Object-based classification of hyperspectral data using Random Forest algorithm

This paper presents a new framework for object-based classification of high-resolution hyperspectral data. This multi-step framework is based on multi-resolution segmentation (MRS) and Random Forest classifier (RFC) algorithms. The first step is to determine of weights of the input features while using the object-based approach with MRS to processing such images. Given the high number of input features, an automatic method is needed for estimation of this parameter. Moreover, we used the Variable Importance (VI), one of the outputs of the RFC, to determine the importance of each image band. Then, based on this parameter and other required parameters, the image is segmented into some homogenous regions. Finally, the RFC is carried out based on the characteristics of segments for converting them into meaningful objects. The proposed method, as well as, the conventional pixel-based RFC and Support Vector Machine (SVM) method was applied to three different hyperspectral data-sets with various spectral and spatial characteristics. These data were acquired by the HyMap, the Airborne Prism Experiment (APEX), and the Compact Airborne Spectrographic Imager (CASI) hyperspectral sensors. The experimental results show that the proposed method is more consistent for land cover mapping in various areas. The overall classification accuracy (OA), obtained by the proposed method was 95.48, 86.57, and 84.29% for the HyMap, the APEX, and the CASI data-sets, respectively. Moreover, this method showed better efficiency in comparison to the spectral-based classifications because the OAs of the proposed method was 5.67 and 3.75% higher than the conventional RFC and SVM classifiers, respectively.     

Crop classification from full-year fully-polarimetric L-band UAVSAR time-series using the Random Forest algorithm

Accurate and timely information on the distribution of crop types is vital to agricultural management, ecosystem services valuation and food security assessment. Synthetic Aperture Radar (SAR) systems have become increasingly popular in the field of crop monitoring and classification. However, the potential of time-series polarimetric SAR data has not been explored extensively, with several open scientific questions (e.g. the optimal combination of image dates for crop classification) that need to be answered. In this research, the usefulness of full year (both 2011 and 2014) L-band fully-polarimetric Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) data in crop classification was fully investigated over an agricultural region with a heterogeneous distribution of crop categories. In total, 11 crop classes including tree crops (almond and walnut), forage crops (grass, alfalfa, hay, and clover), a spring crop (winter wheat), and summer crops (corn, sunflower, tomato, and pepper), were discriminated using the Random Forest (RF) algorithm. The SAR input variables included raw linear polarization channels as well as polarimetric parameters derived from Cloude-Pottier (CP) and Freeman-Durden (FD) decompositions. Results showed clearly that the polarimetric parameters yielded much higher classification accuracies than linear polarizations. The combined use of all variables (linear polarizations and polarimetric parameters) produced the maximum overall accuracy of 90.50 % and 84.93 % for 2011 and 2014, respectively, with a significant increase of approximately 8 percentage points compared with linear polarizations alone. The variable importance provided by the RF illustrated that the polarimetric parameters had a far greater influence than linear polarizations, with the CP parameters being much more important than the FD parameters. The most important acquisitions were the images dated during the peak biomass stage (July and August) when the differences in structural characteristics between most crops were the largest. At the same time, the images in spring (April and May) and autumn (October) also contributed to the crop classification since they respectively provided unique information for discriminating fruit crops (almond and walnut) as well as summer crops (corn, sunflower, and tomato). As a result, the combined use of only four acquisitions (dated May, July, August, and October for 2011 and April, June, August, and October for 2014) was adequate to achieve a nearly-optimal overall accuracy. In light of the promising classification accuracies demonstrated in this research, it becomes increasingly viable to provide accurate and up-to-date crops inventories over large areas based solely on multitemporal polarimetric SAR.     

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.     

An evaluation of Guided Regularized Random Forest for classification and regression tasks in remote sensing

New Earth observation missions and technologies are delivering large amounts of data. Processing this data requires developing and evaluating novel dimensionality reduction approaches to identify the most informative features for classification and regression tasks. Here we present an exhaustive evaluation of Guided Regularized Random Forest (GRRF), a feature selection method based on Random Forest. GRRF does not require fixing a priori the number of features to be selected or setting a threshold of the feature importance. Moreover, the use of regularization ensures that features selected by GRRF are non-redundant and representative. Our experiments based on various kinds of remote sensing images, show that GRRF selected features provides similar results to those obtained when using all the available features. However, the comparison between GRRF and standard random forest features shows substantial differences: in classification, the mean overall accuracy increases by almost 6% and, in regression, the decrease in RMSE almost reaches 2%. These results demonstrate the potential of GRRF for remote sensing image classification and regression. Especially in the context of increasingly large geodatabases that challenge the application of traditional methods.     

