吉林省西部新开发水田遥感动态监测研究

为了响应土地整理开发项目的政策,吉林省西部地区开发了大面积的水田。为掌握新开发水田的动态变化,有必要对吉林省地区进行土地利用分类。以吉林省前郭县为例,利用2010年环境卫星数据进行不同土地利用分类方法的比较,进而对另外3景图像进行信息提取,分析4个年份新开发水田的分布及变化情况。结果表明,支持向量机法比最大似然法的耕地分类精度高约5%,其产品和用户精度分别为95%和84%。加入纹理信息没有显著提高分类精度。2009~2012年水田面积分别增加-67.7 km2,111.7 km2和265.01 km2。     

Random point sampling to detect gain and loss in tree canopy cover in response to urban densification

Urban tree canopy cover (UTC) is a simple, and common, measure of urban forest resource. Urban infill development is likely to lead to losses in UTC under private tenure, at a time when local governments are setting ambitious targets to increase UTC overall. Simple, statistically rigorous methods are required to benchmark and track change in UTC, whilst identifying which land-use types or tenures experience change.We estimated UTC in six Melbourne suburbs in 2010 and 2015 by randomly sampling 2000 points across public land, public streetscapes and private land. We were able to detect a net change in UTC of <2% over five years to a 95% level of confidence. A significant net decrease in UTC (−2.4%) was only detected in one of the six suburbs. Two suburbs had a net increase in UTC by +2.7% over five years. On private land, there was often areas of UTC loss, but this was generally offset by canopy gain in other areas of the private realm as well as in streetscapes and public land. Losses in UTC on private land were mainly due to tree removal, with or without subsequent construction works.This study describes a simple, but statistically rigorous, method to quantify UTC change and the drivers of change in different land-use types and tenure. Despite studying two suburbs will high rates of infill development, only one suburb showed evidence of net UTC decrease. The ‘dynamic equilibrium’ in UTC, whereby canopy losses area approximately offset by concurrent canopy gain, means that ambitious targets being set by local governments to increase UTC may be difficult to achieve without changes in tree protection and infill development policy and planning.     

A computer vision system for oocyte counting using images captured by smartphone

This work proposes a computer vision procedure for counting Twospot astyanax (Astyanax bimaculatus) oocytes in Petri dishes using images captured by smartphone. First, the proposed procedure uses simple linear iterative clustering (SLIC) to divide the images into groups of pixels (superpixels). Then, based on their color and space characteristics, the images are classified into light background, dark background, dirt, or oocyte by a machine learning algorithm. Five different types of machine learning algorithms were tested: support vector machines (SVM), decision trees using the algorithm J48 and random forest, k-nearest neighbors (k-NN), and Naive Bayes. To train the algorithms, 8.578 superpixels were classified by an expert into oocyte (n = 354), dirtiness (n = 651), dark background (n = 3.622), and light background (n = 3.951). Of the five learning algorithms, SVM obtained the best result with 97% correct oocyte recognition. Given the wide availability of smartphones, we therefore conclude that the presented procedure can be a valuable tool in future experiments and studies on fertilization and hatching success in Twospot astyanax.     

Detecting interactive effects of N fertilization and heat stress on maize productivity by remote sensing techniques

The objective of this study was to compare the performance of two different remotely sensed techniques in detecting the effects of terminal heat stress and N fertilization on final maize aerial biomass (AB) and grain yield (GY). The study was conducted under field conditions for two consecutive growing seasons. Six N treatments combining three doses [0, 100, 200 Kg N ha⁻¹] and two timings [at V4 and at 15 days before silking] were applied. Within each N treatment three heat treatments were applied (pre-flowering, post-flowering and the control treatment at ambient air temperature). Remote sensing measurements were taken with a multispectral band camera to measure the normalized difference vegetation index (NDVI) and a digital Red/Green/Blue (RGB) camera to measure the normalized green red difference index (NGRDI). Both indices failed to predict the GY of pre-flowering heat-treated plants due to grain set establishment problems that could not be detected by vegetation indices which are designed to capture differences in green canopy area. In contrast, both the NGRDI and the NDVI correlated positively with GY and AB in the control heat treatment and to a lesser extent in the post-flowering heat treatment. Under the control heat treatment, the NGRDI exhibited higher correlations with AB and GY than the NDVI across the N fertilization treatments. Since the NGRDI is formulated based only on the reflectance in the visible regions (VIS) of the spectrum (Green and Red) without dependence on the near infrared regions (NIR), it performs better than the NDVI. This is because it overcame the reported saturation patterns at high leaf area index and was more efficient at capturing even small differences in leaf colour (chlorophyll content) due to the different applied N treatments. Also, the NGRDI seemed to be a more seasonally independent parameter than the NDVI, which is more affected by temporal variability within the field, and thus the NGRDI predicted AB and GY better than the NDVI when combining the data of the two growing seasons.     

