Using local transition probability models in Markov random fields for forest change detection

Change detection based on the comparison of independently classified images (i.e. post-classification comparison) is well-known to be negatively affected by classification errors of individual maps. Incorporating spatial-temporal contextual information in the classification helps to reduce the classification errors, thus improving change detection results. In this paper, spatial-temporal Markov Random Fields (MRF) models were used to integrate spatial-temporal information with spectral information for multi-temporal classification in an attempt to mitigate the impacts of classification errors on change detection. One important component in spatial-temporal MRF models is the specification of transition probabilities. Traditionally, a global transition probability model is used that assumes spatial stationarity of transition probabilities across an image scene, which may be invalid if areas have varying transition probabilities. By relaxing the stationarity assumption, we developed two local transition probability models to make the transition model locally adaptive to spatially varying transition probabilities. The first model called locally adjusted global transition model adapts to the local variation by multiplying a pixel-wise probability of change with the global transition model. The second model called pixel-wise transition model was developed as a fully local model based on the estimation of the pixel-wise joint probabilities. When applied to the forest change detection in Paraguay, the two local models showed significant improvements in the accuracy of identifying the change from forest to non-forest compared with traditional models. This indicates that the local transition probability models can present temporal information more accurately in change detection algorithms based on spatial-temporal classification of multi-temporal images. The comparison between the two local transition models showed that the fully local model better captured the spatial heterogeneity of the transition probabilities and achieved more stable and consistent results over different regions of a large image scene.     

A change detection model based on neighborhood correlation image analysis and decision tree classification

This study introduces a change detection model based on Neighborhood Correlation Image (NCI) logic. It is based on the fact that the same geographic area (e.g., a 3 × 3 pixel window) on two dates of imagery will tend to be highly correlated if little change has occurred, and uncorrelated when change occurs. Computing the piecewise correlation between two data sets provides valuable information regarding the location and numeric change value derived using contextual information within the specified neighborhood. Various neighborhood configurations (i.e., multi-level NCIs) were explored in the study using high spatial resolution multispectral imagery: smaller neighborhood sizes provided some detailed change information (such as a new patios added to an existing building) at the cost of introducing some noise (such as changes in shadows). Larger neighborhood sizes were useful for removing this noise but introduced some inaccurate change information (such as removing some linear feature changes). When combined with image classification using a machine learning decision tree (C5.0), classifications based on multi-level NCIs yielded superior results (e.g., using a 3-pixel circular radius neighborhood had a Kappa of 0.94), compared to the classification that did not incorporate NCIs (Kappa = 0.86).     

Using random walker for knowledge transfer in classifying multi-temporal VHR images

Acquiring land cover types from very high resolution (VHR) images is of great significance to many applications and has been intensively studied for many years. The difficulties in image classification and the high frequencies of remote sensing image acquisition make it urgent to develop efficient knowledge transfer approaches for understanding multi-temporal VHR images. This letter proposed a knowledge transfer approach that uses the label information of the existing VHR images to classify multi-temporal images. The approach was implemented in three steps: object-based change detection, knowledge transfer of label information, and random walker (RW) classification. The proposed approach was tested by two datasets with each having two temporal images acquired on the same geographical areas. The experimental results showed that the proposed approach outperformed the support vector machine (SVM) algorithm in classifying multi-temporal images and can reduce the influence of spectral confusions on image classification.     

