采用双模态联合表征学习方法识别作物病害

基于深度卷积神经网络的视觉识别方法在病害诊断中表现出色,逐渐成为了研究热点。但是,基于深度卷积神经网络建立的视觉识别模型通常只利用了图像模态的数据,导致模型的识别准确率和鲁棒性,都依赖训练数据集的规模和标注的质量。构建开放环境下大规模的病害数据集并完成高质量的标注,通常需要付出巨大的经济和技术代价,限制了基于深度卷积神经网络的视觉识别方法在实际应用中的推广。该研究提出了一种基于图像与文本双模态联合表征学习的开放环境下作物病害识别模型(bimodalNet)。该模型在病害图像模态的基础上,进行了病害文本模态信息的嵌入,利用两种模态病害信息间的相关性和互补性,实现了病害特征的联合表征学习。最终bimodalNet在较小的数据集上取得了优于单纯的图像模态模型和文本模态模型的效果,最优模型组合在测试集的准确率、精确率、灵敏度、特异性和F1值分别为99.47%、98.51%、98.61%、99.68%和98.51%。该研究证明了利用病害图像和病害文本的双模态表征学习是解决开放环境下作物病害识别的有效方法。     

Projecting the suitability of global and local habitats for myrtle rust (Austropuccinia psidii) using model consensus

Myrtle rust (caused by Austropuccinia psidii) affects more than 500 known host species in the Myrtaceae family. Three different modelling approaches (CLIMEX, MaxEnt and Multi-Model Framework) were used to project the habitat suitability for myrtle rust at both global and local scales. Current data on the global occurrence of myrtle rust were collected from online literature and expert solicitation. Long-term averages of climate data (1960–1990) were sourced from WorldClim and CliMond websites. Recent reports of myrtle rust in New Zealand were used for validation of model outputs but not in model training and testing. The model outputs were combined into a consensus model to identify localities projected to be suitable for myrtle rust according to two or three models (hotspots). In addition to the locations where the pathogen is currently present, all models successfully projected independent occurrence data in New Zealand suitable for establishment of the pathogen. Climate suitability for the pathogen was primarily related to temperature followed by rainfall in MaxEnt and the CLIMEX model. The results confirmed the optimum temperature range of this pathogen in the literature (15–25 °C). Additional analysis of the precipitation variables indicated that excessive rain (more than 2000 mm in warmest quarter of the year) combined with high temperatures (>30 °C) constrain pathogen establishment. The results of the current study can be useful for countries such as New Zealand, China, South Africa and Singapore where the pathogen has not fully spread or established.     

基于生活史特征的数据有限条件下渔业资源评估方法比较

渔业资源评估是开展渔业资源管理,维系渔业可持续发展的基础工作。传统的渔业资源评估方法需要统计产量、资源丰度指数甚至年龄结构等大量数据,由于调查经费和数据的缺乏,全球仅1%的鱼种进行过系统性的资源评估。近年来,在数据有限(data-limited)条件下如何开展资源评估已日益成为学术界的关注热点。本文将基于生活史特征的评估方法分为仅需要生活史参数,需要产量数据和生活史参数,需要产量数据、生活史参数及体长或年龄数据等3大类,分别从方法、数据要求、输出结果及局限性进行了系统回顾分析,提供了关于生活史特征参数的常见估算方法,并就其中两种模型对北大西洋大青鲨(Prionace glauca)的可持续渔获量进行了初步评估与比较。最后,对数据缺乏模型的使用及模型在中国近海渔业资源评估中的运用提出了建议。     

基于SEBAL模型评估干旱半干旱区人工灌丛植被对陆表蒸散的影响

评估人工灌丛植被重建对干旱半干旱区陆地生态系统蒸散的影响,不仅能揭示植被变化与水文过程的耦合机理,又可为区域生态治理与水资源管理提供科学指导。该研究利用Landsat-8OLI/TIR遥感影像及气象数据等驱动SEBAL模型,反演宁夏盐池县的年内不同日期的陆表蒸散,结合目视解译选取的人工灌丛区与对照草地,评估了人工灌丛植被对陆表蒸散的影响。结果表明:1)SEBAL模型的蒸散反演精度与站点观测结果较为一致,可用于干旱半干旱区蒸散反演及空间特征研究;2)盐池县人工灌丛植被区日平均蒸散为1.20mm/d,高于对照草地1.17mm/d的日平均蒸散量,即干旱半干旱区人工种植灌木林增加了生态系统水分消耗,但不同季节和不同生物地理条件下的蒸散增强作用存在差异,蒸散增强在8月份最大,而3、4月份呈现负效应;3)人工灌丛的密度越大、植被盖度越高,对陆表蒸散的增强作用越强,特别在NDVI0.4的高盖度情况下蒸散增强作用更加明显。由此可知,在水资源紧缺的干旱半干旱区开展以灌木树种为主的植被重建需在合理的生态水文阈值范围内开展,才能构建出稳定可持续的人工生态系统。     

