单变量特征选择的苏北地区主要农作物遥感识别

遥感识别多源特征综合和特征优选是提高遥感影像分类精度的关键技术。农作物遥感识别中,识别特征的相对单一和数量过多均会导致作物识别精度不理想。随机森林(random forests)采用分类与回归树(CART)算法来生成分类树,结合了bagging和随机选择特征变量的优点,是一种有效的分类方法。单变量特征选择(univariate feature selection)能够对每一个待分类的特征进行测试,衡量该特征和响应变量之间的关系,根据得分舍弃不好的特征,优选得到的特征用于分类。本文基于随机森林和单变量特征选择,利用多时相光谱信息、植被指数信息、纹理信息及波段差值信息,设计多组分类实验方案,对江苏省泗洪县的高分一号(GF-1)和环境一号(HJ-1A)影像进行分类研究,旨在选择最佳的分类方案对实验区主要农作物进行识别和提取。实验结果表明:(1)多源信息综合的农作物分类精度明显高于单一的原始光谱特征分类,说明不同类型特征的引入能改善分类效果;(2)基于单变量特征选择算法的优选特征分类效果最佳,总体精度97.07%,Kappa系数0.96,表明了特征优选在降低维度的同时,也保证了较高的分类精度。随机森林和单变量特征选择结合的方法可以提高遥感影像的分类精度,为农作物的识别和提取研究提供了有效的方法。     

面向对象影像分类中基于最大化互信息的特征选择

高分辨率影像面向对象分割后产生了大量的光谱、形状以及纹理特征,如何抽取出最佳特征子集是遥感影像识别的重要问题.本文利用最大化互信息统计独立准则抽取最优特征子集,提高了面向对象遥感影像分类精度.基本过程包含以下3个方面:首先,利用eCoginition软件对高分辨遥感影像进行对象分割;然后,基于互信息最大关联、最小冗余准则(mRMR)获取优选的特征子集;最后,基于支持向量机分类器完成影像分类.以福建省漳州市QuickBird数据为例的实验表明,该方法能够有效提高遥感影像的分类精度,平均误分率降低了约4%.     

基于混合式特征选择的高分五号影像农田识别

精准农田识别是农作物估产和粮食安全评估的基础。遥感数据作为农田识别的重要数据源,可提供动态、快速的监测结果。高光谱数据在农田识别分类方面具有巨大的应用潜力,但其中的冗余波段影响了分类效率和分类精度。因此,本研究提出了一种适用于高光谱数据农田分类的混合式特征选择算法。首先,基于变量的重要性排序或约束程度,按步长逐步进行降维;其次,寻找分类精度骤减的转折点,并将其对应的变量作为特征子集;最后,利用序列后向选择SBS（Sequential Backward Selection）方法搜索最优分类特征子集。本研究利用GF-5高光谱数据,共研究了3种降维方法（随机森林RF（Random Forest）、互信息MI（Multi-Information）和L1正则化（L1 regularization））和3种分类算法（随机森林、支持向量机SVM（Support Vector Machine）和K近邻KNN（K-Nearest Neighbor））的组合在农田分类中的表现。结果表明,基于L1正则化法得到的特征子集自相关性较低,并且包含的红边和近红外波段有效提高了农田、森林和裸土的区分度。在不同分类模型比较中发现,SVM在高维空间中表现出非常好的抗噪能力,分类精度高于RF和KNN。而RF在低维空间中的泛化能力要高于SVM和KNN。相比于第一步降维得到的特征子集,使用SBS搜索得到的最优特征子集均提高了分类精度。最终,具有23维输入的L1-SVM-SBS分类模型得到了最高的总体分类精度（94.64%）和农田召回率（95.83%）。本研究为高光谱数据特征优选提供了一种新思路,筛选出了更具代表性的特征波段,提高了农田分类精度,对高光谱遥感分类研究具有参考价值。     

基于随机森林和Landsat8 OLI影像的脐橙果园种植信息提取

针对从中分辨率遥感影像中提取果园种植信息所面临的特征维数高、提取精度低等问题,采用随机森林算法进行特征优选,进而提取脐橙果园种植信息.以重庆奉节为研究区,首先选取春、秋、冬3季Landsat 8 OLI遥感数据生成光谱特征、植被指数特征和纹理特征,并结合地形因子构建初始特征集;再通过随机森林算法确定优选特征集;最后基于随机森林分类器,利用初始特征集和优选特征集进行脐橙果园的提取实验,并对各实验结果进行了精度评估和对比分析.研究结果表明,基于随机森林算法构建的优选特征集能在降低特征维度的情况下获得较高的精度,总体精度和Kappa系数分别为90.71％和0.89.     