Improved land surface emissivities over agricultural areas using ASTER NDVI

Land surface emissivity retrieval over agricultural regions is important for energy balance estimations, land cover assessment and other related environmental studies. The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) produces images of sufficient spatial resolution (from 15 m to 90 m) to be of use in agricultural studies, in which fields of crops are too small to be well-resolved by low resolution sensors. The ASTER project generates land surface emissivity images as a Standard Product (AST05) using the Temperature/Emissivity Separation (TES) algorithm. However, the TES algorithm is prone to scaling errors in estimating emissivities for surfaces with low spectral contrast if the atmospheric correction is inaccurate. This paper shows a comparison between the land surface emissivity estimated with the TES algorithm and from a simple approach using the Normalized Difference Vegetation Index (NDVI) for five ASTER images (28 June 2000, 15 August 2000, 31 August 2000, 28 April 2001 and 02 August 2001) of the agricultural area of Barrax (Albacete, Spain). The results indicate that differences are < 1% for ASTER band 13 (10.7 μm) and < 1.5% for band 14 (11.3 μm), but > 2% for bands 10 (8.3 μm), 11 (8.6 μm) and 12 (9.1 μm). The emissivities for the five ASTER bands were tested against in situ measurements carried out with the CIMEL CE 312-2 field radiometer, the NDVI method giving root mean square errors (RMSE) < 0.005 over vegetated areas and RMSE < 0.015 over bare soil, and the TES algorithm giving RMSE ∼ 0.01 for vegetated areas but RMSE > 0.03 over bare soil. The errors and inconsistencies for ASTER bands 13 and 14 are within those anticipated for TES, but the greater errors for bands 10-12 suggest the presence of problems related to atmospheric compensation and model assumptions about soil spectra. The NDVI method uses visible/near-infrared data co-acquired with the thermal images to estimate vegetation cover and, hence, provides an independent constraint on emissivity. The success of this approach suggests that it may be useful for daytime images of agricultural or other heavily vegetated areas, in which the TES algorithm has occasional failures.     

Impact of lossy compression on mapping crop areas from remote sensing

This study measures the effect of lossy image compression (JPEG 2000 and JPG) on the digital classification of crop areas. The results provide new insights into the influence of compression on the quality of the cartography produced. Both a multitemporal and a single-date classification approach were analysed. With the multitemporal approach, it is possible to use high compression ratios (CRs), up to 20:1 or even 100:1, and the overall accuracy of the classification is similar to that obtained with the original images. Moreover, the classified area is similar or even greater (fewer pixels are uncertain). For a single-date approach, it is only advisable to use 3D-JPEG 2000 at CRs up to 20:1. The optimum CR is also affected by landscape fragmentation (fragmented images tolerate less compression) and the classification method (hybrid classifiers are affected less than the maximum likelihood and minimum distance classifiers). Finally, classifications from compressed images have less ‘salt and pepper’ effect than those obtained from the originals, especially when JPEG 2000 (3D or not) is used.     

Hidden Markov Models for crop recognition in remote sensing image sequences

This work proposes a Hidden Markov Model (HMM) based technique to classify agricultural crops. The method uses HMM to relate the varying spectral response along the crop cycle with plant phenology, for different crop classes, and recognizes different agricultural crops by analyzing their spectral profiles over a sequence of images. The method assigns each image segment to the crop class whose corresponding HMM delivers the highest probability of emitting the observed sequence of spectral values. Experimental analysis was conducted upon a set of 12 co-registered and radiometrically corrected LANDSAT images of region in southeast Brazil, of approximately 124.100 ha, acquired between 2002 and 2004. Reference data was provided by visual classification, validated through extensive field work. The HMM-based method achieved 93% average class accuracy in the identification of the correct crop, being, respectively, 10% and 26% superior to multi-date and single-date alternative approaches applied to the same data set.     

Generation of dense time series synthetic Landsat data through data blending with MODIS using a spatial and temporal adaptive reflectance fusion model

