ON IMAGE CLASSIFICATION: CITY IMAGES VS. LANDSCAPES

Grouping images into semantically meaningful categories using low-level visual features is a challenging and important problem in content-based image retrieval. Based on these groupings, effective indices can be built for an image database. In this paper, we show how a specific high-level classification problem (city images vs landscapes) can be solved from relatively simple low-level features geared for the particular classes. We have developed a procedure to qualitatively measure the saliency of a feature towards a classification problem based on the plot of the intra-class and inter-class distance distributions. We use this approach to determine the discriminative power of the following features: color histogram, color coherence vector, DCT coefficient, edge direction histogram, and edge direction coherence vector. We determine that the edge direction-based features have the most discriminative power for the classification problem of interest here. A weighted k-NN classifier is used for the classification which results in an accuracy of 93.9% when evaluated on an image database of 2716 images using the leave-one-out method. This approach has been extended to further classify 528 landscape images into forests, mountains, and sunset/sunrise classes. First, the input images are classified as sunset/sunrise images vs forest & mountain images (94.5% accuracy) and then the forest & mountain images are classified as forest images or mountain images (91.7% accuracy). We are currently identifying further semantic classes to assign to images as well as extracting low level features which are salient for these classes. Our final goal is to combine multiple 2-class classifiers into a single hierarchical classifier.     

TM digital processing of a tropical forest region in southeastern Mexico

A humid forest in the neotropical area of Los Tuxtlas, in southeastern Mexico has been used as a test area (900km²) for classification of landscape and vegetation by means of Landsat Thematic Mapper (TM) data, aerial photography and 103 ground samples. The area presents altitudinal variations from sea level to 1640m, providing a wide variety of vegetation types. A hybrid (supervised/unsupervised) classification approach was used, defining spectral signatures for 14 clustering areas with data from the reflective bands of the TM. The selected clustering areas ranged from vegetation of the highlands and the rain forest to grassland, barren soil, crops and secondary vegetation. The digital classification compared favourably with results from aerial photography and with those from a multivariate analysis of the 103 ground data. The statistical evaluation (error matrix) of the classified image indicated an overall 84·4 per cent accuracy with a kappa coefficient of agreement of 0·83. A geographical information system (GIS) was used to compile a land unit and a vegetation map. The TM data allowed for delineation of boundaries in the land unit map, and for a finer differentiation of vegetation types than those identified during field work. Digital value patterns of several information classes are shown and discussed as an indirect guide of the spectral behaviour of vegetation of highlands, rain forest, secondary vegetation and crops. The method is considered applicable to the inventory of other forested areas, especially those with significant variations in vegetation.     

Spatial pattern analysis for monitoring protected areas

The management of diverse biota within protected areas is affected by both land cover change within a protected area and habitat loss and fragmentation in the surrounding landscape. Satellite images provide a synoptic view of land cover patterns, but the use of such imagery requires careful consideration of sensor type, resolution, extent, and the metrics used to quantify ecologically significant change. We examined these factors for landscape monitoring applications in four small National Parks near Washington, DC: Antietam National Battlefield, Catoctin Mountain Park, Prince William Forest Park and Rock Creek Park. Using 4 m Ikonos, 10 m SPOT, 15 m pan-sharpened Landsat ETM+ and 30 m Landsat ETM+ imagery, the parks and surrounding areas were mapped to National Land Cover system classes. For each park, we examined four methods for defining map extent, including park administrative boundaries, two variable buffer widths, and watershed boundaries, and then analyzed patterns of forest habitat for the maps using a graph theoretic approach (critical dispersal threshold distance) and common landscape metrics (number of patches, percent forest, forest edge density, and forest area-weighted mean patch size). As expected, landscape metrics for maps derived at differing resolutions varied significantly, but map extent often yielded even larger differences. We found that for most applications, coarser scale data (e.g., 30 m Landsat) are adequate for characterizing landscape pattern, although ultimately data from multiple sensors may be appropriate or necessary based on different objectives of landscape monitoring (e.g., mapping single trees vs. forest stands) and the scale at which a resource of interest interacts with the larger landscape (e.g., birds vs. herptiles). Our results provide a strong caution regarding the practical issues associated with combining data sources from multiple satellite sensors for monitoring applications.     

