Movements and behaviors of swordfish in the Atlantic and Pacific Oceans examined using pop‐up satellite archival tags

Swordfish are highly specialized top‐level predators that have been challenging to study. In this paper, data from 31 pop‐up satellite archival tags attached to swordfish from (i) the eastern Pacific, (ii) central Pacific, and (iii) western North Atlantic‐Caribbean were analyzed. Common across locations was a pronounced diel vertical pattern with daytime hours spent primarily below the thermocline and nighttime hours spent in warmer waters, close to the surface. One exception to this pattern was periodic daytime basking events which were most common in cooler waters off California. Maximum daytime depths were significantly correlated with light penetration as measured by the diffuse attenuation coefficient at 490 nm. Temperature did not appear to influence daytime depths, and swordfish tolerated both extremely low temperatures (4°C) and rapid and dramatic temperature changes (>20°C). Temperature did appear to influence the nighttime depths in the Pacific where fish typically remained in the surface mixed layer. In contrast, in the warm tropical Atlantic this was not the case, and nighttime depths were much deeper. In all areas, nighttime depth increased around the full moon. Given the parallels between the vertical movement patterns of swordfish and those of the deep sound scattering layer we suggest that swordfish vertical distribution patterns, especially during daytime, are influenced largely by resource availability. At night, when swordfish are typically targeted by fisheries, both ambient light and temperature influence movements. Understanding vertical movement patterns of swordfish can help evaluate gear vulnerability, improve population assessments, and potentially reduce fisheries bycatch.     

基于深度学习与多尺度特征融合的烤烟烟叶分级方法

为实现烤烟等级的快速准确识别,降低人工分级中主观因素对分级结果的影响,提高烟叶分级的准确性和一致性,提出一种基于烤烟RGB图像和深度学习的多尺度特征融合的烟叶图像等级分类方法,采用ResNet50提取烟叶图像特征,并引入基于注意力机制的SE模块(压缩激发模块),增强不同通道特征的重要程度;同时,采用FPN(特征金字塔网络)对提取的由浅及深不同层级的烟叶特征进行融合,以实现烟叶多尺度特征的表达。采集皖南地区6068个烤烟的正面和背面图像用于建模和分析。结果表明,提出的烟叶分级方法的分级正确率比经典CNN(卷积神经网络)高出5.21%,分级模型在新批次7个等级烟叶上的分级正确率为80.14%,相邻等级的分级正确率为91.50%。因此,采用RGB图像结合深度学习技术可实现烤烟烟叶等级的良好识别,可为烤烟烟叶收购等级评价提供一种新方法。     

基于轻量型残差网络的自然场景水稻害虫识别

准确识别水稻害虫对水稻及时采取防护和治理措施具有重要意义,该研究以自然场景中水稻害虫图像为研究对象,针对水稻害虫图像的颜色纹理与背景相近以及同类害虫形态差异较大等特点,设计了一个由特征提取、全局优化以及局部优化模块构成的轻量型残差网络（Light Weight Residual Network,LW-ResNet）用于水稻害虫识别。在特征提取模块通过增加卷积层数以及分支数对残差块进行改进,有效提取自然场景中水稻害虫图像的深层全局特征并使用全局优化模块进行优化;局部优化模块通过设计轻量型注意力子模块关注害虫的局部判别性特征。LW-ResNet网络在特征提取模块减少了残差块的数量,在注意力子模块中采用深度可分离卷积减少了浮点运算量,从而实现了模型的轻量化。试验结果表明,所设计的LW-ResNet网络在13类水稻害虫图像的测试数据集上达到了92.5%的识别准确率,高于VGG16、ResNet、AlexNet等经典卷积神经网络模型,并且LW-ResNet网络的参数量仅为1.62×106个,浮点运算量仅为0.34×109次,低于MobileNetV3轻量级卷积神经网络模型。该研究成果可用于移动端水稻害虫的自动识别。     

Destratification in deep fish culture reservoirs in Israel: Practice and modelling

Thermal stratification negatively affects water quality, and thus is a threatto fish culture in deep water ponds and reservoirs. To prevent and controlit, a device based on an underwater propeller that vertically transportsoxygen rich warm water from the epilimnion down to the anoxic and colderhypolimnion was assembled. The empirical results so far obtained in Israelidual purpose reservoirs for field crop irrigation and fish culture aresummarized herein, and a mathematical model of the device's effect onreservoir thermal structure is presented.     

