0 引言
为了兼顾遥感影像分类方法的精度与效率,本文提出了一种基于影像窗口子区的耕地类型自动识别算法。该算法通过对影像的规则划分获取一定大小的影像窗口子区,在提取多光谱和多层次特征的基础上,利用机器学习算法,自动识别影像窗口子区耕地和非耕地类型。以东北地区的高空间分辨率遥感影像为例,建立实验数据集以构建该区域特征库,实现耕地和非耕地类型的非监督自动识别,大大提高了分类算法的自动化程度; 同时也对影像窗口子区尺度和样本容量进行了探讨。
1 算法原理及流程
结合高空间分辨率遥感数据特点,本文提出的基于影像窗口子区的耕地类型自动识别算法分为3个步骤: 首先,通过获取一定数量典型地类的纯净窗口子区构建实验数据集; 然后,对实验数据集进行多光谱、多层次特征提取,建立不同地类的特征库; 最后,采用机器学习分类算法实现各地类的自动识别。
1.1 实验数据集构建
高空间分辨率遥感影像数据中,典型地物类型会在一定的窗口尺度内保持良好的均一性,这样尺度的窗口称为纯净窗口子区,该尺度可作为该类型的空间展布尺度[14],可用一定数量影像窗口子区中纯净窗口子区所占比例表示该尺度下地物的纯净度。依照耕地类型的空间展布尺度,通过随机对遥感影像的规则切分,获取一定数量的影像窗口子区,构建实验数据集,作为样本集。
1.2 特征库建立
构建了实验数据集之后,将对其影像窗口子区进行多光谱、多层次特征提取以建立特征库,作为对影像进行识别分类的依据。
多光谱特征是针对遥感影像多波段数据特点,按照波段数将遥感影像拆分为多个灰度图像,将多波段多图像的光谱特征作为特征库建立的基础。
多层次特征是由对影像的初级特征到高级特征的抽象所得[17,18],包括尺度不变特征变换(scale-invariant feature transform,SIFT)、稀疏编码和最大化池化。SIFT是对影像特征的初级提取,反映了影像的关键点和局部特征[19]。在其基础上,进行稀疏编码表示,获得更高层次的特征[20]。其求解过程包括训练阶段和编码阶段: 训练阶段是对训练数据图像的SIFT特征进行稀疏编码,同时求解稀疏编码约束函数中的稀疏编码表示和字典; 编码阶段是利用学习得到的最优字典,对测试数据图像的SIFT特征进行稀疏编码,求解其稀疏编码表示。得到稀疏编码表示后,通常需要对其进行最大化池化,将向量中各维相应的最大分量组成一个特征向量来表示该影像,获得多层次特征。
1.3 地类识别
图1
2 数据源与实验结果
实验数据源选择0.4 m空间分辨率的Pleiades卫星遥感数据,共4个波段,分别为蓝光波段B1、绿光波段B2、红光波段B3和近红外波段B4。
图2
图3
图3
研究区遥感影像
(Pleiades B3(R),B2(G),B1(B)合成)
Fig.3
Remote sensing images in study areas
2.1 地物类型和样本选取结果
由土地利用/土地覆盖数据,得到研究区内耕地、林地和居民地类型所占面积比例分别约为65%,35%和5%,以此作为样本选取的依据。
为获取纯净窗口子区大小,通过统计窗口子区尺度与纯净度的关系曲线,获得不同尺度下地物的纯净度。本文随机选择256像元
图4
2.2 样本容量选取结果
Tab.2 Sample size of each type(个)
| 数据集 | 耕地 | 林地 | 居民地 | 合计 |
|---|---|---|---|---|
| 训练/验证数据集A | 400 | 150 | 22 | 572 |
| 训练/验证数据集B | 270 | 100 | 15 | 385 |
| 训练/验证数据集C | 200 | 80 | 11 | 291 |
| 训练/验证数据集D | 140 | 50 | 7 | 197 |
| 训练/验证数据集E | 100 | 38 | 5 | 143 |
| 训练/验证数据集F | 68 | 28 | 4 | 100 |
| 测试数据集 | 600 | 250 | 33 | 883 |
图5
图5
样本容量与分类精度关系
Fig.5
Relationship between sample size and classification accuracy