Linking small pelagic fish and cetacean distribution to model suitable habitat for coastal dolphin species,Delphinus delphis and Tursiops truncatus,in the Greek Seas (Eastern Mediterranean)

1. A large‐scale assessment of the summertime suitable habitat for Delphinus delphis (short‐beaked common dolphin) and Tursiops truncatus (common bottlenose dolphin) in Greek Seas (Eastern Mediterranean) was conducted using data from dedicated and opportunistic cetacean surveys and published data records.
2. Using a presence/absence approach, generalized additive models were applied to define a suite of environmental, bathymetric and biotic factors that best describe common and bottlenose dolphin spatial distribution, during early (May, June, July) and late (August, September) summer.
3. A geographic information system (GIS) was used to integrate sightings data with environmental characteristics, distance from the coast and sardine probability of presence. These variables were considered as good proxies for defining species‐suitable habitat within the study area's coastal environment.
4. The final selected models were used to produce annual probability maps of the presence of the species in the entire Greek Seas, as a measure of habitat suitability. Based on the mean probability and standard deviation maps for the study period GIS techniques were subsequently used to determine the persistent (areas with high mean and low variation) and occasional (high mean and high variation) habitat of each species.
5. Results showed that there was a high probability of common dolphin presence in areas with a high probability of sardine presence. For bottlenose dolphin, higher probability of the presence of species occurred in areas closer to the shore, with a high probability of sardine presence and with high concentrations of chlorophyll‐a.
6. In both seasons, the North Aegean Sea and the Inner Ionian Sea Archipelago were indicated as the most suitable areas for common dolphin distribution. Persistent habitat areas of the bottlenose dolphin included enclosed seas, continental shelf waters, and waters surrounding islands. The indicated suitable areas are discussed along with deficiencies of the models and future implications for conservation.

Copyright © 2016 John Wiley & Sons, Ltd.     

Is it possible to differentiate between environmental and fishery effects on abundance‐biomass variation? A case study of blackspot seabream (Pagellus bogaraveo) in the Strait of Gibraltar

We assessed the potential for simulation and modelling of the blackspot seabream (Pagellus bogaraveo) population in the Strait of Gibraltar to discriminate the environmental effects of fishery impacts. A discrete biomass–abundance dynamic model was implemented to obtain a simulated monthly time series of blackspot seabream biomass. On this simulated time series, autoregressive integrated moving average (ARIMA) models were fitted. The best ARIMA fit provided a significant correlation of 0.76 and persistence index higher than 0.85. The proportion of variance non‐explained by the ARIMA models was correlated with a time series of sea surface temperature (SST) and North Atlantic Oscillation (NAO). The analysis of global, annual and winter correlation between the proportion of variance not explained by the ARIMA models and environmental variables showed that significant associations were not detected over the full time series. Our analysis therefore suggests that overexploitation is the main factor responsible for the commercial depletion of blackspot seabream in the Strait of Gibraltar.     

浙江省耕地多功能价值时空变化与权衡-协同关系

耕地多功能价值的时空演变与权衡-协同关系测度对于深化耕地可持续利用和管理具有重要意义。该研究以经济快速发展地区浙江省为例,采用价值量化方法对2000、2010和2015年浙江省耕地多功能进行价值评估并分析其时空变化特征。运用Spearman秩相关系数法、双变量空间自相关模型以及可拓展随机性环境影响评估(StochasticImpactsby Regression on Population,Affluence and Technology,STIRPAT)模型探究耕地多功能之间的权衡-协同关系及影响因素。结果表明:1)2000—2015年浙江省耕地多功能总价值下降,地均耕地多功能价值呈现先下降后上升的趋势。其中气体调节、水源涵养和社会保障功能价值下降明显,食物生产和美学景观功能价值增加;耕地多功能总价值在空间上呈现北高南低的分布格局,2000—2015年大部分县市耕地多功能总价值均有不同程度的下降,西南山地丘陵区耕地多功能总价值有所提升。2)浙江省耕地多功能之间主要表现为协同关系,各项功能之间的协同-权衡关系存在空间异质性;2000—2015年,浙江省耕地多功能协同关系总体减弱,在空间上主要表现为高值协同区减少。3)城镇居民可支配收入和地均农业机械总动力对耕地多功能总价值有负面影响,农村居民可支配收入增加有助于耕地多功能总价值的提升。该研究成果可为科学划定耕地利用与保护区,促进耕地多功能的协同利用和提升耕地资源价值提供科学依据。     

