Temporal optimisation of image acquisition for land cover classification with Random Forest and MODIS time-series

The analysis and classification of land cover is one of the principal applications in terrestrial remote sensing. Due to the seasonal variability of different vegetation types and land surface characteristics, the ability to discriminate land cover types changes over time. Multi-temporal classification can help to improve the classification accuracies, but different constraints, such as financial restrictions or atmospheric conditions, may impede their application. The optimisation of image acquisition timing and frequencies can help to increase the effectiveness of the classification process. For this purpose, the Feature Importance (FI) measure of the state-of-the art machine learning method Random Forest was used to determine the optimal image acquisition periods for a general (Grassland, Forest, Water, Settlement, Peatland) and Grassland specific (Improved Grassland, Semi-Improved Grassland) land cover classification in central Ireland based on a 9-year time-series of MODIS Terra 16 day composite data (MOD13Q1). Feature Importances for each acquisition period of the Enhanced Vegetation Index (EVI) and Normalised Difference Vegetation Index (NDVI) were calculated for both classification scenarios. In the general land cover classification, the months December and January showed the highest, and July and August the lowest separability for both VIs over the entire nine-year period. This temporal separability was reflected in the classification accuracies, where the optimal choice of image dates outperformed the worst image date by 13% using NDVI and 5% using EVI on a mono-temporal analysis. With the addition of the next best image periods to the data input the classification accuracies converged quickly to their limit at around 8–10 images. The binary classification schemes, using two classes only, showed a stronger seasonal dependency with a higher intra-annual, but lower inter-annual variation. Nonetheless anomalous weather conditions, such as the cold winter of 2009/2010 can alter the temporal separability pattern significantly. Due to the extensive use of the NDVI for land cover discrimination, the findings of this study should be transferrable to data from other optical sensors with a higher spatial resolution. However, the high impact of outliers from the general climatic pattern highlights the limitation of spatial transferability to locations with different climatic and land cover conditions. The use of high-temporal, moderate resolution data such as MODIS in conjunction with machine-learning techniques proved to be a good base for the prediction of image acquisition timing for optimal land cover classification results.     

Open land cover from OpenStreetMap and remote sensing

OpenStreetMap (OSM) tags were used to produce a global Open Land Cover (OLC) product with fractional data gaps available at osmlanduse.org. Data gaps in the global OLC map were filled for a case study in Heidelberg, Germany using free remote sensing data, which resulted in a land cover (LC) prototype with complete coverage in this area. Sixty tags in the OSM were used to allocate a Corine Land Cover (CLC) level 2 land use classification to 91.8% of the study area, and the remaining gaps were filled with remote sensing data. For this case study, complete are coverage OLC overall accuracy was estimated 87%, which performed better than the CLC product (81% overall accuracy) of 2012. Spatial thematic overlap for the two products was 84%. OLC was in large parts found to be more detailed than CLC, particularly when LC patterns were heterogeneous, and outperformed CLC in the classification of 12 of the 14 classes. Our OLC product represented data created in different periods; 53% of the area was 2011–2016, and 46% of the area was representative of 2016–2017.     

遥感图像土地覆盖分类中多源特征数据选择研究

多源特征数据可以提高遥感图像的分类精度,选择合适的特征数据十分重要。利用基尼指数对多尺度纹理信息、主成分变换前三分量、地形数据等特征进行选择,选出最佳特征子集。利用支持向量机、神经网络分类法、最大似然法分别对全部特征数据和最佳特征子集结合多光谱数据进行分类。实验结果表明:基尼指数可以有效地对多源特征数据进行选择,特征选择可以提高分类器效率,提高分类精度。     

