Nesrine Chehata

Terrain Modeling From Lidar Range Data in Natural Landscapes: A Predictive and Bayesian Framework

The Earths topography, including vegetation and human-made features, reduced to a virtual 3-D representation is a key geographic layer for any extended development or risk management project. Processed from multiple aerial images or from airborne lidar systems, the 3-D topography is first represented as a point cloud. This paper deals with the generation of digital terrain models (DTMs) in natural landscapes. We present a global methodology for estimating the terrain height by deriving a predictive filter paradigm. Under the assumption that the terrain topography (elevation and slope) is regular in a neighboring system, a predictive filter combines linearly the predicted topographic values and the effective measured values. In this paper, such a filter is applied to 3-D lidar data which are known to be of high elevation accuracy. The algorithm generates an adaptive local geometry wherein the elevation distribution of the point cloud is analyzed. Since local terrain elevations depend on the local slope, a predictive filter is first applied on the slopes and then on the terrain elevations. The algorithm propagates through the point cloud following specific rules in order to optimize the probability of computing areas containing terrain points. Considered as an initial surface, the previous DTM is finally regularized in a Bayesian framework. Our approach is based on the definition of an energy function that manages the evolution of a terrain surface. The energy is designed as a compromise between a data attraction term and a regularization term. The minimum of this energy corresponds to the final terrain surface. The methodology is discussed, and some conclusive results are presented on vegetated mountainous areas.

Contribution of airborne full-waveform lidar and image data for urban scene classification

Airborne lidar systems have become an alternative source for the acquisition of altimeter data. In addition to multi-echo laser scanner systems, full-waveform systems are able to record the whole backscattered signal for each emitted laser pulse. These data provide more information about the structure and the physical properties of the surface. This paper is focused on the classification of full-waveform lidar and airborne image data on urban scenes. Random forests are used since they provide an accurate classification and run efficiently on large datasets. Moreover, they provide measures of variable importance for each class. This is crucial to analyze the relevance of each feature for the classification of urban scenes. Random Forests provide more accurate results than Support Vector Machines with an overall accuracy of 95.75%. The most relevant features show the contribution of lidar waveforms for classifying dense urban scenes and improve the classification accuracy for all classes.

Comparison of Pleiades and LiDAR Digital Elevation Models for Terraces Detection in Farmlands

Among the most evident anthropogenic modifications of the landscape, terraces related to agricultural activities are ubiquitous structures that constitute important investments worldwide, and they recently acquired a new relevance to modern concerns about land-use management and erosion control. Conservation agriculture and terraces management are an application with great potentialities for Satellite Earth observation and the derived high-resolution topography. Due to its high agility, the Pleiades satellite constellation provides new, high-resolution digital elevation models (DEMs) with a submetric resolution that could be potentially useful for this task, and their application in a farmland context is nowadays an open research line. This work provides a first analysis, performing an automatic terrace mapping from DEMs obtained from Pleiades images, as compared to LiDAR DEMs. Two existing methods are considered: 1) the fast line segment detector (LSD) algorithm and 2) a geomorphometric method based on surface curvature. Despite the lower performances of Pleiades DEMs with respect to that of the LiDAR models, the results indicate that the Pleiades models can be used to automatically detect terrace slopes greater than 2 m with a detection rate of more than 80% of the total length of the terraces. In addition, the results showed that when using noisy DEMs, the geomorphometric method is more robust, and it slightly outperforms the LSD algorithm. These results provide a first analysis on how effective Pleiades DEMs can be as an alternative to LiDAR DEMs, also highlighting the future challenges for monitoring large extents in a farmland context.

Support Vectors Selection for Supervised Learning Using an Ensemble Approach

Support Vector Machines (SVMs) are popular for pattern classification. However, training a SVM requires large memory and high processing time, especially for large datasets, which limits their applications. To speed up their training, we present a new efficient support vector selection method based on ensemble margin, a key concept in ensemble classifiers. This algorithm exploits a new version of the margin of an ensemble-based classification and selects the smallest margin instances as support vectors. Our experimental results show that our method reduces training set size significantly without degrading the performance of the resulting SVMs classifiers.

