion. The collected images (3 images per batch with and without water) were labelled by domain experts using Labelbox platform and then employed to train a deep learning model able to detect eggs and assign label based on reflectance data. This preliminary phase was necessary to identify (spectral) bands of interest.
The prototype of the vertical incubator was developed starting from the conventional structure made of a plexiglass cylinder, implemented with a motorized screw conveyor, set at a 45° angle inside the cylinder, to lift the embryos from the bottom to the top of the incubator. At the exit of the screw conveyor, a dedicated slide mechanism was placed to accurately convey embryos from the incubation area to the classification and sorting area. The vision system, responsible for embryos' classification, was constituted by high-resolution Basler USB 3.0 RGB camera. Image acquisition was managed by a GPU-equipped computer to accelerate the inference time of the deep learning model used for detecting healthy and diseased embryos based on You Only Look Once (YOLO) model (v11). The sorting was carried out by a specialized device developed for ensure the precise removal of non-viable or infected eggs based on data acquired by the imaging system. Finally, the prototype was also integrated with a Pacific system for the continuous monitoring of water parameters (pH, temperature, dissolved oxygen, and CO2), and a water flow meter. The prototype was tested under real farming conditions at the Societ�� Agricola Troticoltura F.lli Leonardi (Saone, Italy). Eggs were manually obtained from broodstock and, immediately after the fertilization, were divided in two different batches of 600 newly fertilized embryos each (one for the prototype and one for a traditional vertical incubator). Both batches were maintained in their respective incubator system up to the achievement of the eyed-egg stage. During this period, embryos located in the prototype, daily underwent one complete passage through the screw conveyor and camera inspection for selection. The non-viable embryos were discarded by the sorting system, collected and counted. After this testing period, eyed embryos from both batches were transferred to separate conventional hatchery systems until hatching. This experimental set up was run in triplicate.
Results and Discussion
The prototype efficiently lifted the embryos without affecting their integrity and the cameras discriminated them with an overall accuracy of 94%. The RGB colour images allowed for more detailed analysis of the surface. The prototype completely automated the whole embryo selection process, reducing the need for manual intervention and increasing the selection accuracy. Furthermore, it was possible to continuously monitor both the state of the embryos and the water parameters, ensuring the maintenance of optimal environmental conditions for their development. Results from mortality rates during the early development indicated that the prototype was able to improve up to 5% the number of embryos that achieved the eyed stage, thus reducing losses and improving the whole productivity of rainbow trout farms. The integration of high-quality cameras, powerful deep learning models and the environmental monitoring made the prototype extremely effective and precise, optimizing the success of the incubation process. Finally, at the end of the experimental period, the hatching rate of the embryos reared in the prototype did not show significant differences compared to those maintained in conventional hatchery systems. However, a bottleneck remains: the immediate scalability and adoption of this technology in commercial aquaculture settings.
Funding
Research supported by the "Bringing knowledge and consensus to prevent and reduce Food LOss at the primary production stage. Understanding, measuring, training and adopting (FOLOU)" project, HORIZON-CL6-2022-FARM2FORK-01-08, DOI: 10.3030/101084106.