高分六号红边特征的农作物识别与评估

红边作为植被敏感波段，其红边特征的运用是遥感识别农作物并实现精准农业的高新手段之一。以黑龙江松嫩平原北部为研究区，以国内首个提供红边波段的多光谱高分六号影像和玉米、大豆、水稻总计82859个作物样本同时作为研究对象，从以下几个方面研究了红边波段和红边指数波段等红边特征在农作物识别中的表现，并评估了农作物的识别精度。（1） 通过作物样本辐射亮度值的统计特征，初步显示了在两红边波段0.710 μm和0.750 μm处有比其他波段更好的区分；（2） 根据传统归一化植被指数形式构建了红边归一化植被指数NDVI710和NDVI750，综合两指数在J-M距离表征的作物样本类别区分度上比传统NDVI更显著；（3） 通过多种手段筛选了有效波段并且制定了支持向量机（SVM）框架下4种农作物识别的分类策略，分别在5∶5、6∶4、7∶3、8∶2、9∶1等5套随机样本分割方案下完成研究区域农作物的分类预测。在这20类分类精度中kappa系数均高于0.9609，总体精度高于0.9742；列向上5∶5分割方案的精度最高，8∶2的精度最低；横向上分类精度排序如下：SVM-RFE > SVM-RF > SVM-有红边波段 > SVM-无红边波段，该结果表明了红边指数和红边波段的参与显著地提高了作物的识别精度；（4） 由于水域等其他样本的缺少，SVM-RFE方法和SVM-RF方法的分类图像均存在少量错分现象。但从分类精度和图像细节展示上来看，SVM-RFE方法要优于SVM-RF方法，二者分类图像的交叉验证中kappa系数为0.8060，总体精度为0.8743。总之，高分六号红边特征在作物识别中表现优越，使得识别精度显著提高。后续研究者可开发更多与红边相关的植被指数，充分发挥红边特征在精准农业中的作用。     

基于Relief F与卷积神经网络的湿地植物群落精细分类

湿地是陆地生态系统和水生生态系统之间的重要过渡带,准确高效地获取湿地植物群落分布信息对于保护湿地具有深远的意义。本文以无人机多光谱影像为数据源,首先构建包含光谱特征、植被指数和纹理特征的多维特征数据集,并采用Relief F算法进行特征优选,确定最优特征数据集;然后构建基于特征优选的卷积神经网络(CNN)分类模型,对最优特征数据集进行分类,并与基于原始多光谱影像的CNN和随机森林(RF)分类方法进行对比。结果表明：(1)随着特征个数的增加,分类精度先增加后下降,当特征数为32时分类精度最高;(2)窗口为13×13的GLCM提取的信息熵和同质性等纹理特征及GNDVI、MSAVI2、RVI等多光谱植被指数重要性较高;(3)基于最优特征数据集的CNN分类模型,能够有效提取空间光谱信息,抑制“椒盐现象”的产生,分类效果最佳,总体精度达93.40%,与未进行特征优选的RF和CNN分类模型相比分别提高了9.80%和7.40%。     