Terrain modeling from lidar data: Hierarchical K-means filtering and Markovian regularization

Lidar 3D point cloud corresponds to the terrestrial topography, including true ground and objects belonging either to vegetated areas or to human made features. This paper deals with DTM (digital terrain model) production. First step filtering data into ground and off-ground points is based on a multi-resolution coarse-to-fine approach. The K-means algorithm is used in a hierarchical way that provides robust data filtering. The number of cluster splits is used to automatically qualify the filtering reliability. This point is rarely treated in previous works. Secondly, a regularization process over ground points generates an accurate DTM on a regular grid. The fine DTM is processed with ground points without using classical interpolation algorithms. In fact, a Markovian regularization minimizes a global energy that confronts the terrain regularity and the goodness of fit to the data. It also depends on the filtering reliability. Conclusive results are presented on vegetated and mountainous areas and provide realistic terrain models.

A two-pass random forests classification of airborne lidar and image data on urban scenes

Random forests ensemble classifier showed to be suitable for classifying mutlisource data such as lidar and RGB image for urban scene mapping. However, two major problems remain: (1) the class boundaries are not well classified, a common issue in classification (2) the data are highly imbalanced raising another issue more specific to urban scenes. In this paper, we propose a new ensemble method based on the margin paradigm to improve the classification accuracy of minor classes. Random forests classifier is used in a two-pass methodology with an improved capability for classifying imbalanced data.

Use intermediate results of wrapper band selection methods: A first step toward the optimization of spectral configuration for land cover classifications

Intermediate results of two state-of-the-art wrapper feature selection approaches (GA and SFFS) associated to a classifier (linear SVM) applied to hyperspectral data sets were used to derive information about band importance for specific land cover classification problems. The impact of the number of selected bands on classification accuracy was obtained thanks to SFFS, while a band importance measure was derived from intermediate sets of bands tested by GA. Such results are a first step toward the identification of the most suitable spectral bands to design superspectral camera systems dedicated to specific applications (e.g. classification of urban land cover and material maps).

Comparison of VHR panchromatic texture features for tillage mapping

Agricultural practices are major drivers of water flows in cultivated landscapes. Especially, the spatial arrangements and connectivities of tilled/untilled fields have a strong impact onto run off and soil erosion at the landscape and watershed scales. Very high spatial resolution satellite images offer the possibility to classify tilled vs. untilled fields at a large scale. This paper compares the importance of various VHR texture features for tillage mapping. Classical texture features such as coocurrence Haralick descriptors, Gabor and SIFT-based descriptors are studied. The random forest classifier is used to assess feature importance. A 50 cm panchromatic WorldView-I image is used for experiments. Very good classification accuracies of 83.4 % and 94.5 % are reached.

Classification of remote sensing data using margin-based ensemble methods

This work exploits the margin theory to design better ensemble classifiers for remote sensing data. The margin paradigm is at the core of a new bagging algorithm. This method increases the classification accuracy, particularly in case of difficult classes, and significantly reduces the training set size. The same margin framework is used to derive a novel ensemble pruning algorithm. This method not only highly reduces the complexity of ensemble methods but also performs better than complete bagging in handling minority classes. Our techniques have been successfully used for the classification of remote sensing data.

Contribution of band selection and fusion for hyperspectral classification

For some specific land cover classification problems, it may be interesting to design superspectral camera systems with reduced numbers of bands (∼ 20) and optimized band widths. This paper assesses the contribution of band selection and band fusion processes separately and jointly for dimensionality reduction. The proposed approach is fully automatic and based on a wrapper feature selection using Random forest classifier and a similarity-based fusion process. While combining both processes, selection before fusion gave the best results, reducing by almost 91% the number of bands while keeping satisfying accuracies. Results are presented on Indian Pines, Salinas and Pavia Centre hyperspectral datasets.