Aerial tracking of elongated objects in rural environments

In this paper, a new approach to finding and tracking various land cover boundaries such as rivers, agricultural fields, channels and roads for use in visual navigation system of an unmanned aerial vehicle is presented. We use a combination of statistical estimation and optimization techniques for extraction of dominant boundaries in noisy aerial images. A set of perceptual grouping restrictions is used to connect the acquired piecewise boundaries and to find the heading direction of the main boundary. The results are further refined by applying a set of texture and colour cues and eliminating any false hypothesis. To reduce the computation requirements, another approach based on sampled colour values of different land covers is also investigated. Colour characteristics of a set of manually selected windows are compared to select the best attributes needed for discrimination between different land covers in various (natural) lighting conditions. Each frame is then partially scanned and desired environmental features are extracted and classified. The results show that the proposed technique meets the minimum speed and accuracy requirement of aforementioned application and outperforms single-feature object tracking algorithms.

A quantitative assessment of boundaries in automated forest stand delineation using very high resolution imagery

Based on very high resolution satellite images, object-based classification methods can be used to produce large scale maps for forest management. These new products require a method to derive quantitative information about the accuracy and precision of delineated boundaries. This assessment would complement any measure of thematic accuracy derived from the confusion matrix. This study aims to assess the positional quality of the boundaries between different landscape units produced by automated segmentation of IKONOS and SPOT-5 satellite images over temperate forests. A robust method was developed to assess the accuracy and the precision of the forest boundaries, respectively measured by the bias and the standard deviation. The two main sources of positional error, namely residual parallax and automatic segmentation, were independently assessed. Positional errors caused by the residual parallax were quantified using a 3D model. Forest stand boundaries generated by automatic segmentation were compared to corresponding visual delineations. The results showed that residual parallax was the major source of positive bias (area overestimation) along forest/non-forest boundaries and depended on the interactions between forest stand patterns and sensor viewing angles. Due mainly to tree shade, the automatic segmentation also produced a positive bias on forest areas, which remained under 1 m for both IKONOS-2 and SPOT-5 images. Standard deviation did not increase linearly with pixel size and was influenced by the nature of the boundary. Production of 1:20,000 scale forest maps from very high resolution satellite data clearly requires acquisition of near nadir imagery or knowledge of landscape object height for true orthorectification. In these cases, IKONOS-2 segmentation outputs were found to correspond with 1:20,000 scale map specification, and SPOT-5 imagery with 1:30,000 scale.     

Performance of different spectral and textural aerial photograph features in multi-source forest inventory

Most multi-source forest inventory (MSFI) applications have thus far been based on the use of medium resolution satellite imagery, such as Landsat TM. The high plot and stand level estimation errors of these applications have, however, restricted their use in forest management planning. One reason suggested for the high estimation errors has been the coarse spatial resolution of the imagery employed. Therefore, very high spatial resolution (VHR) imagery sources provide interesting data for stand-level inventory applications. However, digital interpretation of VHR imagery, such as aerial photographs, is more complicated than the use of traditional satellite imagery. Pixel-by-pixel analysis is not applicable to VHR imagery because a single pixel is small in relation to the object of interest, i.e. a forest stand, and therefore it does not adequately represent the spectral properties of a stand. Additionally in aerial photographs, the spectral properties of the objects are dependent on their location in the image. Therefore, MSFI applications based on aerial imagery must employ features that are less sensitive to their location in the image and that have been derived using the spatial neighborhood of each pixel, e.g. a square-shaped window of pixels. In this experiment several spectral and textural features were extracted from color-infrared aerial photographs and employed in estimation of forest attributes. The features were extracted from original, normalized difference vegetation index and channel ratio images. The correlations between the extracted image features and forest attributes measured from sample plots were examined. Additionally, the spectral and textural features were used for estimating the forest attributes of sample plots, applying the k nearest neighbor estimation method. The results show that several spectral and textural image features that are moderately or well correlated with the forest attributes. Furthermore, the accuracy of forest attribute estimation can be significantly improved by a careful selection of image features.     

Geometric discriminative features for aerial image retrieval in social media

The aerial image recognition is an important problem in multimedia information retrieval in social media. In this paper, we propose a new approach by integrating aerial image’s local features into a discriminative one which reflects both the geometric property and the color distribution of aerial image. Firstly, each aerial image is segmented into several regions in terms of their color intensities. And region connected graph (RCG), the links between the spatial neighboring regions, is presented to encode the spatial context of aerial images. Secondly, we mine frequent structures in the RCGs corresponding to training aerial images collected from social media. And a set of refined structures are selected among the frequent ones towards being more discriminative and less redundant. Finally, given a new aerial image, its sub-RCGs corresponding to all the refined structures are extracted and quantized into a discriminative feature for aerial image recognition. The experimental results validate the proposed method by providing a more accurate recognition result of the aerial images on different datasets from different social medias.     