通过训练样本采样处理改善小宗作物遥感识别精度

训练样本质量是决定农作物遥感识别精度的关键因素,虽然高空间分辨率卫星的发展有效地解决了农作物遥感识别过程中的混合像元问题,但是当区域内不同作物种植面积差异较大时,训练集中不同类别样本数量往往相差较大,这样的不均衡数据集影响分类器的训练,导致少数类别的识别精度不理想。为研究作物遥感识别过程中的不均衡样本问题,本文基于GF-2号卫星数据,首先挖掘了地物的光谱信息、纹理信息,用特征递归消除RFE (Recursive Feature Elimination)方法进行特征优选,然后从数据处理的角度采用了5种采样算法对不均衡训练集进行处理,最后使用采样后的均衡数据集训练分类器,对比数据采样前后决策树与Adaboost(Adaptive Boosting)两种分类器的识别结果,发现:(1)经过采样处理后两种分类算法明显提升了小宗作物的分类精度;(2)经过ADASYS (Adaptive synthetic sampling)采样处理后,分类器性能提升最多,决策树的Kappa系数提高了14.32%,Adaboost的Kappa系数提高了10.23%,达到最高值0.9336;(3)过采样的处理效果优于欠采样,过采样对分类器的性能提升更多。综上所述,选择合适的采样方法和分类方法是提高不均衡数据集遥感分类精度的有效途径。     

西北旱区遥感影像分类的支持向量机法

针对较大范围、不同时相、不同气候和地貌类型的遥感影像的土地利用现状分类问题,提出了一种结合标准植被指数和纹理特征的支持向量机法。此方法改进了陕西延安、甘肃嘉峪关和青海果洛的遥感影像分类,有效地解决了最大似然法和BP神经网络法的缺陷造成的分类精度不高的问题。分类结果表明:与最大似然法和BP神经网络法相比,结合标准植被指数和纹理特征的支持向量机法的分类总精度最高(97.75%),Kappa系数为0.9691。该方法可为西北旱区遥感影像解译和土地资源可持续发展战略提供方法支撑。     

高光谱影像特征子集选择方法

针对高维遥感数据的降维困难问题,该文提出并构建了一种融合粒子群优化算法全局寻优能力和支持向量机优秀分类性能的高光谱遥感影像特征子集选择与分类方法。通过引入混沌优化搜索技术改进融合粒子群优化算法的全局寻优能力;提出并采用一种基于粒度的网格搜索策略对支持向量机模型参数进行优化;利用二进制融合粒子群优化算法进行特征选择;然后,支持向量机采用该特征子集所对应的训练数据集进行模型参数优化和分类。实验结果表明该方法能有效地提取出用于分类的最佳波段,具有较高的分类精度。为高光谱遥感影像的特征选择与分类探索出了一种可行的方法。     

WorldView-2纹理的森林地上生物量反演

使用高空间分辨率卫星WorldView-2的多光谱遥感影像,构建植被指数和纹理因子等遥感因子与森林地上生物量的关系方程,并计算模型估测精度和均方根误差,探索高分辨率数据的光谱与纹理信息在温带森林地上生物量估测应用中的潜力。以黑龙江省凉水自然保护区温带天然林及天然次生林为研究对象,通过灰度共生矩阵(GLCM)、灰度差分向量(GLDV)及和差直方图(SADH)对高分辨率遥感影像进行纹理信息提取,并利用外业调查的74个样地地上生物量与遥感因子建立参数估计模型。提取的遥感因子包括6种植被指数(比值植被指数RVI、差值植被指数DVI、规一化植被指数NDVI、增强植被指数EVI、土壤调节植被指数SAVI和修正的土壤调节植被指数MSAVI)以及3类纹理因子(GLCM、GLDV和SADH)。为避免特征变量个数较多对估测模型造成过拟合,利用随机森林算法对提取的遥感因子进行特征选择,将最优的特征变量输入模型参与建模估测。采用支持向量回归(SVR)进行生物量建模及验证,结果显示选入模型的和差直方图均值(sadh_mean)、灰度共生矩阵方差(glcm_var)和差值植被指数(DVI)等遥感因子对森林地上生物量有较好的解释效果;植被指数+纹理因子组合的模型获得较精确的AGB估算结果(R2=0.85,RMSE=42.30 t/ha),单独使用植被指数的模型精度则较低(R~2=0.69,RMSE=61.13 t/ha)。     

Per-pixel and object-oriented classification methods for mapping urban land cover extraction using SPOT 5 imagery

