Wenkai Li

A New Method for Segmenting Individual Trees from the Lidar Point Cloud

Light Detection and Ranging (lidar) has been widely applied to characterize the 3-dimensional (3D) structure of forests as it can generate 3Dpoint data with high spatial resolution and accuracy. Individual tree segmentations, usually derived from the canopy height model, are used to derive individual tree structural attributes such as tree height, crown diameter, canopy-based height, and others. In this study, we develop a new algorithm to segment individual trees from the small footprint discrete return airborne lidar point cloud. We experimentally applied the new algorithm to segment trees in a mixed conifer forest in the Sierra Nevada Mountains in California. The results were evaluated in terms of recall, precision, and F-score, and show that the algorithm detected 86 percent of the trees (“recall”), 94 percent of the segmented trees were correct (“precision”), and the overall F-score is 0.9. Our results indicate that the proposed algorithm has good potential in segmenting individual trees in mixed conifer stands of similar structure using small footprint, discrete return lidar data.

Effects of Topographic Variability and Lidar Sampling Density on Several DEM Interpolation Methods

This study aims to quantify the effects of topographic variability (measured by coefficient variation of elevation, CV) and lidar (Light Detection and Ranging) sampling density on the DEM (Digital Elevation Model) accuracy derived from several interpolation methods at different spatial resolutions. Interpolation methods include natural neighbor (NN), inverse distance weighted (IDW), triangulated irregular network (TIN), spline, ordinary kriging (OK), and universal kriging (UK). This study is unique in that a comprehensive evaluation of the combined effects of three influencing factors (CV, sampling density, and spatial resolution) on lidar-derived DEM accuracy is carried out using different interpolation methods. Results indicate that simple interpolation methods, such as IDW, NN, and TIN, are more efficient at generating DEMs from lidar data, but kriging-based methods, such as OK and UK, are more reliable if accuracy is the most important consideration. Moreover, spatial resolution also plays an important role when generating DEMs from lidar data. Our results could be used to guide the choice of appropriate lidar interpolation methods for DEM generation given the resolution, sampling density, and topographic variability.

Delineating Individual Trees from Lidar Data: A Comparison of Vector- and Raster-based Segmentation Approaches

Light detection and ranging (lidar) data is increasingly being used for ecosystem monitoring across geographic scales. This work concentrates on delineating individual trees in topographically-complex, mixed conifer forest across the California’s Sierra Nevada. We delineated individual trees using vector data and a 3D lidar point cloud segmentation algorithm, and using raster data with an object-based image analysis (OBIA) of a canopy height model (CHM). The two approaches are compared to each other and to ground reference data. We used high density (9 pulses/m2), discreet lidar data and WorldView-2 imagery to delineate individual trees, and to classify them by species or species types. We also identified a new method to correct artifacts in a high-resolution CHM. Our main focus was to determine the difference between the two types of approaches and to identify the one that produces more realistic results. We compared the delineations via tree detection, tree heights, and the shape of the generated polygons. The tree height agreement was high between the two approaches and the ground data (r2: 0.93–0.96). Tree detection rates increased for more dominant trees (8–100 percent). The two approaches delineated tree boundaries that differed in shape: the lidar-approach produced fewer, more complex, and larger polygons that more closely resembled real forest structure.

A Positive and Unlabeled Learning Algorithm for One-Class Classification of Remote-Sensing Data

In remote-sensing classification, there are situations when users are only interested in classifying one specific land-cover type, without considering other classes. These situations are referred to as one-class classification. Traditional supervised learning is inefficient for one-class classification because it requires all classes that occur in the image to be exhaustively assigned labels. In this paper, we investigate a new positive and unlabeled learning (PUL) algorithm, applying it to one-class classifications of two scenes of a high-spatial-resolution aerial photograph. The PUL algorithm trains a classifier on positive and unlabeled data, estimates the probability that a positive training sample has been labeled, and generates binary predictions for test samples using an adjusted threshold. Experimental results indicate that the new algorithm provides high classification accuracy, outperforming the biased support-vector machine (SVM), one-class SVM, and Gaussian domain descriptor methods. The advantages of the new algorithm are that it can use unlabeled data to help build classifiers, and it requires only a small set of positive data to be labeled by hand. Therefore, it can significantly reduce the effort of assigning labels to training data without losing predictive accuracy.

A maximum entropy approach to one-class classification of remote sensing imagery

In remote sensing classification there are situations when users are only interested in classifying one specific land type without considering other classes, which is referred to as one-class classification. Traditional supervised learning requires all classes that occur in the image to be exhaustively labelled and hence is inefficient for one-class classification. In this study we investigate a maximum entropy approach (MAXENT) to one-class classification of remote sensing imagery, i.e. classifying a single land class (e.g. urban areas, trees, grasses and soils) from an aerial photograph with 0.3 m spatial resolution. MAXENT estimates the Gibbs probability distribution that is proportional to the conditional probability of being positive. A threshold for generating binary predictions can be determined based on the omission rate of a validation set. The results indicate that MAXENT provides higher classification accuracy than the one-class support vector machine (OCSVM). MAXENT does not require other land classes for training. Its input is only a set of training samples of the specific land class of interest, as well as a set of known constraints on the distribution. Therefore, the effort of manually collecting training data for classification can be significantly reduced.