Mapping vegetation in a late Quaternary landform of the Amazonian wetlands using object-based image analysis and decision tree classification

Fan-shaped morphologies related to late Quaternary residual megafan depositional systems are common features over wide areas in northern Amazonia. These features were formed by ancient distributary drainage systems that are in great contrast to tributary drainage networks that typify the modern Amazon basin. The surfaces of the Amazonian megafans constitute vegetacional mosaic wetlands with different campinarana types. A fine-scale-resolution investigation is required to provide detailed classification maps for the various campinarana and surrounding forest types associated with the Amazonian megafans. This approach remains to be presented, despite its relevance for analysing the relationship between stages of plant succession and sedimentary dynamics associated with the evolution of megafans. In this work, we develop a methodology for classifying vegetation over a fan-shaped megafan palaeoform from a northern Amazonian wetland. The approach included object-based image analysis (OBIA) and data-mining (DM) techniques combining Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images, land-cover fractions derived by the linear spectral mixing model, synthetic aperture radar (SAR) images, and the digital elevation model (DEM) acquired during the Shuttle Radar Topography Mission (SRTM). The DEM, vegetation fraction, and ASTER band 3 were the most useful parameters for defining the forest classes. The normalized difference vegetation index (NDVI), ASTER band 1, vegetation fraction, and the Advanced Land Observing Satellite (ALOS)/Phased Array type L-band Synthetic Aperture Radar (PALSAR) transmitting and receiving horizontal polarization (HH) and transmitting horizontal and receiving vertical polarization (HV) were all effective in distinguishing the wetland classes campinarana and Mauritia. Tests of statistical significance indicated the overall accuracies and kappa coefficients (κ) of 88% and 0.86 for the final map, respectively. The allocation disagreement coefficient of 5% and a quantity disagreement value of 7% further attested the statistical significance of the classification results. Hence, in addition to water, exposed soil, and deforestation areas, OBIA and DM were successful for differentiating a large number of open (forest, wood, shrub, and grass campinaranas), forest (terra firme, várzea, igapó, and alluvial), as well as Mauritia wetland classes in the inner and outer areas of the studied megafan.     

Perception of a geological body using multiple source remotely-sensed data-relative influence of the spectral content and the spatial resolution

Fracturing maps over a granitic dome (Scaër granite, Brittany, France) have been extracted from the most widely available remotely-sensed data and from aerial photographs. Comparison of the different maps obtained allowed the classification of the mapping potential of the different raw and merged images as well as ranking their ability to point out geological features at different scales. Three different types of geological features were pinpointed: a coarse regional fracturing, kilometric plutonic domes and finer geological structures such as circular features within the granitic dome. The best means of revealing each of these three types of geological features, proved to be radar images, multi-spectral data and aerial photographs, respectively. The data providing the largest range of observation and the greatest amount of information on geological structures and soil types were the merged Landsat-TM and SPOT panchromatic images.     

Aerial image dehazing using a deep convolutional autoencoder

Aerial images and videos are extensively used for object detection and target tracking. However, due to the presence of thin clouds, haze or smoke from buildings, the processing of aerial data can be challenging. Existing single-image dehazing methods that work on ground-to-ground images, do not perform well on aerial images. Moreover, current dehazing methods are not capable for real-time processing. In this paper, a new end-to-end aerial image dehazing method using a deep convolutional autoencoder is proposed. Using the convolutional autoencoder, the dehazing problem is divided into two parts, namely, encoder, which aims extract important features to dehaze hazy regions and decoder, which aims to reconstruct the dehazed image using the down-sampled image received from the encoder. In this proposed method, we also exploit the superpixels in two different scales to generate synthetic thin cloud data to train our network. Since this network is trained in an end-to-end manner, in the test phase, for each input hazy aerial image, the proposed algorithm outputs a dehazed version without requiring any other information such as transmission map or atmospheric light value. With the proposed method, hazy regions are dehazed and objects within hazy regions become more visible while the contrast of non-hazy regions is increased. Experimental results on synthetic and real hazy aerial images demonstrate the superiority of the proposed method compared to existing dehazing methods in terms of quality and speed.

