当前位置: X-MOL 学术Fish. Res. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Object and behavior differentiation for improved automated counts of migrating river fish using imaging sonar data
Fisheries Research ( IF 2.4 ) Pub Date : 2021-01-27 , DOI: 10.1016/j.fishres.2021.105883
Jani Helminen , Tommi Linnansaari

Imaging sonars, such as the Adaptive Resolution Imaging Sonar (ARIS; Sound Metrics Corp.) produce continuous stream of sonar video footage, and they are commonly used for counting and sizing migrating fish in rivers. Although automated methods have been developed for processing imaging sonar data, manual analysis of the data is still common in fish population monitoring projects. In this study, we used Echoview software to automatically produce fish counts from long-range (up to 30 m) imaging sonar data in a prominent Atlantic salmon (Salmo salar) river; the Little Southwest Miramichi River, New Brunswick, Canada. We added postprocessing steps to address sources of error that have been reported in previous studies: 1) Major Axis Distance was used to filter out erroneous fish tracks (89 % of dynamic noise and 67 % of milling fish in the test-set) and to calculate the swimming direction (96 % correct), and 2) a logistic regression (target length, average speed, and absolute fish track change in range) was used to predict downstream moving fish from other objects with a test-set accuracy of 84 %. When 15-min tally counts were compared between computer-generated data and multiple human-generated counts, the mean of differences varied between -39 % and 65 % in the upstream counts in different datasets, and different analysis methods were in a good agreement between each other (ICC = 0.79). There were larger differences in the downstream counts where the mean of differences varied between 14 % and 115 % and there was no agreement between the datasets (ICC = 0.03). With a double-tracking method where the fish were tracked twice, the computer analysed the 24 -h datasets in 500−600 min and was slower than human-generated counts that required 200−600 min, however, computer generated-counts can be derived in the background without the presence of a technician and may produce significant savings in personnel cost.



中文翻译:

利用成像声纳数据,对对象和行为进行区分,以提高对river游鱼类的自动化计数

成像声纳,例如自适应分辨率成像声纳(ARIS; Sound Metrics Corp.),产生连续的声纳视频素材流,它们通常用于计算河流中游鱼的数量和大小。尽管已经开发了用于处理成像声纳数据的自动化方法,但是在鱼类种群监测项目中仍然经常对数据进行手动分析。在这项研究中,我们使用Echoview软件自动从著名的大西洋鲑鱼(Salmo salar)的远距离(长达30 m)成像声纳数据中产生鱼类计数)河; 加拿大新不伦瑞克省的西南小米拉米奇河。我们增加了后处理步骤,以解决先前研究中已报告的错误源:1)使用主轴距离来滤除错误的鱼迹(测试集中89%的动态噪声和67%的)鱼),以及计算游泳方向(正确率为96%),以及2)逻辑回归(目标长度,平均速度和绝对的航迹变化范围)用于以84%的测试精度预测其他物体的下游游动鱼类。当比较计算机生成的数据和多个人类生成的计数的15分钟计数时,不同数据集中上游计数的均值差异在-39%和65%之间,并且不同的分析方法之间存在很好的一致性彼此(ICC = 0.79)。下游计数差异更大,平均差异在14%和115%之间,数据集之间没有一致性(ICC = 0.03)。采用双跟踪方法,其中对鱼类进行了两次跟踪,计算机在500-600分钟内分析了24小时数据集,并且比需要200-600分钟的人为计数慢,但是可以导出计算机生成的计数在没有技术人员在场的情况下进行,可能会大大节省人员成本。

更新日期:2021-01-28
down
wechat
bug