植被覆盖区高精度遥感地貌场景分类数据集

地貌数据集是实现地貌自动分类和加深对地貌形态学认识的重要支撑数据之一。当前缺乏高精度地貌成因类数据集,制约了地貌遥感自动解译的发展。本文在中国东北地区以沟—弧—盆体系为主的天山—兴蒙造山系中,针对强烈的构造运动和新生代以来的火山作用、流水作用形成的地貌成因类型,制作了构造地貌、火山熔岩地貌和流水地貌3类场景数据集（GOS10m）。数据集覆盖面积约5000 km²,包括哨兵2号可见光遥感影像、SRTM1 DEM及基于DEM提取的7个地貌形态参数（山体晕渲图、坡度、DEM局部平均中值、标准偏差、坡向—向北方向偏移量、坡向—向东方向偏移量和相对偏离平均值）。单张样本图为64像素×64像素,空间分辨率为10 m。采用多模态深度学习神经网络对数据进行训练并分类,平均测试精度可达到82.63%,表明构建的数据集具有较高的质量。可为地貌成因遥感自动分类研究以及推动遥感地貌智能解译的向前发展,提供数据集支撑。     

Object-based crop classification using multi-temporal SPOT-5 imagery and textural features with a Random Forest classifier

In this study, an object-based image analysis (OBIA) approach was developed to classify field crops using multi-temporal SPOT-5 images with a random forest (RF) classifier. A wide range of features, including the spectral reflectance, vegetation indices (VIs), textural features based on the grey-level co-occurrence matrix (GLCM) and textural features based on geostatistical semivariogram (GST) were extracted for classification, and their performance was evaluated with the RF variable importance measures. Results showed that the best segmentation quality was achieved using the SPOT image acquired in September, with a scale parameter of 40. The spectral reflectance and the GST had a stronger contribution to crop classification than the VIs and GLCM textures. A subset of 60 features was selected using the RF-based feature selection (FS) method, and in this subset, the near-infrared reflectance and the image acquired in August (jointing and heading stages) were found to be the best for crop classification.     

Fusion of PolSAR and PolInSAR data for land cover classification

The main research goal of this study is to investigate the complementarity and fusion of different frequencies (L- and P-band), polarimetric SAR (PolSAR) and polarimetric interferometric (PolInSAR) data for land cover classification. A large feature set was derived from each of these four modalities and a two-level fusion method was developed: Logistic regression (LR) as ‘feature-level fusion’ and the neural-network (NN) method for higher level fusion. For comparison, a support vector machine (SVM) was also applied. NN and SVM were applied on various combinations of the feature sets.     

Impact of feature selection on the accuracy and spatial uncertainty of per-field crop classification using Support Vector Machines

Crop mapping is one major component of agricultural resource monitoring using remote sensing. Yield or water demand modeling requires that both, the total surface that is cultivated and the accurate distribution of crops, respectively is known. Map quality is crucial and influences the model outputs. Although the use of multi-spectral time series data in crop mapping has been acknowledged, the potentially high dimensionality of the input data remains an issue. In this study Support Vector Machines (SVM) are used for crop classification in irrigated landscapes at the object-level. Input to the classifications is 71 multi-seasonal spectral and geostatistical features computed from RapidEye time series. The random forest (RF) feature importance score was used to select a subset of features that achieved optimal accuracies. The relationship between the hard result accuracy and the soft output from the SVM is investigated by employing two measures of uncertainty, the maximum a posteriori probability and the alpha quadratic entropy. Specifically the effect of feature selection on map uncertainty is investigated by looking at the soft outputs of the SVM, in addition to classical accuracy metrics. Overall the SVMs applied to the reduced feature subspaces that were composed of the most informative multi-seasonal features led to a clear increase in classification accuracy up to 4.3%, and to a significant decline in thematic uncertainty. SVM was shown to be affected by feature space size and could benefit from RF-based feature selection. Uncertainty measures from SVM are an informative source of information on the spatial distribution of error in the crop maps.     