To have sustainable management and proper decision-making, timely acquisition and analysis of surface features are necessary. Traditional pixel-based analysis is the popular way to extract different categories, but it is not comparable by the achievements that can be achieved through the object-based method that uses the additional characteristics of features in the process of classification. In this paper, three types of classification were used to classify SPOT 5 satellite image in mapping land cover; Support vector machine (SVM) pixel-based, SVM object-based and Decision Tree (DT) pixel-based classification. Normalised Difference Vegetation Index and the brightness value of two infrared bands (NIR and SWIR) were used in manually developed DT classification. The classification of the SVM (pixel based) was generated using the selected groups of pixels that represent the selected features. In addition, the SVM (object based) was implemented by using radial-based function kernel. The classified features were oil palm, rubber, urban area, soil, water and other vegetation. The study found that the overall classification of the DT was the lowest at 69.87% while those of SVM (pixel based) and SVM (object based) were 76.67 and 81.25%, respectively.     

基于多源国产高分卫星时空信息的米级分辨率耕地提取

及时准确地获取耕地空间分布数据对于农业生产管理、产量估算、种植结构调整等具有重要意义。目前的耕地提取多基于多时相中低分辨率影像或单时相高分辨率影像,难以满足耕地破碎,农作物种植模式复杂的区域精度需求。基于此,本研究通过协同国产高分一号（GF-1）、高分二号（GF-2）和高分六号（GF-6）卫星影像,探索米级分辨率尺度下的耕地高精度提取方法。该方法以深度神经网络UNet为基础,通过协同GF-1/6的多时相优势和GF-2影像的高空间分辨率构建了CEUNet（Cropland Extraction UNet）模型,以充分挖掘耕地的时相特征和空间几何特征。同时,将基于CEUNet模型提取的米级耕地结果分别与基于UNet和多源不同分辨率遥感影像的语义分割（UNet_m）、基于UNet和单时相高分辨率影像的语义分割（UNet_s）、基于对象的随机森林分类（OBIA）、基于像元的随机森林分类（RF）提取的耕地结果展开对比,分析所提出的方法在不同区域的适宜性。结果表明,基于CEUNet模型提取的米级耕地总体精度达到92.92%,且基于CEUNet提取的耕地的逐像元验证结果在平均F1-Score值上相较于基于对象和基于像元的随机森林分类分别提升了0.21和0.21,相较于UNet_m和UNet_s分别提升了0.04和0.11,其中针对地块破碎,景观异质性高等区域,CEUNet相较于UNet_m和UNet_s提升了0.09和0.26。本研究提出的CEUNet模型能够充分发挥多源国产高分卫星数据的空间和时间优势,两者结合能够快速、高效地提取不同农业景观及不同种植模式的耕地空间分布信息。     

基于NDVI城镇土地利用变化检测探讨

通过归一化植被指数（NDVI）结合影像差值，对经过辐射校正后的1993年和2005年武汉地区TM影像进行了土地利用变化检测，与监督法分类进行比较，得出NDVI更宜于实现变化信息探测和提取。     

Mapping Wetland Areas Using Landsat-Derived NDVI and LSWI: A Case Study of West Songnen Plain,Northeast China

Increasing interest in wetlands for environmental management requires an understanding of the location, spatial extent, and configuration of the resource. The National Wetlands Inventory is the most commonly used data source for this information. However, its accuracy is limited in some contexts, such as agricultural and forested wetlands. An large number of studies have mapped wetlands worldwide from the perspective of land use and land cover change. However, information on the actual wetland planting areas annually is limited, which greatly impacts ongoing research. In this case study of the West Songnen Plain, we developed a simple algorithm for the quick mapping of wetlands by utilizing their unique physical features, such as annual display of phenological land-cover change of exposed soils, shallow flooding water, and plants from multi-temporal Landsat images. Temporal variations of the Normalized Difference Vegetation Index (NDVI) and Land Surface Water Index (LSWI) derived from Landsat images in 2010 for wetlands at different growth stages were analyzed. Results show that during the ante-tillering phase, the NDVI value (above zero) is lower than the LSWI value of paddies because of flooding of shallow water; during the reproductive and ripening phases, the NDVI value is higher than the LSWI value (above zero); and during the post-harvest wetland planting phase, the NDVI value is still higher than the LSWI value, but the LSWI value is negative. Wetland areas can be detected using one or two images in the optimum time window. The algorithm based on the difference of NDVI and LSWI values derived from Landsat images was used to extract the actual wetland planting area. Validated alongside statistical data, the algorithm showed high accuracy. Therefore, this algorithm highlights the unique features of wetlands and can help in mapping the actual wetland area annually on a regional scale. Results further indicate that the new method has a classification accuracy of 92 %. In comparison, two traditional methods based on Landsat-7/ETM registered accuracy rates of only 83 % and 87 % respectively.     