基于迁移学习的无人机影像耕地信息提取方法

随着精准农业技术的发展,对农作物用地信息快速、准确提取的需求越来越高。同时,无人机技术以其方便、高效、具有低空云下飞行能力等优势被广泛应用于自然资源的调查中。但无人机影像普遍光谱信息较为匮乏,因此很难准确、快速地提取出耕地信息。基于此,提出了一种利用迁移学习机制的耕地提取方法(TLCLE)。首先,利用深度卷积神经网络(DCNN)剔除线状地物(道路、田埂等),然后,通过引入迁移学习机制将DCNN特征训练过程中得到的特征提取方法迁移到耕地提取中,最后,将所提方法与利用易康(e Cognition)软件进行耕地提取(ECLE)结果进行对比。研究结果表明:对于实验影像1、2,TLCLE方法耕地提取总体精度分别为91.9%、88.1%,ECLE方法总体精度分别为90.3%、88.3%,2种方法提取精度相当,在保证耕地地块完整、连续性上TLCLE方法优于ECLE方法。     

基于卷积神经网络的小麦产量预估方法

小麦产量是评估农业生产力的重要指标之一,针对小麦产量人工预估困难,提出将卷积神经网络运用于小麦产量预估,为农业生产力的预估提供参考,指导农业生产管理决策。利用无人机分别在河南省新乡、漯河两地进行图片采集,并以之构建麦穗数据集,分为正样本(麦穗)和负样本(叶子和背景)。针对小麦常规的生理形态和生长环境,设计卷积神经网络识别模型,以图像金字塔构建多尺度滑动窗口,以非极大值抑制(NMS)去除重叠率较高的目标框,实现对单位面积内麦穗的计数,并利用随机采样的方式对大田麦穗进行单位面积图像采样,以采样图像中麦穗数量的平均值作为产量预估基准,进一步实现麦穗产量预估。随机抽取100幅不同小麦图片进行测试,与人工计数结果进行对比,准确率达到97.30%,漏检率为0.34%,误检率为2.36%,误差率为2.70%。试验结果表明,此方法能够克服环境中的多种噪声干扰,能够在不同光照条件下对麦穗进行计数和产量预估。     

改进SSD的灵武长枣图像轻量化目标检测方法

针对加载预训练模型的传统SSD（Single Shot MultiBox Detector）模型不能更改网络结构,设备内存资源有限时便无法使用,该研究提出一种不使用预训练模型也能达到较高检测精度的灵武长枣图像轻量化目标检测方法。首先,建立灵武长枣目标检测数据集。其次,以提出的改进DenseNet网络为主干网络,并将Inception模块替换SSD模型中的前3个额外层,同时结合多级融合结构,得到改进SSD模型。然后,通过对比试验证明改进DenseNet网络和改进SSD模型的有效性。在灵武长枣数据集上的试验结果表明,不加载预训练模型的情况下,改进SSD模型的平均准确率（mAP,mean Average Precision）为96.60%,检测速度为28.05帧/s,参数量为1.99×106,比SSD模型和SSD模型（预训练）的mAP分别高出2.02个百分点和0.05个百分点,网络结构参数量比SSD模型少11.14×106,满足轻量化网络的要求。即使在不加载预训练模型的情况下,改进SSD模型也能够很好地完成灵武长枣图像的目标检测任务,研究结果也可为其他无法加载预训练模型的目标检测任务提供新方法和新思路。     

基于序列信息的土壤CT图像超分辨率重建

受部分容积效应影响,土壤计算机断层扫描（Computed Tomography,CT）图像存在孔隙边界模糊现象,影响土壤孔隙结构研究的准确性。针对该问题,该研究提出基于序列信息的生成式对抗网络（Sequence information Generative Adversarial Network,SeqGAN）,实现土壤CT图像的超分辨率重建。针对土壤CT序列图像具有较高相似性的特点,SeqGAN法引入序列卷积块挖掘前后图像的序列信息,并将多重特征增强融合于目标图像中;利用多层残差块提取图像特征,构建残差块输入和输出的直接连接,以减少模型退化;利用对抗网络实现损失间接反馈,提高模型的特征学习能力。在序列相似性较高的土壤图像数据集验证了该方法性能。结果表明,SeqGAN法均方误差比次优方法GAN降低25%,峰值信噪比提升1.4 dB,结构相似性提升0.2%。重建的土壤图像具有较高准确率和清晰度,可为后续土壤物理学研究提供准确的数据基础。     

融合光谱混合分解与面向对象的土地利用/覆被分类

错综复杂的土地利用模式和破碎的地物斑块制约了土地利用/覆被分类的精度和效率。一方面,混合像元模糊了地物的光谱信息,影响了分类精度。另一方面,如何高效利用地物的光谱、形状和纹理特征是当前土地利用/覆被分类的研究热点。为了提高基于遥感技术的土地利用/覆被分类精度,该研究基于Sentinel-2A遥感影像,开展融合光谱混合分解与面向对象的土地利用/覆被分类研究。首先,基于地物的光谱、形状和纹理特征,在3个分割尺度通过NDWI（Normalized Difference Water Index)、NDVI（Normalized Difference Vegetation Index）、SBL（Soil Background Level）等8个特征参数构建了不同地物信息的提取规则。其次,利用光谱混合分解模型提取研究区基质（SL;岩石和土壤）、植被（GV;光合作用叶片）和暗色物质（DA;阴影和水）3类通用端元。最后,尝试融合3端元光谱特征优化地物信息提取规则。研究结果表明：1）基于构建的光谱、形状和纹理的地物信息提取规则,使用模糊函数、阈值法进行土地利用/覆被分类,获得了较高的分类精度,总体精度为80.83%,Kappa系数为0.76。2）融合3端元的光谱特征的提取规则将分类精度提升至90.00%,Kappa系数提升至0.88。3）具有明确物理意义的3端元的融入增强了像元内各组分信息的差异性,弥补了传统光谱指数对植被与土壤间的亮度信息解析度不足的缺陷。该方法能充分利用影像的光谱信息,是一种由易到难、对不确定因素进行逐层剥离的土地利用/覆被信息提取技术。因此,对中高分辨率的多光谱遥感影像十分友好,在土地利用/覆被的精细化分类中有较大应用潜力。