基于随机森林的地理要素面向对象自动解译方法

面向地理对象影像分析(GEOBIA)技术取得了显著的进展,代表了遥感影像解译的发展范式,其主要目标是发展智能化分析方法。随机森林机器学习方法是一种相对新的、数据驱动的非参数分类方法,具有自动特征优选、自动模型构建等优势,为智能化分析提供了有效手段。充分利用GEOBIA及随机森林机器学习的优势,提出了基于随机森林的地理要素面向对象自动解译方法,阐述了随机森林面向对象分类方法的技术流程,为设计和实现该方法提供了详细指导,有助于指导用户优选特征和构建分类模型。通过与支持向量机分类的对比实验证明,该方法可以自动进行特征优选及分类模型的构建,利用较少的特征得到较高的分类精度,在不损失性能的前提下减少了计算量和内存使用,能够为大范围、大区域地理要素自动解译提供先验知识及自动化的手段。     

Update and spatial extension of strategic forest inventories using time series remote sensing and modeling

Up-to-date forest inventory information relating the characteristics of managed and natural forests is fundamental to sustainable forest management and required to inform conservation of biodiversity and assess climate change impacts and mitigation opportunities. Strategic forest inventories are difficult to compile over large areas and are often quickly outdated or spatially incomplete as a function of their long production cycle. As a consequence, automated approaches supported by remotely sensed data are increasingly sought to provide exhaustive spatial coverage for a set of core attributes in a timely fashion. The objective of this study was to demonstrate the integration of current remotely-sensed data products and pre-existing jurisdictional inventory data to map four forest attributes of interest (stand age, dominant species, site index, and stem density) for a 55 Mha study region in British Columbia, Canada. First, via image segmentation, spectrally homogenous objects were derived from Landsat surface-reflectance pixel composites. Second, a suite of Landsat-based predictors (e.g., spectral indices, disturbance history, and forest structure) and ancillary variables (e.g., geographic, topographic, and climatic) were derived for these units and used to develop predictive models of target attributes. For the often difficult classification of dominant species, two modelling approaches were compared: (a) a global Random Forests model calibrated with training samples collected over the entire study area, and (b) an ensemble of local models, each calibrated with spatially constrained local samples. Accuracy assessment based upon independent validation samples revealed that the ensemble of local models was more accurate and efficient for species classification, achieving an overall accuracy of 72% for the species which dominate 80% of the forested areas in the province. Results indicated that site index had the highest agreement between predicted and reference (R² = 0.74, %RMSE = 23.1%), followed by stand age (R² = 0.62, %RMSE = 35.6%), and stem density (R² = 0.33, %RMSE = 65.2%). Inventory attributes mapped at the image-derived unit level captured much finer details than traditional polygon-based inventory, yet can be readily reassembled into these larger units for strategic forest planning purposes. Based upon this work, we conclude that in a multi-source forest monitoring program, spatially localized and detailed characterizations enabled by time series of Landsat observations in conjunction with ancillary data can be used to support strategic inventory activities over large areas.     

面向遥感影像智能分类的海量样本数据采集方法

以地理国情监测高分辨率遥感影像及高精度地表覆盖分类产品为数据源,提出了一种面向遥感影像智能分类、基于位置匹配技术的全国尺度海量样本数据采集方法。根据数据源特征,研究了县域采集数量权重设置、坐标投影转换、栅格灰度重采样、无效样本数据过滤、地表覆盖分类码映射、样本数据命名标识、特定地表覆盖类型样本数据采集等关键技术,构建了位置匹配的遥感影像数据与分类标签数据组成的样本数据对,开发了样本数据自动采集软件。利用该方法,以县级行政区划为单元,实现了全国尺度海量样本数据采集。选取其中5个县域的成果,评估了方法的实用性及运算性能。研究表明：该方法提升了生产全国尺度海量样本数据的计算响应速度;采集的样本数据能够满足遥感影像智能分类对样本源高质量、大规模的需求,提升了遥感影像分类与预测的准确度。     

A Boosted Genetic Fuzzy Classifier for land cover classification of remote sensing imagery

A Boosted Genetic Fuzzy Classifier (BGFC) is proposed in this paper, for land cover classification from multispectral images. The model comprises a set of fuzzy classification rules, which resemble the reasoning employed by humans. Fuzzy rules are generated in an iterative fashion, incrementally covering subspaces of the feature space, as directed by a boosting algorithm. Each rule is able to select the required features, further improving the interpretability of the obtained model. After the rule generation stage, a genetic tuning stage is employed, aiming at improving the cooperation among the fuzzy rules, thus increasing the classification performance attained after the first stage. The BGFC is tested using an IKONOS multispectral VHR image, in a lake-wetland ecosystem of international importance. For effective classification, we consider advanced feature sets, containing spectral and textural feature types. Comparative results with well-known classifiers, commonly employed in remote sensing tasks, indicate that the proposed system is able to handle multi-dimensional feature spaces more efficiently, effectively exploiting information from different feature sources.     