本文主要研究对象为耕地类型,因此将林地和居民地进行合并,作为非耕地类型。从图5中可以看出,随着训练/验证数据集样本容量的增加,居民地分类精度大幅上升,林地和非耕地分类精度有一定幅度的提高,耕地分类精度也有上升趋势。为保障各类型识别精度要求,训练/验证数据集样本容量选择为572,耕地分类精度为97.0%,林地分类精度为81.6%,居民地分类精度为63.6%,非耕地分类精度为79.5%,总体精度为90.8%。
2.3 不同特征库条件下的自动识别精度
在确定了影像窗口子区尺度和样本容量后,选择了2套不同特征库的构建方案进行对比实验,分别对多光谱数据的多波段多图像以及合成单图像进行特征提取构建特征库,并对构建的特征库进行了多次精度验证。各方案模型在验证阶段的分类精度均值和方差如表3所示。
表3 各方案模型分类精度均值和方差
Tab.3
| 图像类型 | 均值 | 方差 |
|---|---|---|
| 多波段多图像 | 0.840 | 0.031 |
| 合成单图像 | 0.750 | 0.043 |
从表3可以看出,选择多波段多图像的自动识别方法精度较高,在验证阶段达到84.0%,远高于合成单图像的自动识别方法。
2.4 基于特征库的耕地类型识别精度
利用训练所得模型,对测试数据集进行识别分类,并将类型识别结果与已知标签对比获取其分类精度。各方案模型在测试阶段的不同地物类型识别精度如表4所示。
表4 各方案模型识别精度
Tab.4
| 图像类型 | 耕地 | 非耕地 | 总体 |
|---|---|---|---|
| 多波段多图像 | 0.962 | 0.795 | 0.908 |
| 合成单图像 | 0.950 | 0.703 | 0.871 |
从表3可见,多波段多图像的自动识别方法总体精度较高,在测试阶段达到了90.8%,具有较为理想的自动识别效果。利用该方法,以构建某一区域特征库为基础,实现对耕地类型的非监督分类,大大增强了分类方法的自动化程度。
3 结论与讨论
本文提出了一种遥感影像自动识别耕地类型的机器学习算法,基于对影像的规则切分获取一定大小的影像窗口子区,通过提取其多光谱和多层次特征,利用机器学习算法,实现耕地和非耕地类型的自动判别。通过研究,得到了以下结论:
1)与传统基于像元和面向对象分析方法相比,本文算法仅以规则切分后的影像纯净窗口子区作为研究对象进行识别,无需对单一像元进行分析处理,也避免了对高空间分辨率影像进行复杂的分割操作。研究表明,窗口子区尺度和样本容量的选择是决定分类精度与效率的主要参量,即窗口子区尺度选择的纯净度和样本容量选择的代表性是分类精度的关键,为了兼顾工作效率,应选择满足纯净度要求的最大窗口子区尺度和满足分类精度要求的最小样本容量。本文中研究区典型地物纯净窗口子区的尺度选择为200 m
2)通过对图像提取多光谱和多层次特征信息,构建特征库,利用机器学习算法,实现了对高空间分辨率遥感影像任意窗口子区耕地类型的自动识别,精度达到了90.8%。通过不同特征库条件下的对比实验发现,训练过程中所采用的策略会对分类精度产生影响,在只利用合成单图像进行多层次特征提取与识别分类时,精度下降到了87.1%。可见基于多光谱和多层次特征的地类识别机器学习算法可充分利用多光谱数据特征,提高分类模型的性能。
3)通过特征库的构建,使遥感影像分类过程中主要关注纯净窗口子区的规则切分以获取样本对象,同时本文算法所选取的训练样本并不局限于同一景影像中,因此适用于某一区域内大量遥感数据的自动化分类处理。在得到影像块数据集后进行特征提取,即可利用特征库进行地物类别自动识别,简化了分类阶段的训练过程,实现非监督的耕地识别,提高了分类算法的自动化程度,同时也可以用于从遥感影像中某一种纯净地物类型的快速提取。
本文研究也发现了一些不足之处,如识别的精度受到训练样本代表性及纯净度的限制。若训练样本中耕地和非耕地的混合样本数超过15%,不满足窗口子区纯净度要求时,会导致获取特征的不稳定,从而对分类精度有一定影响。因此,在训练阶段需要采用一定训练样本的选取策略才能保证分类结果的精度。今后的研究工作将着重探讨纯净窗口子区的自动获取与快速高效的自动分类研究。
参考文献
全球30 m空间分辨率耕地遥感制图研究
[J].
Research on global 30 m spatial resolution remote sensing mapping of cultivated land
[J].
1990—2013年东北地区耕地时空变化遥感分析
[J].
