The large volumes of spatial data and their products made it necessary to research new data compression techniques. Much research has been focused on developing compression methods for aerial and satellite imagery [1,2]. As a result, a number of image compression techniques have been developed for these types of imagery. These compression techniques can be generally categorized in two classes: 1) one that reduces the number of bits and creates a numerically-identical replica of the original image; and 2) one that creates a much compressed replica of the image, but with much degraded quality [3].

LiDAR sensors generate huge datasets of unstructured point clouds of multiple returns, which may be false signals or correspond to natural or manmade features. Because of its large size, LiDAR data has costly storage and computational requirements [4]. To reduce the size of LiDAR data for effective storage and processing, robust compression methods are being researched. Wu and Amaratunga [5] presented wavelet transform-based triangulated networks (WTIN) method for representing large GIS datasets. The WTIN method, which is based on the second generation wavelet theory, can be used to produce multi-resolution representations of the data. Although this method has produced compact multi-resolution representations of LiDAR data with acceptable quality, more efficient quantization schemes are much needed than the simple threshold operation adopted in the method to compress the data [6,7]. Ali and Mehrabian [8] developed a novel computational paradigm for picking sample points to create a triangulated irregular network (TIN) model from LiDAR data of a flat terrain. This method can be used to remove unwanted and redundant points in over-sampled smooth terrain surfaces and small high resolution objects resulting in a compact TIN model.

The method uses the Voronoi diagram to evaluate the local density of the LiDAR points and identify clusters within the data. Then, points in the same proximity with elevations within a threshold are selected. The Voronoi tree concept is then used to delete the selected points and update the Voronoi diagram. The final TIN is then built using a randomized incremental algorithm. The methods described above can help produce a compressed LiDAR dataset. However, none of these methods can be used to remove unwanted and redundant points in the LiDAR data set. The method presented herein can be used to remove unwanted and redundant LiDAR point; producing a much compressed LiDAR dataset through spatial clustering and optimal plane fitting.

The method adopted in this research consists of (1) classification and spatial clustering and (2) optimal firstorder plane fitting of the LiDAR dataset. The schematic diagram of Figure 1 summarizes the major steps of our compression method.

First-order plane-fitting helps in identifying the type of surface defined by every LiDAR point and its neighbors, and therefore the removal of unwanted, redundant LiDAR points, and fake LiDAR returns becomes possible. The technical approach adopted in this study is explained below:

The LiDAR data used in the work was acquired for a study area in the north east region of the City of Venice, which is located in Sarasota County, Florida, United

States (Figure 3). The LiDAR dataset for the study area was produced in 2005 for the Southwest Florida Water Management District (SWFWMD) as part of the Manatee/Little Manatee LiDAR Survey project. The data set comprised of 3-D bare earth mass points delivered in the LAS file format based upon the District’s 5000’ × 5000’ grid structure. The LiDAR data (Figure 4) was collected using a Leica LS-50 LiDAR system integrated with an inertial measuring unit (IMU) and a dual frequency GPS receiver. Positional accuracy is 0.75-ft root mean square (RMSE), which satisfies the National Standard for Spatial Data Accuracy (NSSDA) standard for 2-foot contours (scale of 1:12,000). Bare earth LiDAR masspoint data has a vertical accuracy of 0.3-feet root mean square (RMSE). Projected Coordinate System is North American Datum (NAD) of 1983; State Plane System of Florida West and the Vertical datum is the North American Vertical Datum (NAVD) of 1988.

Classification of the study area ortho-imagery was performed using the Bayesian maximum likelihood classification (BMLC) method. The process started by computing statistics for selected training sites of land cover classes and used the results of the statistical summary to

classify the image. Histogram analysis was performed to locate ortho-image clusters using intensity and distance metrics (Figure 5). The classified ortho-image was sharpened and then a vectorized layer was created using the run graph method resulting in a polygon layer (Figure 6).

The resulted vector layer was then used to initiate a sweeping spatial clustering algorithm in order to identify clusters in the LiDAR dataset following [11]. The algorithm is based on the sweep-line concept which is widely known in computational geometry and computer graphics. The identified clusters in the LiDAR dataset as processed with this algorithm are shown in Figure 7. Optimal planes fitted to the LiDAR dataset (Figure 8) were then generated using the principal component analysis method performed on the clustered LiDAR dataset of Figure 6.

As it can be seen in Figure 7, the sweeping spatial clustering algorithm managed to identify clusters in the LiDAR dataset successfully by separating smooth objects such as buildings tops, pavement, etc. from vegetation and water bodies, even in where these classes are in the same local neighborhood. The spatial clustering of the

LiDAR data in this way helped facilitate the execution of the optimal plane fitting.

Although the optimal planes shown in Figure 8 have been created from the clustered LiDAR dataset shown in Figure 7; we shouldn’t expect the two outcomes to be similar because the optimal plane fitting process by design doesn’t necessary consider the classes in the study area. Optimal plane fitting considers the local neighbored of every LiDAR point regardless of the class to which that point belongs. A keep-or-remove criterion has then been established to remove every LiDAR point that doesn’t fall on one of those optimal planes. All LiDAR points in the study area, which are not located on these optimal planes, have been removed; resulting in a muchcompressed LiDAR dataset that has a smaller size on disk. Specifically, the size of the original LiDAR dataset was 703.32MB and that of the compressed dataset was 578.13MB; resulting in a compression ratio of 17.8%.

A LiDAR data compression method was presented in this ongoing research work based on spatial clustering and optimal plane fitting. The method has produced a compression ratio of 17.8% for the LiDAR dataset of the study area, which is promising. The issue this ongoing study is trying to address however is not only the development of a LiDAR compression method with low computational demands. The objective is to develop a compression method that can be applied at the LiDAR acquisition stage that only records the LiDAR points that are on these optimal planes. If this goal is achieved, it will help to design a LiDAR sensor in the future that will only record points that are located on these planes.