A New Accuracy Assessment Method for One-Class Remote Sensing Classification

In one-class remote sensing classification, users are only interested in classifying one specific land type (positive class), without considering other classes (negative class). Previous researchers have proposed different one-class classification methods without requiring negative data. An appropriate accuracy measure is usually needed to tune free parameters/threshold and to evaluate the classification result. However, traditional accuracy measures, such as the kappa coefficient and F-measure (F), require both positive and negative data, and hence, they are not applicable for positive-only data. In this paper, we investigate a new accuracy assessment method that does not require negative data. Two new statistics Fpb (proxy of F-measure based on positive-background data) and Fcpb (prevalence-calibrated proxy of F-measure based on positive-background data) can be calculated from a modified confusion matrix, where the observed negative data are replaced by background data. To investigate the effectiveness of the new method, we produced different one-class classification results using two scenes of aerial photograph, and the accuracy values were evaluated by Fpb, Fcpb, kappa coefficient, and F. The effectiveness of F pb in model and threshold selection was investigated as well. Experimental results show that the behaviors of Fpb, Fcpb, F, and kappa coefficient are similar, and they all rank the models by accuracy similarly. In model and threshold selection, the classification accuracy values produced by maximizing Fpb and F are similar, and they are higher than those produced by setting an arbitrary rejection fraction. Therefore, we conclude that the new method is effective in model selection, threshold selection, and accuracy assessment, and it will have important applications in one-class remote sensing classification since negative data are not needed.

SRTM DEM Correction in Vegetated Mountain Areas through the Integration of Spaceborne LiDAR, Airborne LiDAR, and Optical Imagery

The Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) is one of the most complete and frequently used global-scale DEM products in various applications. However, previous studies have shown that the SRTM DEM is systematically higher than the actual land surface in vegetated mountain areas. The objective of this study is to propose a procedure to calibrate the SRTM DEM over large vegetated mountain areas. Firstly, we developed methods to estimate canopy cover from aerial imagery and tree height from multi-source datasets (i.e., field observations, airborne light detection and ranging (LiDAR) data, Geoscience Laser Altimeter System (GLAS) data, Landsat TM imagery, climate surfaces, and topographic data). Then, the airborne LiDAR derived DEM, covering ~5% of the study area, was used to evaluate the accuracy of the SRTM DEM. Finally, a regression model of the SRTM DEM error depending on tree height, canopy cover, and terrain slope was developed to calibrate the SRTM DEM. Our results show that the proposed procedure can significantly improve the accuracy of the SRTM DEM over vegetated mountain areas. The mean difference between the SRTM DEM and the LiDAR DEM decreased from 12.15 m to −0.82 m, and the standard deviation dropped by 2 m.

Predicting potential distributions of geographic events using one-class data: concepts and methods

One common problem with geographic data is that, for a specific geographic event, only occurrence information is available; information about the absence of the event is not available. We refer to these specific types of geospatial data as geographic one-class data (GOCD). Predicting the potential spatial distributions that a particular geographic event may occur from GOCD is difficult because traditional binary classification methods that require availability of both positive and negative training samples cannot be used. The objective of this research is to define GOCD and propose novel approaches for modelling potential spatial distributions of geographic events using GOCD. We investigate the effectiveness of one-class support vector machine (OCSVM), maximum entropy (MAXENT) and the newly proposed positive and unlabelled learning (PUL) algorithm for solving GOCD problems using a case study: species distribution modelling from synthetic data. Our experimental results indicate that generally OCSVM, MAXENT and PUL are effective in modelling the GOCD. Each method has advantages and disadvantages, but PUL seems to be the most promising method.

Space-time analyses for forecasting future incident occurrence: a case study from Yosemite National Park using the presence and background learning algorithm

To address a spatiotemporal challenge such as incident prevention, we need information about the time and place where incidents have occurred in the past. Using geographic coordinates of previous incidents in coincidence with spatial layers corresponding to environmental variables, we can produce probability maps in geographic and temporal space. Here, we evaluate spatial statistic and machine learning approaches to answer an important space-time question: where and when are wildland search and rescue (WiSAR) incidents most likely to occur within Yosemite National Park (YNP)? We produced a monthly probability map for the year 2011 based on the presence and background learning (PBL) algorithm that successfully forecasts the most likely areas of WiSAR incident occurrence based on environmental variables (distance to anthropogenic and natural features, vegetation, elevation, and slope) and the overlap with historic incidents from 2001 to 2010. This will allow decision-makers to spatially allocate resources where and when incidents are most likely to occur. In the process, we not only answered questions related to a real-world problem but also used novel space-time analyses that give us insight into machine learning principles. The GIScience findings from this applied research have major implications for best practices in future space-time research in the fields of epidemiology and ecological niche modeling.

One-Class Classification of Airborne LiDAR Data in Urban Areas Using a Presence and Background Learning Algorithm

Automatic classification of light detection and ranging (LiDAR) data in urban areas is of great importance for many applications such as generating three-dimensional (3D) building models and monitoring power lines. Traditional supervised classification methods require training samples of all classes to construct a reliable classifier. However, complete training samples are normally hard and costly to collect, and a common circumstance is that only training samples for a class of interest are available, in which traditional supervised classification methods may be inappropriate. In this study, we investigated the possibility of using a novel one-class classification algorithm, i.e., the presence and background learning (PBL) algorithm, to classify LiDAR data in an urban scenario. The results demonstrated that the PBL algorithm implemented by back propagation (BP) neural network (PBL-BP) could effectively classify a single class (e.g., building, tree, terrain, power line, and others) from airborne LiDAR point cloud with very high accuracy. The mean F-score for all of the classes from the PBL-BP classification results was 0.94, which was higher than those from one-class support vector machine (SVM), biased SVM, and maximum entropy methods (0.68, 0.82 and 0.93, respectively). Moreover, the PBL-BP algorithm yielded a comparable overall accuracy to the multi-class SVM method. Therefore, this method is very promising in the classification of the LiDAR point cloud.