Use of field reflectance data for crop mapping using airborne hyperspectral image

Recent developments in hyperspectral remote sensing technologies enable acquisition of image with high spectral resolution, which is typical to the laboratory or in situ reflectance measurements. There has been an increasing interest in the utilization of in situ reference reflectance spectra for rapid and repeated mapping of various surface features. Here we examined the prospect of classifying airborne hyperspectral image using field reflectance spectra as the training data for crop mapping. Canopy level field reflectance measurements of some important agricultural crops, i.e. alfalfa, winter barley, winter rape, winter rye, and winter wheat collected during four consecutive growing seasons are used for the classification of a HyMAP image acquired for a separate location by (1) mixture tuned matched filtering (MTMF), (2) spectral feature fitting (SFF), and (3) spectral angle mapper (SAM) methods. In order to answer a general research question “what is the prospect of using independent reference reflectance spectra for image classification”, while focussing on the crop classification, the results indicate distinct aspects. On the one hand, field reflectance spectra of winter rape and alfalfa demonstrate excellent crop discrimination and spectral matching with the image across the growing seasons. On the other hand, significant spectral confusion detected among the winter barley, winter rye, and winter wheat rule out the possibility of existence of a meaningful spectral matching between field reflectance spectra and image. While supporting the current notion of “non-existence of characteristic reflectance spectral signatures for vegetation”, results indicate that there exist some crops whose spectral signatures are similar to characteristic spectral signatures with possibility of using them in image classification.     

Directional effects on NDVI and LAI retrievals from MODIS: A case study in Brazil with soybean

The Moderate Resolution Imaging Spectroradiometer (MODIS) is largely used to estimate Leaf Area Index (LAI) using radiative transfer modeling (the “main” algorithm). When this algorithm fails for a pixel, which frequently occurs over Brazilian soybean areas, an empirical model (the “backup” algorithm) based on the relationship between the Normalized Difference Vegetation Index (NDVI) and LAI is utilized. The objective of this study is to evaluate directional effects on NDVI and subsequent LAI estimates using global (biome 3) and local empirical models, as a function of the soybean development in two growing seasons (2004–2005 and 2005–2006). The local model was derived from the pixels that had LAI values retrieved from the main algorithm. In order to keep the reproductive stage for a given cultivar as a constant factor while varying the viewing geometry, pairs of MODIS images acquired in close dates from opposite directions (backscattering and forward scattering) were selected. Linear regression relationships between the NDVI values calculated from these two directions were evaluated for different view angles (0–25°; 25–45°; 45–60°) and development stages (<45; 45–90; >90 days after planting). Impacts on LAI retrievals were analyzed. Results showed higher reflectance values in backscattering direction due to the predominance of sunlit soybean canopy components towards the sensor and higher NDVI values in forward scattering direction due to stronger shadow effects in the red waveband. NDVI differences between the two directions were statistically significant for view angles larger than 25°. The main algorithm for LAI estimation failed in the two growing seasons with gradual crop development. As a result, up to 94% of the pixels had LAI values calculated from the backup algorithm at the peak of canopy closure. Most of the pixels selected to compose the 8-day MODIS LAI product came from the forward scattering view because it displayed larger LAI values than the backscattering. Directional effects on the subsequent LAI retrievals were stronger at the peak of the soybean development (NDVI values between 0.70 and 0.85). When the global empirical model was used, LAI differences up to 3.2 for consecutive days and opposite viewing directions were observed. Such differences were reduced to values up to 1.5 with the local model. Because of the predominance of LAI retrievals from the MODIS backup algorithm during the Brazilian soybean development, care is necessary if one considers using these data in agronomic growing/yield models.     