DOI:10.11975/j.issn.1002-6819.2016.07.001
URL
[本文引用: 1]
为揭示1990-2013年东北地区耕地变化规律,以Landsat TM/ETM+/OLI遥感影像为数据源,采用面向对象与目视解译相结合的分类方法,提取东北地区1990年、2000年和2013年耕地信息,辅以气候、地形、社会经济等数据,分析耕地时空变化特征及其驱动因素。结果表明:面向对象分类与目视解译相结合的方法能够快速、准确地完成耕地空间分布信息提取,总体分类精度达93%以上;2013年东北地区现有耕地4.17×105 km2,占东北地区土地总面积的33.51%,其中旱田比例为86.24%;1990-2013年耕地总量增加,2000-2013年比1990-2000年耕地增加速度减小,空间上,耕地呈现南减北增的趋势,新增耕地重心逐渐北移,建设用地侵占耕地速度加快,以省会城市为中心,向东北部蔓延;对湿地、林地、草地开垦减弱,对裸地和海面的农业开发增强,退耕还湿、还草、还林效果明显,耕地变化活跃区处于年降水量在500~700 mm、年均气温在2~4℃、坡度小于3°、高程小于200 m的地理环境中;旱田改为水田规模扩大,旱田质心变化不显著,水田质心向东北移207.41 km。23 a间,虽然气候变化对东北地区耕地变化有一定影响,但政策调控、经济发展、科技进步、人口增长、城市扩张等人文因素仍然是耕地变化和时空分异特征的主要驱动因素。农业发展转型、提高单位面积产量和劳动生产效率,将是未来东北地区耕地发展的必由之路。
Spatio-temporal dynamics analysis of cropland in Northeast China during 1990—2013 based on remote sensing
[J].
A combined fuzzy pixel-based and object-based approach for classification of high-resolution multispectral data over urban areas
[J].
DOI:10.1109/TGRS.2003.815972
URL
[本文引用: 1]
In this paper, we present an object-based approach for urban land cover classification from high-resolution multispectral image data that builds upon a pixel-based fuzzy classification approach. This combined pixel/object approach is demonstrated using pan-sharpened multispectral IKONOS imagery from dense urban areas. The fuzzy pixel-based classifier utilizes both spectral and spatial information to discriminate between spectrally similar road and building urban land cover classes. After the pixel-based classification, a technique that utilizes both spectral and spatial heterogeneity is used to segment the image to facilitate further object-based classification. An object-based fuzzy logic classifier is then implemented to improve upon the pixel-based classification by identifying one additional class in dense urban areas: nonroad, nonbuilding impervious surface. With the fuzzy pixel-based classification as input, the object-based classifier then uses shape, spectral, and neighborhood features to determine the final classification of the segmented image. Using these techniques, the object-based classifier is able to identify buildings, impervious surface, and roads in dense urban areas with 76%, 81%, and 99% classification accuracies, respectively.
基于逐像元和面向对象分类方法的分析研究
[J].
Analysis and research on the the pixel-based and object oriented classification
[J].
基于多特征对象的高分辨率遥感影像分类方法及其应用
[J].
A method for classification of high resolution remotely sensed images based on multi-feature objects and its application
[J].
基于影像融合和面向对象技术的植被信息提取研究
[J].
DOI:10.11873/j.issn.1004-0323.2014.1.0164
URL
[本文引用: 1]
Vegetation is an important part in ecological system and indicating certain landscapes,It is a meaningful work to obtain detailed information of vegetation using GeoEye image with its abundant spatial and spectral information.This study fused the panchromatic band and multispectral bands with four image fusion methods,Image segmentation is the first and critical procedure in the workflow of object\|oriented image analysis,discrepancy between reference polygons and corresponding segment is used to assess segmentation quality in this study.We extracted the vegetation information using classification feature which is selected from the perspective of remote sensing image cognition and geographical understanding.The results showed that Gram\|Schmidt(GS)method is the most effective in fusing panchromatic bands and multispectral bands,And object\|oriented classification is effective in high resolution remote sensing information extraction,the overall accuracy is up to 90.3%.this research provided an effective method for vegetation information extraction.
Study on vegetation information extraction based on object-oriented image analysis
[J].
面向对象的高分辨率影像单值分类耕地提取方法研究
[J].
Method research on cultivated land extraction based on object one-class classification of high-spatial-resolution images
[J].
A multiscale framework for landscape analysis:Object-specific analysis and upscaling
[J].
DOI:10.1023/A:1013101931793
URL
[本文引用: 1]
Landscapes are complex systems that require a multiscale approach to fully understand, manage, and predict their behavior. Remote sensing technologies represent the primary data source for landscape analysis, but suffer from the modifiable areal unit problem ( MAUP ). To reduce the effects of MAUP when using remote sensing data for multiscale analysis we present a novel analytical and upscaling framework based on the spatial influence of the dominant objects composing a scene. By considering landscapes as hierarchical in nature, we theorize how a multiscale extension of this object-specific framework may assist in automatically defining critical landscape thresholds, domains of scale, ecotone boundaries, and the grain and extent at which scale-dependent ecological models could be developed and applied through scale.