An elliptical basis function network for classification of remote sensing images

An elliptical basis function (EBF) network is employed in this study for the classification of remotely sensed images. Though similar in structure, the EBF network differs from the well-known radial basis function (RBF) network by incorporating full covariance matrices and employing the expectation-maximization (EM) algorithm to estimate the basis functions. Since remotely sensed data often take on mixture-density distributions in the feature space, the network not only possesses the advantage of the RBF mechanism, but also utilizes the EM algorithm to compute the maximum likelihood estimates of the mean vectors and covariance matrices of a Gaussian mixture distribution in the training phase. Experimental results show that the EM-based EBF network is more effective in training and simpler in structure than an RBF network constructed for the same task.The research was supported by grant 40101021 from the Natural Science Foundation of China, and grant 2002AA135230 from Hi-Tech research and development program of China. The authors would like to thank the reviewers for their valuable comments.     

Methods for improving accuracy and extending results beyond periods covered by traditional ground-truth in remote sensing classification of a complex landscape

Successful development of approaches to quantify impacts of diverse landuse and associated agricultural management practices on ecosystem services is frequently limited by lack of historical and contemporary landuse data. We hypothesized that ground truth data from one year could be used to extrapolate previous or future landuse in a complex landscape where cropping systems do not generally change greatly from year to year because the majority of crops are established perennials or the same annual crops grown on the same fields over multiple years. Prior to testing this hypothesis, it was first necessary to classify 57 major landuses in the Willamette Valley of western Oregon from 2005 to 2011 using normal same year ground-truth, elaborating on previously published work and traditional sources such as Cropland Data Layers (CDL) to more fully include minor crops grown in the region. Available remote sensing data included Landsat, MODIS 16-day composites, and National Aerial Imagery Program (NAIP) imagery, all of which were resampled to a common 30 m resolution. The frequent presence of clouds and Landsat7 scan line gaps forced us to conduct of series of separate classifications in each year, which were then merged by choosing whichever classification used the highest number of cloud- and gap-free bands at any given pixel. Procedures adopted to improve accuracy beyond that achieved by maximum likelihood pixel classification included majority-rule reclassification of pixels within 91,442 Common Land Unit (CLU) polygons, smoothing and aggregation of areas outside the CLU polygons, and majority-rule reclassification over time of forest and urban development areas. Final classifications in all seven years separated annually disturbed agriculture, established perennial crops, forest, and urban development from each other at 90 to 95% overall 4-class validation accuracy. In the most successful use of subsequent year ground-truth data to classify prior year landuse, an overall 57-class accuracy of 75% was achieved despite the omission of 10 entire classes, most of which were annually disturbed or perennial crops grown on very few fields. Synthetic ground-truth data for the 2004 harvest year based on the most common landuse classes over the following 7 years classified 49 of 57 categories at an overall accuracy of 96% in a final version that included CLU polygon majority rule, default smoothing and aggregation, and forcing of urban development and forest from multi-year majority-rule.     

Multi-temporal remote sensing data applied in automatic land cover update using iterative training sample selection and Markov Random Field model

Automatic land cover update was an effective means to obtain objective and timely land cover maps without human disturbance. This study investigated the efficacy of multi-temporal remote sensing data and advanced non-parametric classifier on improving the classification accuracy of the automatic land cover update approach integrating iterative training sample selection and Markov Random Fields model when the historical remote sensing data were unavailable. The results indicated that two-temporal remote sensing data acquired in one crop growth season could significantly improve the classification accuracy of the automatic land cover update approach by approximately 3–4%. However, the support vector machine (SVM) classifier was not suitable to be integrated in the automatic land cover update approach, because the huge initially selected training samples made the training of the SVM classifier unrealizable.     