Landsat imagery with a 30 m spatial resolution is well suited for characterizing landscape-level forest structure and dynamics. While Landsat images have advantageous spatial and spectral characteristics for describing vegetation properties, the Landsat sensor's revisit rate, or the temporal resolution of the data, is 16 days. When considering that cloud cover may impact any given acquisition, this lengthy revisit rate often results in a dearth of imagery for a desired time interval (e.g., month, growing season, or year) especially for areas at higher latitudes with shorter growing seasons. In contrast, MODIS (MODerate-resolution Imaging Spectroradiometer) has a high temporal resolution, covering the Earth up to multiple times per day, and depending on the spectral characteristics of interest, MODIS data have spatial resolutions of 250 m, 500 m, and 1000 m. By combining Landsat and MODIS data, we are able to capitalize on the spatial detail of Landsat and the temporal regularity of MODIS acquisitions. In this research, we apply and demonstrate a data fusion approach (Spatial and Temporal Adaptive Reflectance Fusion Model, STARFM) at a mainly coniferous study area in central British Columbia, Canada. Reflectance data for selected MODIS channels, all of which were resampled to 500 m, and Landsat (at 30 m) were combined to produce 18 synthetic Landsat images encompassing the 2001 growing season (May to October). We compared, on a channel-by-channel basis, the surface reflectance values (stratified by broad land cover types) of four real Landsat images with the corresponding closest date of synthetic Landsat imagery, and found no significant difference between real (observed) and synthetic (predicted) reflectance values (mean difference in reflectance: mixed forest x? = 0.086, σ = 0.088, broadleaf x? = 0.019, σ = 0.079, coniferous x? = 0.039, σ = 0.093). Similarly, a pixel based analysis shows that predicted and observed reflectance values for the four Landsat dates were closely related (mean r² = 0.76 for the NIR band; r² = 0.54 for the red band; p < 0.01). Investigating the trend in NDVI values in synthetic Landsat values over a growing season revealed that phenological patterns were well captured; however, when seasonal differences lead to a change in land cover (i.e., disturbance, snow cover), the algorithm used to generate the synthetic Landsat images was, as expected, less effective at predicting reflectance.     

Synergistic use of QuickBird multispectral imagery and LIDAR data for object-based forest species classification

This study evaluated the synergistic use of high spatial resolution multispectral imagery (i.e., QuickBird, 2.4 m) and low-posting-density LIDAR data (3 m) for forest species classification using an object-based approach. The integration of QuickBird multispectral imagery and LIDAR data was considered during image segmentation and the subsequent object-based classification. Three segmentation schemes were examined: (1) segmentation based solely on the spectral image layers; (2) segmentation based solely on LIDAR-derived layers; and (3) segmentation based on both the spectral and LIDAR-derived layers. For each segmentation scheme, objects were generated at twelve different scales in order to determine optimal scale parameters. Six categories of classification metrics were generated for each object based on spectral data alone, LIDAR data alone and the combination of both data sources. Machine learning decision trees were used to build classification rule sets. Quantitative segmentation quality assessment and classification accuracy results showed the integration of spectral and LIDAR data, in both image segmentation and object-based classification, improved the forest classification compared to using either data source independently. Better segmentation quality led to higher classification accuracy. The highest classification accuracy (Kappa = 91.6%) was acquired when using both spectral- and LIDAR-derived metrics based on objects segmented from both spectral and LIDAR layers at scale parameter 250, where best segmentation quality was achieved. Optimal scales were analyzed for each segmentation-classification scheme. Statistical analysis of classification accuracies at different scales revealed that there was a range of optimal scales that provided statistically similar accuracy.     

Comparison of three SeaWiFS atmospheric correction algorithms for turbid waters using AERONET-OC measurements