植被指数和纹理特征在影像解译中的应用

以QuickBird影像为例,分别分析了植被指数和纹理特征两大解译标志,植被指数主要介绍了比值植被指数、归一化植被指数、修正型土壤调整植被指数、差值植被指数四种,纹理特征主要基于灰度共生矩阵进行分析,并通过对比分类实验验证了这两大解译标志在遥感影像分类中的作用。     

Evaluating and comparing Sentinel 2A and Landsat-8 Operational Land Imager (OLI) spectral indices for estimating fire severity in a Mediterranean pine ecosystem of Greece

The main purpose of this study is to explore the relationship between three field-based fire severity indices (Composite Burn Index-CBI, Geometrically structure CBI, weighted CBI) and spectral indices derived from Sentinel 2A and Landsat-8 OLI imagery on a recent large fire in Thasos, Greece. We employed remotely sensed indices previously used from the remote sensing fire community (Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR), differenced NDVI, differenced NBR, relative differenced NBR, Relativized Burn Ratio) and seven Sentinel 2A-specific indices considering the availability of spectral information recorded in the red-edge spectral region. The statistical correlation indicated a slightly stronger relationship between the differenced NBR and the GeoCBI for both Sentinel 2A (r = 0.872) and Landsat-8 OLI (r = 0.845) imagery. Predictive local thresholds of dNBR values showed slightly higher classification accuracy for Sentinel 2A (73.33%) than Landsat-8 OLI (71.11%), suggesting the adequacy of Sentinel 2A for forest fire severity assessment and mapping in Mediterranean pine ecosystems. The evaluation of the classification thresholds calculated in this study over other fires with similar pre-fire conditions could contribute in the operational mapping and reconstruction of the historical patterns of fire severity over the Eastern Mediterranean region.     

Estimating crop net primary production using national inventory data and MODIS-derived parameters

National estimates of spatially-resolved cropland net primary production (NPP) are needed for diagnostic and prognostic modeling of carbon sources, sinks, and net carbon flux between land and atmosphere. Cropland NPP estimates that correspond with existing cropland cover maps are needed to drive biogeochemical models at the local scale as well as national and continental scales. Existing satellite-based NPP products tend to underestimate NPP on croplands. An Agricultural Inventory-based Light Use Efficiency (AgI-LUE) framework was developed to estimate individual crop biophysical parameters for use in estimating crop-specific NPP over large multi-state regions. The method is documented here and evaluated for corn (Zea mays L.) and soybean (Glycine max L. Merr.) in Iowa and Illinois in 2006 and 2007. The method includes a crop-specific Enhanced Vegetation Index (EVI), shortwave radiation data estimated using the Mountain Climate Simulator (MTCLIM) algorithm, and crop-specific LUE per county. The combined aforementioned variables were used to generate spatially-resolved, crop-specific NPP that corresponds to the Cropland Data Layer (CDL) land cover product. Results from the modeling framework captured the spatial NPP gradient across croplands of Iowa and Illinois, and also represented the difference in NPP between years 2006 and 2007. Average corn and soybean NPP from AgI-LUE was 917 g C m⁻² yr⁻¹ and 409 g C m⁻² yr⁻¹, respectively. This was 2.4 and 1.1 times higher, respectively, for corn and soybean compared to the MOD17A3 NPP product. Site comparisons with flux tower data show AgI-LUE NPP in close agreement with tower-derived NPP, lower than inventory-based NPP, and higher than MOD17A3 NPP. The combination of new inputs and improved datasets enabled the development of spatially explicit and reliable NPP estimates for individual crops over large regional extents.     

支撑向量机及其遥感影像空间特征提取和分类的应用研究

提出了基于支撑向量机（SVM）的遥感影像空间特征提取的新方法，并以SPOT全色波段影像上城市特征信息的提取为应用实例，并与人工神经网络（ANN）等特征提取方法进行综合比较，认为SVM方法不但能够获得比较高的分类精度，而且在学习速度、自适应能力、特征空间高维不限制、可表达性等方面具有优势。