面向对象的高分辨率遥感影像信息提取技术
[D].
Object-oriented Information Extraction Technology of High Resolution Remote Sensing Image
[D].
一种基于支撑向量机的遥感影像不完全监督分类新方法
[J].
Partially supervised classification of remotely sensed imagery using support vector machines
[J].
加权纹理特征在高分辨率遥感影像耕地再分类中的应用
[J].
Sub-classification of farmland with weighted textural features in high resolution remote sensing images
[J].
Tracing the evolution of 2010 Merapi volcanic deposits (Indonesia) based on object-oriented classification and analysis of multi-temporal,very high resolution images
[J].
DOI:10.1016/j.rse.2015.09.028
URL
[本文引用: 1]
About 40% of the devastated upper catchment was recolonized by vegetation between 2010 and 2012. The recovery also took place in the forested valley margins affected by ash-cloud surges. The morphometric analysis of the initial drainage network, digitized from the 2011 2012 images, demonstrated (1) the resurfacing of pristine 2010 PDC deposits by runoff and (2) incision or remobilization by lahars. It took two years following the eruption in the rugged upper catchment devastated by high-energy surges to fully develop the hydrographic network. It is, however, still rudimentary on gently sloping fans created by overbank PDC deposits in the middle valley, thus suggesting the importance of slope gradient, grain size, permeability and thickness of deposits. As much as 35% of the 2010 PDC deposits, emplaced in the vicinity of the river channels, were remobilized by lahars over the two post-eruption rainy seasons and also by constant mining activities. Studies on the erosion of the pyroclastic deposits after 2012 need to concentrate on the upper reach of the catchment on the south flank.
遥感图像典型地物特征提取的尺度效应研究
[D].
The Study of Scale Effects in Typical Objects’ Feature Extraction from Remote Sensing Image
[D].
训练样本对TM尺度小麦种植面积测量精度影响研究(Ⅰ)——训练样本与分类方法间分类精度响应关系研究
[J].
The effects of training samples on the wheat planting area measure accuracy in TM scale(Ⅰ):The accuracy response of different classifiers to training samples
[J].
延怀盆地景观格局动态过程研究
[D].
Research on the Dynamic Process of Landscape Pattern in Yan Huai Basin
[D].
面向机器学习的高分辨率光学遥感影像识别方法
[D].
High Spatial Rseolution Remote Sensing Images Recognition by Means of Machine Learning
[D].
Distinctive image features from scale-invariant keypoints
[J].
DOI:10.1023/B:VISI.0000029664.99615.94
URL
[本文引用: 1]
This paper presents a method for extracting distinctive invariant features from images that can be used to perform reliable matching between different views of an object or scene. The features are invariant to image scale and rotation, and are shown to provide robust matching across a substantial range of affine distortion, change in 3D viewpoint, addition of noise, and change in illumination. The features are highly distinctive, in the sense that a single feature can be correctly matched with high probability against a large database of features from many images. This paper also describes an approach to using these features for object recognition. The recognition proceeds by matching individual features to a database of features from known objects using a fast nearest-neighbor algorithm, followed by a Hough transform to identify clusters belonging to a single object, and finally performing verification through least-squares solution for consistent pose parameters. This approach to recognition can robustly identify objects among clutter and occlusion while achieving near real-time performance.
Linear spatial pyramid matching using sparse coding for image classification
Toward an optimal SVM classification system for hyperspectral remote sensing images
[J].
DOI:10.1109/TGRS.2006.880628
URL
[本文引用: 1]
Recent remote sensing literature has shown that support vector machine (SVM) methods generally outperform traditional statistical and neural methods in classification problems involving hyperspectral images. However, there are still open issues that, if suitably addressed, could allow further improvement of their performances in terms of classification accuracy. Two especially critical issues are: 1) the determination of the most appropriate feature subspace where to carry out the classification task and 2) model selection. In this paper, these two issues are addressed through a classification system that optimizes the SVM classifier accuracy for this kind of imagery. This system is based on a genetic optimization framework formulated in such a way as to detect the best discriminative features without requiring the a priori setting of their number by the user and to estimate the best SVM parameters (i.e., regularization and kernel parameters) in a completely automatic way. For these purposes, it exploits fitness criteria intrinsically related to the generalization capabilities of SVM classifiers. In particular, two criteria are explored, namely: 1) the simple support vector count and 2) the radius margin bound. The effectiveness of the proposed classification system in general and of these two criteria in particular is assessed both by simulated and real experiments. In addition, a comparison with classification approaches based on three different feature selection methods is reported, i.e., the steepest ascent (SA) algorithm and two other methods explicitly developed for SVM classifiers, namely: 1) the recursive feature elimination technique and 2) the radius margin bound minimization method