The use of satellites to monitor the color of the ocean requires effective removal of the atmospheric signal. This can be performed by extrapolating the aerosol optical properties in the visible from the near-infrared (NIR) spectral region assuming that the seawater is totally absorbant in this latter part of the spectrum. However, the non-negligible water-leaving radiance in the NIR which is characteristic of turbid waters may lead to an overestimate of the atmospheric radiance in the whole visible spectrum with increasing severity at shorter wavelengths. This may result in significant errors, if not complete failure, of various algorithms for the retrieval of chlorophyll-a concentration, inherent optical properties and biogeochemical parameters of surface waters.This paper presents results of an inter-comparison study of three methods that compensate for NIR water-leaving radiances and that are based on very different hypothesis: 1) the standard SeaWiFS algorithm (Stumpf et al., 2003; Bailey et al., 2010) based on a bio-optical model and an iterative process; 2) the algorithm developed by Ruddick et al. (2000) based on the spatial homogeneity of the NIR ratios of the aerosol and water-leaving radiances; and 3) the algorithm of Kuchinke et al. (2009) based on a fully coupled atmosphere-ocean spectral optimization inversion. They are compared using normalized water-leaving radiance nL_w in the visible. The reference source for comparison is ground-based measurements from three AERONET-Ocean Color sites, one in the Adriatic Sea and two in the East Coast of USA.Based on the matchup exercise, the best overall estimates of the nL_w are obtained with the latest SeaWiFS standard algorithm version with relative error varying from 14.97% to 35.27% for λ = 490 nm and λ = 670 nm respectively. The least accurate estimates are given by the algorithm of Ruddick, the relative errors being between 16.36% and 42.92% for λ = 490 nm and λ = 412 nm, respectively. The algorithm of Kuchinke appears to be the most accurate algorithm at 412 nm (30.02%), 510 (15.54%) and 670 nm (32.32%) using its default optimization and bio-optical model coefficient settings.Similar conclusions are obtained for the aerosol optical properties (aerosol optical thickness τ(865) and the Ångström exponent, α(510, 865)). Those parameters are retrieved more accurately with the SeaWiFS standard algorithm (relative error of 33% and 54.15% for τ(865) and α(510, 865)).A detailed analysis of the hypotheses of the methods is given for explaining the differences between the algorithms. The determination of the aerosol parameters is critical for the algorithm of Ruddick et al. (2000) while the bio-optical model is critical for the algorithm of Stumpf et al. (2003) utilized in the standard SeaWiFS atmospheric correction and both aerosol and bio-optical model for the coupled atmospheric-ocean algorithm of Kuchinke. The Kuchinke algorithm presents model aerosol-size distributions that differ from real aerosol-size distribution pertaining to the measurements. In conclusion, the results show that for the given atmospheric and oceanic conditions of this study, the SeaWiFS atmospheric correction algorithm is most appropriate for estimating the marine and aerosol parameters in the given turbid waters regions.     

基于智能CASE库的多时相SAR影像分类方法-土地利用及变化监测

针对SAR影像分类,提出了一种基于智能案例(CASE)库多时相SAR影像分类方法。该方法主要分为4部分:SAR影像预处理;智能CASE的建构;基于CASE相似度匹配的SAR影像分类;分类后处理。在智能CASE建构期间,引入时空分析技术去除“伪”CASE,从而保证了CASE库中CASE信息的可靠性。接着,在基于CASE匹配的SAR影像分类过程中,采用分层相似度评价的方法,消除CASE特征相互之间的混叠效应。最后,采用面向对象的方法进行影像分类后处理。该方法有效地考虑了分类地块的形状因子,使分类结果更精确、更符合逻辑性。以2000年(4景,包含4个季度)和2004年(3景,包含3个季度)的多时相SAR影像作为实验数据,结果表明,使用我们提出的方法能达到较好的SAR影像分类结果,分类总体精度达到85%～90%,这为利用多时相SAR影像实施土地利用和变化监测(Land Use and Land Cover Change,LULC)奠定了良好基础。     

A Resource Aware MapReduce Based Parallel SVM for Large Scale Image Classifications

Machine learning techniques have facilitated image retrieval by automatically classifying and annotating images with keywords. Among them support vector machines (SVMs) are used extensively due to their generalization properties. However, SVM training is notably a computationally intensive process especially when the training dataset is large. This paper presents RASMO, a resource aware MapReduce based parallel SVM algorithm for large scale image classifications which partitions the training data set into smaller subsets and optimizes SVM training in parallel using a cluster of computers. A genetic algorithm based load balancing scheme is designed to optimize the performance of RASMO in heterogeneous computing environments. RASMO is evaluated in both experimental and simulation environments. The results show that the parallel SVM algorithm reduces the training time significantly compared with the sequential SMO algorithm while maintaining a high level of accuracy in classifications.     