The potential of linear discriminative Laplacian eigenmaps dimensionality reduction in polarimetric SAR classification for agricultural areas

In this paper, the linear discriminative Laplacian eigenmaps (LDLE) dimensionality reduction (DR) algorithm is introduced to C-band polarimetric synthetic aperture radar (PolSAR) agricultural classification. A collection of homogenous areas of the same crop class usually presents physical parameter variation, such as the biomass and soil moisture. Furthermore, the local incidence angle also impacts a lot on the same crop category when the vegetation layer is penetrable with C-band radar. We name this phenomenon as the “observed variation of the same category” (OVSC). The most common PolSAR features, e.g., the Freeman–Durden and Cloude–Pottier decompositions, show an inadequate performance with OVSC. In our research, more than 40 coherent and incoherent PolSAR decomposition models are stacked into the high-dimensionality feature cube to describe the various physical parameters. The LDLE algorithm is then performed on the observed feature cube, with the aim of simultaneously pushing the local samples of the same category closer to each other, as well as maximizing the distance between local samples of different categories in the learnt subspace. Finally, the classification result is obtained by nearest neighbor (NN) or Wishart classification in the reduced feature space. In the simulation experiment, eight crop blocks are picked to generate a test patch from the 1991 Airborne Synthetic Aperture Radar (AIRSAR) C-band fully polarimetric data from of Flevoland test site. Locality preserving projections (LPP) and principal component analysis (PCA) are then utilized to evaluate the DR results of the proposed method. The classification results show that LDLE can distinguish the influence of the physical parameters and achieve a 99% overall accuracy, which is better than LPP (97%), PCA (88%), NN (89%), and Wishart (88%). In the real data experiment, the Chinese Hailaer nationalized farm RadarSat2 PolSAR test set is used, and the classification accuracy is around 94%, which is again better than LPP (90%), PCA (88%), NN (89%), and Wishart (85%). Both experiments suggest that the LDLE algorithm is an effective way of relieving the OVSC phenomenon.     

面向对象的旱区植被遥感精细分类研究

针对旱区植被分类尺度过大、种群无法准确提取的问题,该文提出了面向对象的CFS-RF分类模型,即利用CFS算法对先验样本数据集进行特征优选,结合随机森林构建分类规则,完成分类过程。以新疆阿勒泰为研究区,利用GF-2数据,通过CFS、ReliefF两种不同特征选择方法和J48、SVM、RF 3种分类算法构造出6种面向对象分类方案来实现小尺度植被种群提取。结果表明,经过特征选择,上述分类方案的精度和效率均得到了提升。其中,CFS-RF算法最优,总体精度达到92.41%,Kappa系数为0.90,更适用于旱区植被遥感精细分类。     

An adaptive unimodal subclass decomposition (AUSD) learning system for land use and land cover classification using high-resolution remote sensing

Land use and land cover classification is an important application of remote-sensing images. The performances of most classification models are largely limited by the incompleteness of the calibration set and the complexity of spectral features. It is difficult for models to realize continuous learning when the study area is transferred or enlarged. This paper proposed an adaptive unimodal subclass decomposition (AUSD) learning system, which comprises two-level iterative learning controls: The inner loop separates each class into several unimodal Gaussian subclasses; the outer loop utilizes transfer learning to extend the model to adapt to supplementary calibration set collected from enlarged study areas. The proposed model can be efficiently adjusted according to the variability of spectral signatures caused by the increasingly high-resolution imagery. The classification result can be obtained using the Gaussian mixture model by Bayesian decision theory. This AUSD learning system was validated using simulated data with the Gaussian distribution and multi-area SPOT-5 high-resolution images with 2.5-m resolution. The experimental results on numerical data demonstrated the ability of continuous learning. The proposed method achieved an overall accuracy of over 90% in all the experiments, validating the effectiveness as well as its superiority over several widely used classification methods.     

