基于Deeplabv3+模型的成都平原水产养殖水体信息提取

为应用深度学习和遥感影像实现养殖水体信息的快速提取,以成都平原为研究区,以Sentinel 2A和高分6号多光谱影像为数据源,基于国产开源深度学习平台PaddlePaddle训练Deeplabv3+语义分割模型,构建遥感影像的水体语义分割模型,用于提取成都平原养殖水体信息。Deeplabv3+方法的总体精度和Kappa系数分别达到94.14%和0.88,均高于归一化差分水体指数法和最大似然监督分类法;模型对阴影和建筑物等误分为水体的抑制效果较好,而对小面积和细小线状水体信息的提取则受影像分辨率影响,效果无明显改进;成都平原2018年和2020年养殖水体面积分别为22.3 khm²和28.6 khm²,其验证区青白江区、新津县和广汉市养殖水体面积的泛化提取结果验证误差均≤±10%。该研究结果可为应用深度学习平台建立遥感影像的水体语义分割模型及提取水产养殖水体信息提供参考。     

基于Xception-CEMs神经网络的植物病害识别

随着深度学习技术与农业的密切融合,越来越多的研究将深度学习技术用于农业病虫害检测,提高农产品产量和质量。本文提出一种新颖的基于Xception模型的植物病害识别方法。了解到植物病害图像会受到不确定环境因素的干扰而减小图像信息。在Xception的基础上,提出一种新的通道扩增模块,采用带有通道分配权重的多尺度深度卷积与组卷积结合,增强空间和通道的特征提取效率;在网络中采用通道扩张-保持-再扩张-压缩的新策略,进一步优化通道特征提取;引入密集连接方式,提高在同尺寸的特征图之间特征重用。试验数据集由10种不同植物的50类图像组成,分别包括10种健康植物和27种病害,其中对13种病害进行了两种程度的分类。本文的方法在这些类别上可以获得91.9%的准确率,88.7%的精确率,82.45%的召回率以及85.33%的F1值。本文的算法有更小的模型复杂度和参数量,计算量为29.33 M,为Xception的66.4%。参数量为14.05 M,为Xception的66.9%。因此,Xception-CEMs能够有效对病虫害进行识别,有利于农业智能化发展。     

基于小样本学习的植物病害识别研究*

为在仅有少量训练样本条件下获得较高的植物病害分类精度,采用小样本学习模型作为病害分类器,在匹配网络、原型网络和关系网络3种典型小样本学习算法框架下分别采用Conv4、Conv6、ResNet10、ResNet18和ResNet34 5种浅层网络作为特征提取网络,在PlantVillage植物病害数据集上对病害识别性能进行对比试验。在1shot条件下,匹配网络、原型网络和关系网络对植物叶片病害识别的平均准确率分别为72.29%、72.43%和69.45%;其中原型网络+ResNet34为表现最好的组合,病害识别准确率达到了77.60%。在5shot条件下,匹配网络、原型网络和关系网络平均准确率分别为87.11%、87.50%和82.92%,各种网络病害识别准确率比1shot条件均有明显提升;原型网络+ResNet34依旧是表现最佳的组合方式,识别准确率达到89.66%。上述试验结果表明,通过优选小样本学习框架和特征提取网络的组合方式,对于少量样本的病害也能取得较好的识别效果。     

关于农村基层学习型党组织建设现状、问题的调查及对策思考

在农村基层创建学习型党组织,对于引导农村党员干部开拓发展思路,提升新农村建设能力,巩固党在农村的执政基础,加快农村小康社会的建设进程具有重要意义。但是,农村地区由于学习资源有限,人才短缺等原因,在农村基层创建学习型党组织还存在诸多问题,也面临很多困难。本文以调查为基础,分析了当前农村基层学习型党组织建设存在认识不足、人才缺乏、学习内容单调、学习手段单一等突出问题,并从组织保障、人才支持、创新手段、健全制度等方面提出了对策。     

吉林省民营高等学校资源配置效率研究

民营高等学校是我国重要的人才培养渠道,自20世纪80年代以来,民营高等教育进入了飞速发展时期。吉林省民营高校经过多年的改革和探索已有了跨越式的发展,但如何使其从粗放型发展向内涵式发展转变,实现教育资源合理配置是关键。本文从调查研究吉林省民营高等学校资源配置的实际情况出发,采用数据包络分析方法（DEA）,评价民营高等学校资源配置效率问题,可以看出,吉林省民营高等院校在资源配置上依然存在着非效率的情况,本文据此提出相应建议。     

The importance of safety signals in animal handling and training

This review considers the importance of safety for various species of domestic animals and explains how the need for safety may motivate them to offer a variety of unwelcome responses. We argue that the value of safety to animals is often overlooked by trainers and handlers. As a result, animal owners, handlers, trainers, and veterinarians are regularly injured, and training may fail. Reinforced responses that increase the animal's perceived sense of safety but simultaneously endanger the safety of handlers or trainers may lead to the inadvertent training of dangerous responses. This review offers suggestions about how safety can be used effectively and humanely as a resource in operant training. Training calmness in the presence of a specific stimulus that is associated with safety may deserve closer attention. The ethics of creating the need for safety in domestic animals as part of a training regime are discussed. It emerges that the highly prized attribute of so-called trust in animal–trainer dyads may, at least sometimes, be a manifestation of trainers acting as safety signals. Similarly, animals said to have confidence in and regard for their handlers may value the relative safety they afford.     