Effects of spatial and spectral resolution in estimating ecosystem α-diversity by satellite imagery

Remote sensing represents a powerful tool to derive quantitative and qualitative information about ecosystem biodiversity. In particular, since plant species richness is a fundamental indicator of biodiversity at the community and regional scales, attempts were made to predict species richness (spatial heterogeneity) by means of spectral heterogeneity. The possibility of using spectral variance of satellite images for predicting species richness is known as Spectral Variation Hypothesis. However, when using remotely sensed data, researchers are limited to specific scales of investigation. This paper aims to investigate the effects of scale (both as spatial and spectral resolution) when searching for a relation between spectral and spatial (related to plant species richness) heterogeneity, by using satellite data with different spatial and spectral resolution. Species composition was sampled within square plots of 100 m² nested in macroplots of 10,000 m². Spectral heterogeneity of each macroplot was calculated using satellite images with different spatial and spectral resolution: a Quickbird multispectral image (4 bands, spatial resolution of 3 m), an Aster multispectral image (first 9 bands used, spatial resolution of 15 m for bands 1 to 3 and 30 m for bands 4 to 9), an ortho-Landsat ETM+ multispectral image (bands 1 to 5 and band 7 used; spatial resolution, 30 m), a resampled 60 m Landsat ETM+ image.Quickbird image heterogeneity showed a statistically highly significant correlation with species richness (r = 0.69) while coarse resolution images showed contrasting results (r = 0.43, r = 0.67, and r = 0.69 considering the Aster, Landsat ETM+, and the resampled 60 m Landsat ETM+ images respectively). It should be stressed that spectral variability is scene and sensor dependent. Considering coarser spatial resolution images, in such a case even using SWIR Aster bands (i.e. the additional spectral information with respect to Quickbird image) such an image showed a very low power in catching spectral and thus spatial variability with respect to Landsat ETM+ imagery. Obviously coarser resolution data tend to have mixed pixel problems and hence less sensitive to spatial complexity. Thus, one might argue that using a finer pixel dimension should inevitably result in a higher level of detail. On the other hand, the spectral response from different land-cover features (and thus different species) in images with higher spectral resolution should exhibit higher complexity.Spectral Variation Hypothesis could be a basis for improving sampling designs and strategies for species inventory fieldwork. However, researchers must be aware on scale effects when measuring spectral (and thus spatial) heterogeneity and relating it to field data, hence considering the concept of scale not only related to a spatial framework but even to a spectral one.     

Evaluation of annual forest disturbance monitoring using a static decision tree approach and 250 m MODIS data

Research on change detected has largely focused on method development and evaluation in a temporally dependent manner where training and validation data are from the same temporal period. Monitoring over several change periods needs to account for increased variability resulting from possible combinations of atmosphere, sensor, and surface conditions. Training a change method for each monitoring period (i.e. a dynamic approach) is an option, but can be costly to develop the needed training datasets and may not be warranted if sufficient accuracy can be obtained without retraining (i.e. a static approach). In this research the potential of change detection using a static approach suitable for near-real time annual monitoring was evaluated. The research assessed the influence of feature set size, radiometric normalization, incorporation of temporal information, and change object size and sub-pixel fraction on accuracy. The static approach was based on a decision tree developed using 250 m MODIS data from 2005 to 2006 and applied annually for the period 2001-2005. Change results between years were combined and compared to reference data representing change from 2001 to 2005 to evaluate monitoring performance. Results revealed high accuracy for the decision tree change model development from 2005 to 2006 (bootstrap cross-validation KAPPA = 0.91), with lower accuracy (KAPPA = 0.80) when applied for monitoring from 2001 to 2005. Radiometric normalization increased monitoring accuracy (KAPPA = 0.86). Further improvement was achieved with the incorporation of temporal contextual tests to combine the 2001-2005 inter-annual change maps (KAPPA = 0.90), but required a time lag of 1 year. An alternative temporal test that was not restricted by the 1 year time lag produced slightly lower accuracy (KAPPA = 0.88). Evaluation of the effect of object size on detection accuracy showed that accuracy for objects less than 7 pixels was strongly related to object size, with objects less than 3 pixels having low detection rates. The effect of sub-pixel change fraction was found to be dependent on object size with larger objects reducing detection error across the range of fractions evaluated.     

