Ethical statement. The study confirm that all methods were carried out in accordance with relevant guidelines and regulations. All methods were approved by the KASETSART UNIVERSITY Institutional Animal Care and Use Committee (ACKU 66-FIS-005), and this study followed Arrive guidelines (https://arriveguidelines.org).
Study site, fish and system design. Data were collected at the Fish Bear Farm, a red tilapia farm in the Mae Klong River, at Tha Muang district, Kanchanaburi province, Thailand (13°58'15”N 99°34'46”E). The data were collected in 8 cages each with dimensions of 5 x 5 x 2.5 m (width x length x depth) for 1 culture cycle (approximately 4-6 months). Then, the data were divided into a separate training and validation data set (5 cages) and a test data set (3 cages). The red tilapia fish were released in each cage to obtain a density of 1,500 fish/cage (24 fish/m3).
The fish were fed by staff with a pellet feed containing 30 % of protein (SPM 042R; S.P.M. Feedmill Co., Ltd; Thailand) until they were satiated. One day before the UAV flight, 20 fish in each cage were randomly selected and weighed using a digital scale (CST-CDR-3; CST Instruments Thailand Ltd.; Bangkok, Thailand) as illustrated in Figure 1.
Measurement of water quality and wind sampling. Prior to capturing aerial photographs using the UAV, various water quality parameters were assessed, including dissolved oxygen (DO), water temperature (Temp), pH, transparency (Tran), Alkalinity (ALK) and total ammonia nitrogen (TAN). The DO and Temp levels were measured using a YSI Pro20i instrument (YSI; USA), while the pH was determined using a YSI pH100A instrument (YSI; USA). Transparency was evaluated using a 2-color disc (Secchi disc), while ALK and TAN levels were monitored in the laboratory following the guidelines outlined by the American Public Health Association28. Additionally, wind speed was recorded using an anemometer (model AM-4836; Comcube co Ltd.; Thailand) at a height not exceeding 3.00 meters above the cage, as per the device's limitations.
Unmanned aerial vehicle (UAV) and Machine vision system (MVS). The UAV or drone used in the study was the DJI Air 2S (Mavic). All adjustments of the UAV and camera were set to ‘default’ (Table 1) and the internal storage of UAV was 8 GB. The UAV was controlled by the pilot using a DJI smart controller (DJI 13 store authorized dealer, Thailand Co Ltd.; Thailand). Image acquired by the UAV were processed using the Rapidminer software (Altair Engineering Inc.; Thailand) run using an Intel (R) Core (TM) i7-9750H CPU @ 2.60GHz 2.59GHz, RAM 16 GB, 64-bit laptop workstation.
The UAV elevation above the water surface was 3.5 m, which was the lowest practical elevation that did not cause changes in fish swimming behavior when the UAV was used to capture animation in the morning and evening before feeding27 as illustrated in Figure 1.
Table 1. Specification of DJI Air 2S (Mavic).
Specification
|
Value
|
Flight Time
|
34 minutes
|
Range
|
10,000 meters
|
Transmission System
|
OcuSync 2.0
|
Weight
|
1.3 Ibs
|
Folded Size
|
7.1 x 3.8 x 3.3 inches
|
Max Speed
|
42.5 MPH
|
Obstacle Avoidance
|
3-Directions cameras and IR
|
Special Features
|
4K/60, HDR, 48 MP Photos
|
Phone Charging
|
Available
|
Takeoff and Landing Light
|
Available
|
Internal Storage
|
8 GB
|
Our Favorite Feature
|
48 MP Camera
|
In the training and validation dataset, the nine datasets were segmented into nine distinct weight ranges (9 Classes). Through the nine UAV flights, each dataset used photographs that underwent processing to achieve a size of 2 x 2 meters (adjusted relative to the cage's actual proportions), as illustrated in Figure 2. This encompassed 50 images per cage, resulting in a cumulative count of 250 images per UAV flight and a total of 2,250 images from the nine flights. Subsequently, these images were imported into the RapidMiner software for processing, with the aim of identifying the most optimal classification model.
For the testing dataset, data from the other three cages were similarly divided into the same nine weight ranges (9 Classes) utilized in the training and validation phase. The process mirrored the previous steps, maintaining a consistent number of images captured per cage as in the training and validation stage (1,350 images). This dataset was then subjected to testing, following the steps outlined in Figure 3.
1. Select operator and use Multiple Color Image Opener (MCIO) to open all images from folders.
2. Select the Edit list button of the Images parameter to assign all 9 classes, then press the assign label to assign a label.
3. Added a sub operator in MCIO, Global Feature Extraction from a Single Image (GFESI), to extract all features from each image and add sub operators in GFESI using Global statistics.
4. Processing using cross-validation:
4.1 The training and validation dataset was examined the predictive performance using the DT, RF, NB, KNN and ANN models with default configuration (Table 2). The split test (training and testing) was used in the ratio of 90:10 and k-folded cross validation was equals to 10 (Figure 3a).
4.2 The testing dataset was used to test the predictive performance of the best model from 4.1 using the same split test and k-folded validation as in 4.1, and then processing it using read excel and apply model (Figure 3b).
Table 2. Default configuration of parameters of each model.
Model
|
Configuration
|
DT
|
Criterion = accuracy, Maximal depth = 10, Apply pruning, Confidence = 0.1, Apply prepruning, Minimal gain = 0.01, Minimal leaf size = 2, Minimal size for split = 4, Number of prepruning alternatives = 3.
|
RF
|
Number of trees = 100, Criterion = accuaracy, Maximal depth = 10, Guess subset ratio, Confidence vote, Enable parallel execution.
|
NB
|
Laplace correction.
|
KNN
|
k = 5, Weighted vote, Mixed measures, Mixed Euclidean distance.
|
ANN
|
Training cycles = 200, Learning = 0.01, Momentum = 0.9, Hidden layer =1, Neuron =2, Shulffle, Normalize.
|
Performance evaluation. In this experiment, we evaluate the segmentation performance of fish postures through the assessment of accuracy, precision, recall, and F1. Accuracy quantifies the ratio of accurately identified samples to the total number of samples. A higher accuracy indicates superior model performance in discerning distinct fish postures. Precision denotes the proportion of correctly identified positive samples among all identified positive samples. Recall quantifies the ratio of correctly identified positive samples to the entirety of positive samples. The F1-Score, often refered to as the balanced score, represents the harmonic mean of precision rate and recall rate.The estimation metrics are defines as:
Where TP (true positive) signifies the count of fishes correctly identified as positive samples that are indeed positive samples; TN (true negative) denotes the number of fish accurately identified as negative samples that are actually negative samples; FP (false positive) corresponds to the number of fish erroneously identified as positive samples when they are, in fact, negative samples; FN (false negative) represents the number of fishes identified as negative samples and actually positive samples.
UAV flight permission. The UAV used in this study was a DJI Air 2S (Mavic) that had been certified for registration of radiocommunication equipment for unmanned aircraft in research, trial, and testing purposes, according to the announcement of the Office of the National Broadcasting and Telecommunications Commissions (certificate no. T040465013010), Thailand.
Ethics approval. The study confirm that all methods were carried out in accordance with relevant guidelines and regulations. All methods were approved by the KASETSART UNIVERSITY Institutional Animal Care and Use Committee (ACKU 66-FIS-005)for Good Scientific Practice.