无人机机载激光雷达提取果树单木树冠信息

定株管理是未来果园精准生产管理的趋势，果树单木树冠信息的提取是定株管理的关键。该研究利用无人机采集的苹果园激光探测与测量数据（Light Detection and Ranging，LiDAR）检测和测量每棵果树的树冠面积和树冠直径，并评价空间分辨率对于果树单木树冠检测与提取的影响。该方法主要包括使用反距离权重插值法间接生成冠层高度模型（Canopy Height Model，CHM）；使用局部极大值滤波算法和标记控制分水岭分割算法（Marked-Controlled Watered Segmentation，MCWS）对果树进行单木树冠检测与提取，通过与参考数据的比较，评估了该方法的精度，并定量分析了空间分辨率对于单木树冠检测与信息提取结果的敏感性。结果表明，该方法有效地实现果树单木树冠检测与信息提取，代表果树检测精度的F1得分为94.86%，树冠轮廓提取准确率为86.39%，树冠面积的提取数据集和参考数据集的线性拟合结果决定系数和归一化均方根误差分别为0.81和20.56%，树冠直径的提取数据集和参考数据集的线性拟合结果决定系数和归一化均方根误差分别为0.85和14.79%，树冠面积和直径不同程度地被高估。此外，冠层高度模型的空间分辨率接近果树平均树冠直径的1/10时精度最高，可以有效检测果树单木树冠及提取树冠轮廓，从而准确提取果树单木树冠信息。     

Alpha-shape算法构建枣树点云三维模型

为了实现枣树智能化修剪作业，该研究提出了基于点云配准的自然光照环境下的果树三维重构方法，并针对传统最近点迭代（Iterative Closest Point，ICP）算法对待配准点云的空间位置要求苛刻的问题，提出了改进的点云配准算法。首先，使用彩色深度（RGB-D）相机采集不同角度下的枣树彩色和深度图像，并通过信息融合实现相应角度下的点云获取。其次，对点云进行背景去除和滤波处理，基于直方图设定分割阈值，提取单株枣树点云，并将放置在树根附近的标靶球作为标记，使用人工标记法进行两站点云初配准。最后，在初配准基础上计算点云的曲面法向量和曲率，由曲率相近的点构成配对点对，使用k维树最近点迭代（k dimensional-tree-Iterative Closest Point，kd-tree-ICP）算法完成精配准，对点云使用Alpha-shape算法面片化，实现表面重构。利用上述方法对多棵枣树进行全局配准并完整重构果树模型。试验结果表明，通过引入初配准，有效提高了点云配准的准确性和稳定性，配准误差均控制在1.0 cm以内，平均配准误差为0.78 cm；重构模型真实感较强，在外观上更加接近真实树，枝干相对误差控制在7%以内。该研究重构模型精度较高，可为枣树智能修剪提供可视化研究基础和技术支持。     

光谱特征变量和BP神经网络构建油用牡丹种子含水率估算模型

为了进一步提高种子含水率的高光谱估算精度，该研究测定了156份油用牡丹种子的近红外吸收光谱及其对应的含水率值，分析了近红外吸收光谱、一阶微分光谱、水分吸收特征参数与含水率的相关关系，构建了基于特征波长吸收光谱、特征波长一阶微分光谱、水分特征吸收参数和BP神经网络的油用牡丹种子含水率估算模型，并对模型进行了验证；再结合一元线性回归（SLR，Single Linear Regression）、逐步多元线性回归（SMLR，Stepwise MultipleLinear Regression）、偏最小二乘回归（PLSR，Partial Least Squares Regression）模型与BP神经网络（BPNN，BP Neural Network）模型进行比较。结果表明：1）油用牡丹种子含水率的吸收光谱特征波长位于1 410、1 900、1990 nm，一阶微分光谱特征波长位于1 150、1 950、2 080 nm；2）以DF2080和AD2140为自变量建立的一元线性回归模型预测效果较优，在能够满足水分估算精度的情况下，是最优的选择方法。3）将优选的特征参数作为输入，实测含水率值作为输出，构建BP神经网络模型，其建模与验模R2分别为0.978和0.973，RMSE分别为0.22%和0.242%，而RPD值分别为6.478和5.889，与其他模型相比，BP神经网络模型的建模及预测精度均最高，是估算油用牡丹种子含水率的最优模型，其次为逐步多元线性回归模型。研究结果表明BP神经网络模型对种子含水率具有更好的预测能力，是估算油用牡丹种子含水率的有效方法。