A colorization algorithm based on local MAP estimation

This paper presents a colorization algorithm which adds color to monochrome images. In this paper, the colorization problem is formulated as the maximum a posteriori (MAP) estimation of a color image given a monochrome image. Markov random field (MRF) is used for modeling a color image which is utilized as a prior for the MAP estimation. The MAP estimation problem for a whole image is decomposed into local MAP estimation problems for each pixel. Using 0.6% of whole pixels as references, the proposed method produced pretty high quality color images with 25.7-32.6 dB PSNR values for eight images.     

Mapping burned areas from Landsat TM/ETM+ data with a two-phase algorithm: Balancing omission and commission errors

Maps of burned area have been obtained from an automatic algorithm applied to a multitemporal series of Landsat TM/ETM+ images in two Mediterranean sites. The proposed algorithm is based on two phases: the first one intends to detect the more severely burned areas and minimize commission errors. The second phase improves burned patches delimitation using a hybrid contextual algorithm based on logistic regression analysis, and tries to minimize omission errors. The algorithm was calibrated using six study sites and it was validated for the whole territory of Portugal (89,000 km²) and for Southern California (70,000 km²). In the validation exercise, 65 TM/ETM+ scenes for Portugal and 35 for California were used, all from the 2003 fire season. A good agreement with the official burned area perimeters was shown, with kappa values close to 0.85 and low omission and commission errors (< 16.5%). The proposed algorithm could be operationally used for historical mapping of burned areas from Landsat images, as well as from future medium resolution sensors, providing they acquire images in two bands of the Short Wave Infrared (1.5-2.2 μm).     

Segmentation of brain MR images using a proper combination of DCS based method with MRF

Manual segmentation of Magnetic Resonance Images (MRI) is a time-consuming process, thus automatic segmentation of brain MR images has attracted more attention in recent years. In this paper, we introduce Dynamic Classifier Selection Markov Random Field (DCSMRF) algorithm for supervised segmentation of brain MR images into three main tissues such as White Matter (WM), Gray Matter (GM) and Cerebrospinal Fluid (CSF). DCSMRF combines a novel ensemble method with the Markov Random Field (MRF) algorithm and tries to obtain the advantages of both algorithms. For the ensemble part of DCSMRF, we propose an ensemble method called Dynamic Classifier System-Weighted Local Accuracy (DCS-WLA) which is a type of Combination of Multiple Classifier (CMC) algorithm. Later, the MRF algorithm is utilized for incorporating spatial, contextual and textural information in this paper. For the MRF section, an energy function based on the output of the DCS-WLA algorithm is proposed, then maximum value for Maximum A Posterior (MAP) criterion is searched to obtain optimal segmentation. The MRF algorithm applies similar to a post processing step in which only a subset of pixels is selected for optimization step. Hence, a vast amount of search space is pruned. Consequently, the computational burden of the proposed algorithm is more tolerable than the conventional MRF-based methods. Moreover, by employing ensemble algorithms, the accuracy and reliability of final results are enhanced compared to the individual methods.     