The slow rise of technology: Computer vision techniques in fish population connectivity

1. Technological advancements in data collection and analysis are producing a new generation of ecological data. Among these, computer vision (CV) has received increased attention for its robust capabilities for rapidly processing large volumes of digital imagery.
2. In marine ecosystems, the study of fish connectivity provides fundamental information for assessing fisheries stocks, designing and implementing protected areas and understanding the impact of habitat loss. While the field of fish connectivity has benefited from technological advancements, the extent to which novel techniques, such as CV, have been utilized has not been assessed. To inform future directions and developments, this study reviewed the current use of CV in fish connectivity research, quantified how the implementation of such technology in fish connectivity research compared with other areas of marine research and described how this field could benefit from CV.
3. The review found that the use of remote camera systems in fish connectivity research is increasing, but the implementation of automated analysis of digital imagery has been slow. Successful implementation and expansion of CV frameworks in aquaculture and coral reef ecology suggest that CV techniques could greatly benefit fish connectivity research.
4. A case study of potential use of CV in fish connectivity research, scaling up optimal foraging models to predict marine population connectivity, highlights how beneficial it could be.
5. The capacity for CV techniques to be adopted alongside traditional approaches, the unparalleled speed, accuracy and reliability of these approaches and the benefits of being able to study ecosystems along multiple spatial–temporal scales, all make CV a valuable tool for assessing connectivity. Ultimately, these technologies can assist data-driven decisions that directly influence the health and productivity of marine ecosystems.

     

Social learning in fishes: a review

Animals acquire skills and knowledge from other animals, and fish are no exception. There is now strong experimental evidence that many species of fish exhibit social learning and traditional behaviours. Here, we review the literature pertaining to social learning in fish, focusing on (i) antipredator behaviour, (ii) migration and orientation, (iii) foraging, (iv) mate choice and (v) eavesdropping.     

The role of learning in fish orientation

Past research on fish orientation and migration has tended to focus on genetically fixed patterns of behaviour. However, in most aquatic environments, the physical landscape as well as biologically important locations will be subject to varying degrees of change, favouring an ability to learn. Here, we review the evidence from field and laboratory‐based experiments that fish can and do use learning and memory to orientate within their natural environments, relying on information from a diverse array of different sources. Comparisons of different species and populations suggest that fish may be predisposed to learn specific associations at specific times or places, appropriate for the particular navigational problems that they are likely to encounter.     

基于特征迁移学习的变工况下轴向柱塞泵故障诊断

不同工况下的轴向柱塞泵故障数据存在分布差异,现有的基于特征迁移学习的变工况故障诊断方法大多只通过单个传感器信号进行分析,具有一定的局限性和片面性。为了利用多传感器信号提高变工况下轴向柱塞泵故障诊断的性能,该研究提出一种耦合分类器子空间嵌入分布自适应（Subspace Embedded Distribution Adaptation with Coupled Classifiers, SEDACC）方法。该方法利用多传感器信号的频谱数据构造主要数据集和辅助数据集,通过子空间对齐（Subspace Alignment, SA）方法将源域和目标域的主要数据投影到公共子空间中,并采用加权条件最大均值差异（Weighted Conditional Maximum Mean Discrepancy, WCMMD）作为度量进行特征分布的适配。同时,基于结构风险最小化（Structural Risk Minimization, SRM）准则在源域标签数据上学习主分类器,根据主分类器对于目标域的预测结果在目标域辅助数据上学习辅助分类器。通过交替和迭代策略不断优化分类器参数,最后对二者进行加权融合得到最终的诊断模型。通过轴向柱塞泵变工况故障诊断试验进行验证,结果表明,当以垂直于端盖的z方向振动信号为主要数据并使用声音信号（或以平行于端盖的x方向的振动信号）作为辅助数据时,SEDACC方法在6种迁移任务中的平均准确率为99.88%（99.46%）,高于其他方法。此外,所提方法在目标工况样本稀少的情况下仍具有较高的诊断精度,当目标域和源域样本数比值为0.2时,6种迁移任务的平均准确率达到92.66%。研究结果可为更完备与准确的机械故障诊断提供参考。