基于ST-LSTM的植物生长发育预测模型

提早预知植物生长发育是智能育种过程的重要组成部分,针对植物表型难以精准预测和模拟的问题,利用植物生长发育的空间和时间依赖性,提出了一种基于时空长短时记忆网络（Spatiotemporal long short-term memory, ST-LSTM）的植物生长发育预测模型,实现植物生长发育的预测。首先,通过微调Mask R-CNN模型实现识别、提取植物掩模,预处理具有时空相关性的植物生长发育图像序列,构建植物生长发育预测数据集。然后,基于ST-LSTM建立植物生长发育预测模型,利用历史生长发育图像序列,融合时空深度特征,预测植物未来的生长发育图像序列。研究结果表明,所提出模型预测的图像序列与生长发育实际图像序列具有较高的一致性和相似性,首个预测时间节点的结构相似度为0.874 1,均方误差为17.10,峰值信噪比为30.83,测试集的冠层叶面积、冠幅和叶片数预测R²分别为0.961 9、0.908 7和0.915 8。该研究实现了基于植物生长发育图像序列的生长发育预测,有效减少了田间反复试验的时间、土地和人力成本,为提高智能育种效率提供了参考。     

基于自适应深度学习的数控机床运行状态预测方法

针对机床状态动态标签及差异化分布数据下的预测适应性差与准确度低问题,结合时序特征关系和模型融合方法,建立自适应混合深度学习模型进行机床状态预测.首先,通过融合最小近邻分类器,设计一种基于权值累积的自适应更新法则,建立具有数据自适应性的状态预测模型.在此基础上,提出一种基于中心损失函数的特征距离度量优化策略,构建综合决策...     

Carbon mineralization in response to nitrogen and litter addition in surface and subsoils in an agroecosystem

More than 50% of global soil organic carbon stocks are stored below 20 cm of soil depth capable of massively altering global C cycle and climate. However, subsoil C dynamics are largely overlooked implicitly assuming that surface and subsoil C dynamics are similar. Here, we compared the soil C dynamics in surface and subsurface soil layers in response to nitrogen and maize leaf litter additions. Soils, sampled from 0 to 5, 15 to 35, 35 to 55 and 55 to 75 cm depths, were incubated at 25°C after adding litter, nitrogen (NH₄NO₃) or litter plus nitrogen. Soil respiration (C mineralization) was measured throughout the incubation period. Litter addition significantly increased C mineralization in all the soil layers. However, the soil CO₂ release relative to control was more than twofold higher in 15–35 and 35–55 cm soil layers than the surface layer. Nitrogen additions significantly decreased C mineralization in 0–15 cm soil, increased in 35–55 cm and had minimal effects in the 15–35 and 55–75 cm layers. Different soil C dynamics in surface and subsurface soil layers found in our study contradict the general assumption that soil C dynamics may be treated similarly along different soil depths.     

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。     

草坪草屑的加工与利用

我国每年修剪草坪产生大量的草坪草屑,对草坪草屑的加工利用,可以对我国能源、饲料、肥料及蛋白资源的短缺起到一定的缓解作用,因此如何对草坪草屑进行合理利用已成为一项重要的研究课题。本文综述了草坪草屑能源化、饲料化、肥料化和深加工4种利用方式,详细介绍了草坪草屑在发酵生产能源、饲喂家畜、堆肥还坪以及加工叶蛋白等方面的研究进展,指出禾本科草坪草屑和豆科草坪草屑应采用不同的处理方式,对各加工方式的投入和产出也应进行进一步的量化核算,同时总结并展望了草坪草屑利用方式的发展前景。     

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...