     

数字表面模型(DSM)参数指示林窗特征的有效性研究

林窗的空间格局不仅对林下物种多样性有重要作用,而且也是量化不同林型结构特征的重要指标.近年来,灵活便捷的小型无人机航拍技术的快速发展为获取高分辨率的森林冠层三维结构信息提供了可低成本获取的途径.利用航拍影像提取林窗斑块来计算景观指数是描述林窗格局有效的传统方法,但提取细小林窗往往难度大,尤其是在需要处理大量航拍数据时会...     

Contour analysis-based matching of ground objects in aerial images

An original approach to matching images is proposed based on the technique of comparison of the structure of contours. The technique for detecting contours is reinforced by combining the use of data from independent contours and contours that are the boundaries of regions. For this purpose, several assumptions and simple rules of general nature are formulated. Contours are identified based on local salient features of their one-dimensional representation. To avoid ambiguity in identifying contours, the technique of dynamic programming is harnessed. All computational procedures are invariant to image rotation. In conclusion the developed method is applied to the problem of matching image fragments in aerial images.     

Use of shadows for extracting buildings in aerial images

Two new methods integrating region growing and edge detection are presented to extract buildings in aerial images. In the domain we are considering, building shadows play an important role in enhancing the detection accuracy and reliability. In both methods, interpretations of buildings are only based on the reliable features (shadows and shapes of buildings in our case) obtained by segmentation and are not affected by artifacts or sementation errors. We also propose algorithms to correct the segmentation errors or to enhance resultant contours based on local edge information (contrast and smoothness along contours) and global shape information. In the first method we start with an edge detection to loacte shadowed boundaries, then combine region growing to find boundaries without shadows. In the second method, we use region growing to obtain a primary segmentation. Edge information is then used to eliminate false boundaries or to modify the contours. Further modification of contour artifacts can be achieved by using domain knowledge (most boundaries are smooth enough to be approximated by straight segments.)     

Location of kimberlites using Landsat Thematic Mapper images and aerial photographs: the Redondao diatreme,Brazil

Abstract

Landsat Thematic Mapper images and aerial photographs were used in the detection of kimberlile-derived materials in the Redondao test site. In this area kimberlite-derived soils show a flora constituted mainly by grasses and shrubs, which differ from the surrounding savanna-park (cerrado) vegetation cover. Band-ratio images were able to distinguish kimberlite-derived materials by enhancing areas with different vegetation covers. However, the coarse spatial resolution of Landsat-TM images compared with the spatial variability of the study area, and the removal of topographic shadowing effects on ratio images blurred several landscape features. To increase discrimination, Landsat Thematic Mapper ratio images were merged with digitized aerial photographs through intensity, hue and saturation (IHS) colour transforms. The resulting merged colour composite highlighted the spatial and spectral features of the study area permitting an accurate definition of the kimberlite-derived materials within the Redondao diatreme.     

Specific object retrieval based on salient regions

In this paper, we present an image retrieval technique for specific objects based on salient regions. The salient regions we select are invariant to geometric and photometric variations. Those salient regions are detected based on low level features, and need to be classified into different types before they can be applied on further vision tasks. We first classify the selected regions into four types including blobs, edges and lines, textures, and texture boundaries, by using the correlations with the neigbouring regions. Then, some specific region types are chosen for further object retrieval applications. We observe that regions selected from images of the same object are more similar to each other than regions selected from images of different objects. Correlation is used as the similarity measure between regions selected from different images. Two images are considered to contain the same object, if some regions selected from the first image are highly correlated to some regions selected from the second image. Two data sets are employed for experiment: the first data set contains human face images of a number of different people and is used for testing the retrieval algorithm on distinguishing specific objects of the same category; and the second data set contains images of different objects and is used for testing the retrieval algorithm on distinguishing objects of different categories. The results show that our method is very effective on specific object retrieval.     

基于图分割与Hausdorff距离的多分辨率影像匹配

针对具有复杂场景的航拍图像提出了一种基于图分割理论与Hausdorff距离的多分辨率影像匹配方法。在高斯金字塔图像模型中,低分辨率的图像通过图分割方法,充分考虑图像中的局部和全局的信息,提取到稳定和完整的图像区域边界,并以区域边界作为待匹配的曲线。再通过计算曲线的统计特性作为图像间待匹配特征,并由信号相关的度量方法粗估计出图像间全局仿射变换参数。利用粗估计的参数在高分辨率层次上进一步通过基于Hausdorff距离的匹配方法搜索到精确的变换参数。实验结果表明,该方法在较大变形和强噪音干扰的情况下对复杂场景的图像也能有效地完成匹配。     