Mapping the distributions of C3 and C4 grasses in the mixed-grass prairies of southwest Oklahoma using the Random Forest classification algorithm

The objective of this paper is to demonstrate a new method to map the distributions of C₃ and C₄ grasses at 30 m resolution and over a 25-year period of time (1988–2013) by combining the Random Forest (RF) classification algorithm and patch stable areas identified using the spatial pattern analysis software FRAGSTATS. Predictor variables for RF classifications consisted of ten spectral variables, four soil edaphic variables and three topographic variables. We provided a confidence score in terms of obtaining pure land cover at each pixel location by retrieving the classification tree votes. Classification accuracy assessments and predictor variable importance evaluations were conducted based on a repeated stratified sampling approach. Results show that patch stable areas obtained from larger patches are more appropriate to be used as sample data pools to train and validate RF classifiers for historical land cover mapping purposes and it is more reasonable to use patch stable areas as sample pools to map land cover in a year closer to the present rather than years further back in time. The percentage of obtained high confidence prediction pixels across the study area ranges from 71.18% in 1988 to 73.48% in 2013. The repeated stratified sampling approach is necessary in terms of reducing the positive bias in the estimated classification accuracy caused by the possible selections of training and validation pixels from the same patch stable areas. The RF classification algorithm was able to identify the important environmental factors affecting the distributions of C₃ and C₄ grasses in our study area such as elevation, soil pH, soil organic matter and soil texture.     

基于知识的分层综合分类法在土地利用/土地覆盖遥感信息提取中的应用

土地利用/土地覆盖数据的获取是研究LUCC的重要基础工作。随着遥感技术的飞速发展,通过遥感提取土地利用/土地覆盖专题信息已成为LUCC研究必不可少的一步。目前遥感专题信息提取水平相对滞后于遥感数据获取,为了提高遥感数据在土地利用/土地覆盖的应用,寻找一种较好的、具有相对适用性的方法是目前遥感应用的一个迫切要求。本文比较了目前比较常用的几种土地利用/土地覆盖遥感信息提取方法,分别以西部干旱区(柴达木盆地)和东部地区(鄱阳湖地区)为例,提出在GIS支持下基于知识的分层综合分类方法,并通过和其他几种常用方法进行比较分析,得到如下结果:在自然环境相差较大的柴达木盆地和鄱阳湖地区,采用了GIS支持下基于知识的分层综合分类方法的提取精度均要比单独采用最大似然法、纹理分析法、神经网络分类法等方法的总体精度高出25%,Kappa系数高出0.2。由此可以说明了该方法对于土地利用/土地覆盖专题信息的提取是可行的,同时它也具有一定的适用性。     

Assessing integration of intensity,polarimetric scattering,interferometric coherence and spatial texture metrics in PALSAR-derived land cover classification

Synthetic aperture radar (SAR) is an important alternative to optical remote sensing due to its ability to acquire data regardless of weather conditions and day/night cycle. The Phased Array type L-band SAR (PALSAR) onboard the Advanced Land Observing Satellite (ALOS) provided new opportunities for vegetation and land cover mapping. Most previous studies employing PALSAR investigated the use of one or two feature types (e.g. intensity, coherence); however, little effort has been devoted to assessing the simultaneous integration of multiple types of features. In this study, we bridged this gap by evaluating the potential of using numerous metrics expressing four feature types: intensity, polarimetric scattering, interferometric coherence and spatial texture. Our case study was conducted in Central New York State, USA using multitemporal PALSAR imagery from 2010. The land cover classification implemented an ensemble learning algorithm, namely random forest. Accuracies of each classified map produced from different combinations of features were assessed on a pixel-by-pixel basis using validation data obtained from a stratified random sample. Among the different combinations of feature types evaluated, intensity was the most indispensable because intensity was included in all of the highest accuracy scenarios. However, relative to using only intensity metrics, combining all four feature types increased overall accuracy by 7%. Producer’s and user’s accuracies of the four vegetation classes improved considerably for the best performing combination of features when compared to classifications using only a single feature type.