SPOT4-VEGETATION中国西北地区土地覆盖制图与验证

利用SPOT4 VEGETATION的遥感数据产品生成的NDVI与NDWI植被指数时间序列图像集 ,通过ISODATA非监督分类方法 ,编制中国西北地区土地覆盖图。以TM影像人工解译结果作为真实值 ,通过对西北五省共计 47个均匀分布且异质性强的 2 5km× 2 5km样本区内的土地覆盖类型及其面积进行统计分析 ,修正了SPOT4 VEGETATION的土地覆盖分类系统 ,建立了各省验证结果的样本统计直方图并计算其回归系数。结果表明SPOT4 VEGETATION中国西北地区土地覆盖图在总体上具有较高的准确性。影响遥感数据自动分类精度 ,造成土地覆盖误判的原因主要来源于两个方面 :即异物同谱和混合像元问题。对于前者通过叠加各种辅助数据如DEM等可以降低误判的机率 ;对于后者运用混合像元分解的各种算法可以提高分类精度     

Contribution of multispectral (optical and radar) satellite images to the classification of agricultural surfaces

The monitoring of different crops (cultivated plots) and types of surface (bare soils, etc.) is a crucial economic and environmental issue for the management of resources and human activity. In this context, the objective of this study is to evaluate the contribution of multispectral satellite imagery (optical and radar) to land use and land cover classification.Object-oriented supervised classifications, based on a Random Forest algorithm, and majority zoning post-processing are used. This study emerges from the experiment on multi-sensor crop monitoring (MCM'10, Baup et al., 2012) conducted in 2010 on a mixed farming area in the southwest of France, near Toulouse. This experiment enabled the regular and quasi-synchronous collection of multi-sensor satellite data and in situ observations, which are used in this study. 211 plots with contrasting characteristics (different slopes, soil types, aspects, farming practices, shapes and surface areas) were monitored to represent the variability of the study area. They can be grouped into four classes of land cover: 39 grassland areas, 100 plots of wheat, 13 plots of barley, 20 plots of rapeseed, and 2 classes of bare soil: 23 plots of small roughness and 16 plots of medium roughness. Satellite radar images in the X-, C- and L-bands (HH polarization) were acquired between 14 and 18 April 2010. Optical images delivered by Formosat-2 and corresponding field data were acquired on 14 April 2010.The results show that combining images acquired in the L-band (Alos) and the optical range (Formosat-2) improves the classification performance (overall accuracy = 0.85, kappa = 0.81) compared to the use of radar or optical data alone. The results obtained for the various types of land cover show performance levels and confusions related to the phenological stage of the species studied, with the geometry of the cover, the roughness states of the surfaces, etc. Performance is also related to the wavelength and penetration depth of the signal providing the images. Thus, the results show that the quality of the classification often increases with increasing wavelength of the images used.     

Evaluating the potential of the red edge channel for C3 (Festuca spp.) grass discrimination using Sentinel-2 and Rapid Eye satellite image data

Integrating the Red Edge channel in satellite sensors is valuable for plant species discrimination. Sentinel-2 MSI and Rapid Eye are some of the new generation satellite sensors that are characterized by finer spatial and spectral resolution, including the red edge band. The aim of this study was to evaluate the potential of the red edge band of Sentinel-2 and Rapid Eye, for mapping festuca C3 grass using discriminant analysis and maximum likelihood classification algorithms. Spectral bands, vegetation indices and spectral bands plus vegetation indices were analysed. Results show that the integration of the red edge band improved the festuca C3 grass mapping accuracy by 5.95 and 4.76% for Sentinel-2 and Rapid Eye when the red edge bands were included and excluded in the analysis, respectively. The results demonstrate that the use of sensors with strategically positioned red edge bands, could offer information that is critical for the sustainable rangeland management.     

Mapping the distribution of the main host for plague in a complex landscape in Kazakhstan: An object-based approach using SPOT-5 XS,Landsat 7 ETM+, SRTM and multiple Random Forests

Plague is a zoonotic infectious disease present in great gerbil populations in Kazakhstan. Infectious disease dynamics are influenced by the spatial distribution of the carriers (hosts) of the disease. The great gerbil, the main host in our study area, lives in burrows, which can be recognized on high resolution satellite imagery. In this study, using earth observation data at various spatial scales, we map the spatial distribution of burrows in a semi-desert landscape.The study area consists of various landscape types. To evaluate whether identification of burrows by classification is possible in these landscape types, the study area was subdivided into eight landscape units, on the basis of Landsat 7 ETM+ derived Tasselled Cap Greenness and Brightness, and SRTM derived standard deviation in elevation.In the field, 904 burrows were mapped. Using two segmented 2.5 m resolution SPOT-5 XS satellite scenes, reference object sets were created. Random Forests were built for both SPOT scenes and used to classify the images. Additionally, a stratified classification was carried out, by building separate Random Forests per landscape unit.Burrows were successfully classified in all landscape units. In the ‘steppe on floodplain’ areas, classification worked best: producer's and user's accuracy in those areas reached 88% and 100%, respectively. In the ‘floodplain’ areas with a more heterogeneous vegetation cover, classification worked least well; there, accuracies were 86 and 58% respectively. Stratified classification improved the results in all landscape units where comparison was possible (four), increasing kappa coefficients by 13, 10, 9 and 1%, respectively.In this study, an innovative stratification method using high- and medium resolution imagery was applied in order to map host distribution on a large spatial scale. The burrow maps we developed will help to detect changes in the distribution of great gerbil populations and, moreover, serve as a unique empirical data set which can be used as input for epidemiological plague models. This is an important step in understanding the dynamics of plague.