Forest fuel treatment detection using multi-temporal airborne lidar data and high-resolution aerial imagery: a case study in the Sierra Nevada Mountains,California

Treatments to reduce forest fuels are often performed in forests to enhance forest health, regulate stand density, and reduce the risk of wildfires. Although commonly employed, there are concerns that these forest fuel treatments (FTs) may have negative impacts on certain wildlife species. Often FTs are planned across large landscapes, but the actual treatment extents can differ from the planned extents due to operational constraints and protection of resources (e.g. perennial streams, cultural resources, wildlife habitats). Identifying the actual extent of the treated areas is of primary importance to understand the environmental influence of FTs. Light detection and ranging (lidar) is a powerful remote-sensing tool that can provide accurate measurements of forest structures and has great potential for monitoring forest changes. This study used the canopy height model (CHM) and canopy cover (CC) products derived from multi-temporal airborne laser scanning (ALS) data to monitor forest changes following the implementation of landscape-scale FT projects. Our approach involved the combination of a pixel-wise thresholding method and an object-of-interest (OBI) segmentation method. We also investigated forest change using normalized difference vegetation index (NDVI) and standardized principal component analysis from multi-temporal high-resolution aerial imagery. The same FT detection routine was then applied to compare the capability of ALS data and aerial imagery for FT detection. Our results demonstrate that the FT detection using ALS-derived CC products produced both the highest total accuracy (93.5%) and kappa coefficient (κ) (0.70), and was more robust in identifying areas with light FTs. The accuracy using ALS-derived CHM products (the total accuracy was 91.6%, and the κ was 0.59) was significantly lower than that using ALS-derived CC, but was still higher than using aerial imagery. Moreover, we also developed and tested a method to recognize the intensity of FTs directly from pre- and post-treatment ALS point clouds.     

无人机可见光遥感影像地物目标提取技术研究

无人机可见光遥感影像中地物目标边界清晰度较低,容易导致地物目标与背景之间的区分度降低,进而难以提取地物目标。为此,提出无人机可见光遥感影像地物目标提取方法。从光谱特征、纹理特征和边缘特征三个方面分析无人机可见光遥感影像特征。结合三种影像特征对无人机可见光遥感影像数据集实行增广处理。对完成增广后的数据集定义影像编码标签,以此确定地物目标增强权重,通过参量化处理地物目标光谱特征,计算光谱吸收指数,获取地物目标提取表达式,从而实现无人机可见光遥感影像地物目标提取。实验结果表明,所提方法能够保证地物目标边界的清晰度,具有较强的地物目标提取能力。     

Generation of a DTM and building detection based on an MPF through integrating airborne lidar data and aerial images

This study presents an approach that uses airborne light detection and ranging (lidar) data and aerial imagery for creating a digital terrain model (DTM) and for extracting building objects. The process of creating the DTM from lidar data requires four steps in this study: pre-processing, segmentation, extraction of ground points, and refinement. In the pre-processing step, raw data are transformed to raster data. For segmentation, we propose a new mean planar filter (MPF) that uses a 3 × 3 kernel to divide lidar data into planar and nonplanar surfaces. For extraction of ground points, a new method to extract additional ground points in forest areas is used, thus improving the accuracy of the DTM. The refinement process further increases the accuracy of the DTM by repeated comparison of a temporary DTM and the digital surface model. After the DTM is generated, building objects are extracted via a proposed three-step process: detection of high objects, removal of forest areas, and removal of small areas. High objects are extracted using the height threshold from the normalized digital surface model. To remove forest areas from among the high objects, an aerial image and normalized digital surface model from the lidar data are used in a supervised classification. Finally, an area-based filter eliminates small areas, such as noise, thus extracting building objects. To evaluate the proposed method, we applied this and three other methods to five sites in different environments. The experiment showed that the proposed method leads to a notable increase in accuracy over three other methods when compared with the in situ reference data.     

Noisy fingerprints classification with directional FFT based features using MLP

A methodology is described for classifying noisy fingerprints directly from raw unprocessed images. The directional properties of fingerprints are exploited as input features by computing one-dimensional fast Fourier transform (FFT) of the images over some selected bands in four and eight directions. The ability of the multilayer perceptron (MLP) for generating complex boundaries is utilised for the purpose of classification. The superiority of the method over some existing ones is established for fingerprints corrupted with various types of distortions, especially random noise.