| Abstrakt: |
Manual measurement of live fish is stressful and may cause injuries or post-release mortality. Therefore, indirect measurement based on image analysis should be developed. In this study, 150 Nile tilapia samples of three different size ranges (0.5-1 g, 20-30 g, and 40-60 g·fish-1) were collected. Each fish was photographed five times from above while freely swimming, and then weighed. Data from 1,500 images (10 images of each fish were analyzed: 5 for whole body and 5 without fins and tail) were manually segmented to extract the view area (V). Based on an 80-20 split test, the data were divided into two sets: training data (120 fish; 1,200 images) and validation data (30 fish; 300 images). The results showed that fish body weight (W) fitted with V without fins and tail achieved a higher coefficient of determination (r2) than whole body. The linear regression model was chosen as the best fit for W estimation based on r2 (0.922-0.958) and several error analyses: root mean square error (RMSE; 1.02±0.86 g), mean absolute error (MAE; 0.90±0.82 g), mean absolute relative error (MARE; 4.57± 4.11%), maximum absolute error (MXAE; 1.76±1.36 g), and maximum relative error (MXRE; 0.12± 0.10%). Our results indicated that utilizing a linear model was ideal and easy to apply. Furthermore, there is no suffering or weight loss associated with this procedure, since it is not necessary to harvest the fish as with traditional methods. This suggests that the findings of this study can be utilized in a subsequent phase to estimate the weight of freely moving fish, and we also favor incorporating our results with unmanned aerial vehicles (UAVs). Furthermore, Artificial Intelligence (AI) will be employed to identify models capable of autonomous operation. [ABSTRACT FROM AUTHOR] |