面向行人移除的街景图像全局优化时空修补方法

针对街景图像中往往包含大量行人等隐私对象的问题,以移除图像中的行人为例,提出一种全局优化的时空图像修补方法.首先利用运动获得结构算法建立参考图像与目标图像之间的对应关系,该过程不依赖场景的简化假设,使得该图像修补方法适合各类复杂场景;然后对待修补区域建立马尔科夫随机场,通过合理设计标号集和能耗函数,把时域和空域修补结合到同一优化过程中,并自动判断何时选择何种修补方式,使修补结果尽量符合实际场景同时又具有较好的视觉一致性.大量实验结果表明,该方法对各种复杂场景的街景图像都能够得到较好的修补效果.     

Mapping biomass and stress in the Sierra Nevada using lidar and hyperspectral data fusion

In this paper, we explored fusion of structural metrics from the Laser Vegetation Imaging Sensor (LVIS) and spectral characteristics from the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) for biomass estimation in the Sierra Nevada. In addition, we combined the two sensors to map species-specific biomass and stress at landscape scale. Multiple endmember spectral mixture analysis (MESMA) was used to classify vegetation from AVIRIS images and obtain sub-pixel fractions of green vegetation, non-photosynthetic vegetation, soil, and shade. LVIS metrics, AVIRIS spectral indices, and MESMA fractions were compared with field measures of biomass using linear and stepwise regressions at stand (1 ha) level. AVIRIS metrics such as water band indices and shade fractions showed strong correlation with LVIS canopy height (r² = 0.69, RMSE = 5.2 m) and explained around 60% variability in biomass. LVIS variables were found to be consistently good predictors of total and species specific biomass (r² = 0.77, RMSE = 70.12 Mg/ha). Prediction by LVIS after species stratification of field data reduced errors by 12% (r² = 0.84, RMSE = 58.78 Mg/ha) over using LVIS metrics alone. Species-specific biomass maps and associated errors created from fusion were different from those produced without fusion, particularly for hardwoods and pines, although mean biomass differences between the two techniques were not statistically significant. A combined analysis of spatial maps from LVIS and AVIRIS showed increased water and chlorophyll stress in several high biomass stands in the study area. This study provides further evidence that lidar is better suited for biomass estimation, per se, while the best use of hyperspectral data may be to refine biomass predictions through a priori species stratification, while also providing information on canopy state, such as stress. Together, the two sensors have many potential applications in carbon dynamics, ecological and habitat studies.     

An effective hybrid approach to remote-sensing image classification

This article presents a hybrid fuzzy classifier for effective land-use/land-cover (LULC) mapping. It discusses a Bayesian method of incorporating spatial contextual information into the fuzzy noise classifier (FNC). The FNC was chosen as it detects noise using spectral information more efficiently than its fuzzy counterparts. The spatial information at the level of the second-order pixel neighbourhood was modelled using Markov random fields (MRFs). Spatial contextual information was added to the MRF using different adaptive interaction functions. These help to avoid over-smoothing at the class boundaries. The hybrid classifier was applied to advanced wide-field sensor (AWiFS) and linear imaging self-scanning sensor-III (LISS-III) images from a rural area in India. Validation was done with a LISS-IV image from the same area. The highest increase in accuracy among the adaptive functions was 4.1% and 2.1% for AWiFS and LISS-III images, respectively. The paper concludes that incorporation of spatial contextual information into the fuzzy noise classifier helps in achieving a more realistic and accurate classification of satellite images.     

A MapReduce-based distributed SVM algorithm for automatic image annotation

Machine learning techniques have facilitated image retrieval by automatically classifying and annotating images with keywords. Among them Support Vector Machines (SVMs) have been used extensively due to their generalization properties. However, SVM training is notably a computationally intensive process especially when the training dataset is large. This paper presents MRSMO, a MapReduce based distributed SVM algorithm for automatic image annotation. The performance of the MRSMO algorithm is evaluated in an experimental environment. By partitioning the training dataset into smaller subsets and optimizing the partitioned subsets across a cluster of computers, the MRSMO algorithm reduces the training time significantly while maintaining a high level of accuracy in both binary and multiclass classifications.