Information extraction and dynamic evaluation of soil salinization with a remote sensing method in a typical county on the Huang-Huai-Hai Plain of China

Monitoring the dynamics of soil salinization is of great importance for agricultural production. This study selected Yucheng County, a typical county on the Huang-Huai-Hai Plain (HHHP) of China, as the study area and evaluated the spatial and temporal variation of soil salinization. Three methods, consisting of principal component analysis (PCA) transformation, tasseled cap (TC) transformation, and optimal band combination (OBC), were used to extract information from an early Landsat multispectral scanner (MSS) image from 1984, and their advantages were compared. In addition, OBC was used on a thematic mapper (TM) image from 2009. An iteratively self-organizing data analysis algorithm was used together with prior knowledge of likely classifications to interpret the MSS and TM images for data classification. Finally, a transfer matrix method was used to assess the spatial and temporal variability of soil salinization and analyze the driving factors of soil salinization. Compared to PCA transformation and OBC, TC transformation was a more effective method for extracting soil salinization information from the MSS sensor. The results indicate that a soil area of approximately 298 km² was affected by salinity in 1984 in Yucheng County, of which 5.40%, 11.96%, and 12.75% were classified as being subject to slight, moderate, and severe salinization, respectively. In 2009, the saline area was reduced to only 146 km², of which 10.70% and 3.75% were characterized by slight to moderate salinization and no severe salinization, respectively. The saline land decreased at an average rate of 6 km² per year. This decrease was probably a result of lower groundwater depth, increased organic fertilizer or crop straw in soil, changed land use type, and increased vegetation coverage.     

Use of airborne multispectral imagery to discriminate and map weed infestations in a citrus orchard

Reliable information on weed abundance and distribution within fields is essential for weed management in agricultural systems. Such information is necessary to adopt localized and variable rates of herbicide spraying, thus reducing chemical waste, crop damage, and environmental pollution. This paper examined the potential of airborne multispectral imagery to discriminate and map weed infestations in an experimental citrus orchard in Japan. Using an airborne digital sensor, multispectral imagery was acquired over the study site on 10 April 2003. The obtained reflectance imagery was analyzed using an image object-based approach in eCognition. After creating image objects on the image, the spectral information for weeds and citrus, represented by corresponding selected sample image objects, was extracted. Significant differences in the spectral characteristics between weeds and citrus were observed in each of the red, green, and blue wavebands. The simple average values of these wavebands were used to classify image objects with the nearest neighbor algorithm. Maps were generated with different classes or levels of class groups. A subsequent accuracy assessment demonstrated that the weeds were successfully discriminated from other image objects with a classification accuracy of 99.07%. Therefore, maps generated based on the classification result could provide valuable information for developing a site-specific weed management program for the study orchard.     

A method for mapping the distribution of willow at a catchment scale using bi-seasonal SPOT5 imagery

This study showed that seasonal imagery acquired at specific stages of phenology can be used to improve the mapping accuracy of invasive willow at a catchment scale. SPOT5 XI (10 m) satellite imagery was acquired for early autumn and winter to represent the phenological stages of leaf cover and leaf fall respectively. Four classification regimes were evaluated using single‐ and bi‐seasonal composite imagery to determine the most accurate method. Significant spectral noise was found in willow populations, especially in the winter image, due to the effects of undergrowth exposure, shadowing, topography and boundary‐mixed pixels. Two noise reduction techniques were applied to the bi‐seasonal composite image to improve the classification results. The noise‐reduced bi‐seasonal composite image was classified using the spectral angle mapper (SAM) algorithm before importation into a geographical information system. Aerial photography was used to reduce the errors of commission associated with misclassification of pastures. The class accuracy achieved for willow using the method described in this study was 77.5% (Kappa =0.87). The high cost of eradicating willow means that managers must establish priorities for control; this technique can provide a powerful tool for prioritizing control programmes and for monitoring results at a catchment scale.     

Mapping Ridolfia segetum patches in sunflower crop using remote sensing

Ridolfia segetum is a frequent umbelliferous weed in sunflower crops in the Mediterranean basin. Field and remote sensing research was conducted in 2003 and 2004 over two naturally infested fields to determine the potential of multispectral imagery for discrimination and mapping of R. segetum patches in sunflower crops. The efficiency of the four wavebands blue (B), green (G), red (R) and near‐infrared (NIR), selected vegetation indices and the spectral angle mapper (SAM) classification method were studied using aerial photographs taken in the late vegetative (mid‐May), flowering (mid‐June) and senescence (mid‐July) crop growth stages. Discrimination efficiency of R. segetum patches in sunflower crops is consistently affected by their phenological stages, in this order: flowering > senescence > vegetative. In both fields, R. segetum patches were efficiently discriminated in mid‐June, corresponding to the flowering phase, by using the waveband G, the ratio R/B or SAM with overall accuracies ranging from 85% to 98%. The application of the median‐filtering algorithm to any of the classified images improved the accuracy. Our results suggest that mapping R. segetum weed patches in sunflower to implement site‐specific weed management techniques is feasible with aerial photography when images are taken from 8 to 10 weeks before harvesting.     

病害胁迫下玉米LAI遥感反演研究

为了在病害发生条件下进行玉米LAI的遥感估算,针对41个不同抗性的玉米自交系品种,通过人工接种方法,获得了不同病害严重程度(1～9级)的LAI数据,同时采集了地面高光谱和无人机多光谱数据,构建了K近邻算法、支持向量机、梯度提升分类树和决策分类树分类模型对病害进行分类,对玉米种质资源抗病性进行了划分。基于不同玉米病害胁迫程度分类结果,采用随机森林回归、梯度提升回归树、极端梯度增强算法、轻量梯度提升机4种机器学习模型对玉米LAI进行反演,讨论了不同模型在病害胁迫下的鲁棒性。研究结果表明,对不同生育期玉米病害程度进行划分,基于地面高光谱识别精度分别为84.72%(梯度提升分类树)、47.67%(支持向量机)、55.05%(K近邻算法)、83.02%(决策分类树)。基于病害分类结果,本文利用无人机多光谱数据估算了不同病情等级胁迫下的玉米LAI。构建了4种集成学习模型对不同病情等级的LAI进行估算,4个LAI反演模型的总体反演精度(rRMSE)分别为：19.11%(梯度提升回归树)、15.94%(轻量梯度提升机)、14.51%(随机森林回归)和15.45%(极端梯度增强算法)。其中极端梯度增强算...     

基于无人机多光谱影像和关键点检测的雪茄烟株数提取

为从无人机遥感影像中准确识别烟草,实现植株定位与计数,以雪茄烟草植株为研究对象,提出一种新的深度学习模型。区别于传统的利用检测框识别目标,本文模型利用少量的关键点学习烟草中心形态学特征,并采用轻量级的编、解码器从无人机遥感影像快速识别烟草并定位计数。首先,提出的模型针对烟草植物形态学特点,通过中心关键点标注的方法,使用高斯函数生成概率密度图,引入更多监督信息。其次,对比不同主干网络在模型中的效果,ResNet18作为主干网络时平均精度大于99.5%,精度和置信度都高于测试的其他主干网络。而MobileNetV2在CPU环境下达到运行效率最优,但平均置信度相对较低。使用损失函数Focal Loss与MSE Loss结合的Union Loss时,平均精度大于99.5%。最后,利用不同波段组合作为训练数据,对比结果发现使用红边波段更有助于模型快速收敛且能够很好地区分烟草和杂草。由于红边波段与植株冠层结构相关,使用红边、红、绿波段时平均精度达到99.6%。本文提出的深度学习模型能够准确地检测无人机遥感影像中的烟草,可为烟草的农情监测提供数据支持。     

基于纹理-颜色特征与植被指数融合的冬小麦LAI估测

准确、快速、无损估测叶面积指数(LAI)对于冬小麦生产管理具有重要意义。利用无人机搭载Prime ALTUM多光谱相机获取冬小麦拔节期、孕穗期、抽穗期、灌浆期多光谱图像,利用LAI-2200C型植物冠层分析仪获取地面LAI数据。通过Pearson相关性分析筛选出25个植被指数,并提取植被指数影像中8种纹理特征：对比度(CON)、熵(ENT)、方差(VAR)、均值(MEA)、协同性(HOM)、相异性(DIS)、二阶矩(SEM)和相关性(COR),以及3种颜色特征：一阶矩(M)、二阶矩(V)和三阶矩(S),再分别利用多元逐步回归模型(MSR)、支持向量回归模型(SVR)和高斯过程回归模型(GPR)构建冬小麦LAI估测模型。结果表明：相对于考虑单一类型变量,考虑结合纹理特征和颜色特征进行估测时模型精度更高;3类模型中,GPR模型估测冬小麦LAI的精度最高;所有模型中,基于纹理-颜色特征与植被指数融合的GPR模型估测冬小麦LAI精度最高(决定系数R²为0.94,均方根误差(RMSE)为0.17 m²/m²,平均绝对误差(MAE)...     

融合多光谱成像与深度学习的作物植株叶绿素检测系统研究

为了满足田间作物长势快速检测与指导变量管理的需求,以玉米为例设计了基于多光谱成像的田间作物植株叶绿素检测系统,包括可见光(RGB)和近红外(Near-infrared, NIR)图像采集模块、主控处理器模块、模型加速模块、显示及电源模块,用于实现玉米植株智能识别与叶绿素指标一体化检测。首先,采集玉米苗期和拔节期冠层图像数据集,比较了植株冠层实例分割与株心目标检测两种深度学习模型,构建了基于MobileDet+SSDLite(Single shot multibox detector lite)轻量化网络的玉米植株定位检测模型,实现玉米植株识别。其次,提取被识别的植株株心RGB-NIR图像,开展RGB和NIR图像匹配与分割,提取R、G、B和NIR灰度值计算植被指数,使用SPXY算法(Sample set portioning based on joint X-Y distances)和连续投影算法(Successive projections algorithm, SPA)分别对数据集进行样本划分及特征变量筛选,选择高斯过程回归(Gaussian process regression, ...     

基于无人机多光谱影像的荒漠草原典型物种识别

【目的】 准确获取草原植物物种空间分布信息是草原生态系统生物多样性监测、群落重构与生态功能维持的重要基础。及时准确获取植物物种空间分布可以为草原植物物种信息提取提供有效技术手段。【方法】 文章以无人机多光谱影像为基础,分别在像元尺度和对象尺度上开展了荒漠草原典型物种的信息提取方法研究。像元尺度上先定义样本计算样本可分离性,在选择不同分类器进行分类。而对象尺度上首先进行遥感影像尺度分割研究,选出最佳分割尺度。在此基础上,提取最优特征变量,并采用阈值分类法提取植被信息。【结果】 高分辨率无人机多光谱数据能够为荒漠草原物种信息提取提供有效数据基础。面向对象影像分析技术的表现最好,总体精度85.16%,Kappa系数0.71,其中短花针茅的制图精度和用户精度分别为97.6%和86%;其次是支持向量机机器学习算法,其总体精度80.40%,Kappa系数0.70,短花针茅的制图精度和用户精度分别为90.08%和76.46%;而传统最大似然分类法的识别精度较低,总体精度为74.68%,Kappa系数0.64,短花针茅的制图精度和用户精度分别为72.40和81.96。【结论】 无人机多光谱数据对于集中连片分布的植被物种的识别能力较强,而对零星分布的物种的识别精度并不理想,但该文结果能够为大尺度草原植物物种识别提供一定参考。     

基于无人机多时相植被指数的冬小麦产量估测

通过无人机搭载多光谱相机,对不同水分亏缺条件下冬小麦多个生育期进行遥感监测,采用不同种类多光谱植被指数表征冬小麦的生长特征,分析了植被指数与冬小麦产量的相关关系,并利用多时相植被指数构建产量估测数据集,采用偏最小二乘回归、支持向量机回归和随机森林回归3种机器学习算法进行冬小麦产量估测。结果表明,随着冬小麦的生长,多个植被指数与产量的相关性不断增强,灌浆末期相关系数达到0.7,植被指数与产量的线性回归决定系数也达到最大。多时相植被指数反映了冬小麦生长的变化特征,进一步提高了冬小麦产量估测精度,采用开花期和灌浆初期的多时相植被指数进行估产比采用单个生育期的植被指数估测产量的精度高,采用偏最小二乘回归模型的估测精度R²提高约0.021,支持向量机回归模型R²提高约0.015,随机森林回归模型R²提高约0.051。采用灌浆末期的多时相植被指数,3种模型均有较高的估测精度,偏最小二乘回归模型估测精度最高时的R²、RMSE分别为0.459、1 822.746 kg/hm²,支持向量机回归模型估测精度最高时的R²、RMSE分别为0.540、1 676.520 kg/hm²,随机森林回归模型估测精度最高时的R²、RMSE分别为0.560、1 633.896 kg/hm²,本文数据集训练的随机森林回归模型估测精度最高,且稳定性更好。