No abstract

Deep convolutional neural networks (DCNNs) have been used to achieve state-of-the-art performance on many computer vision tasks (e.g., object recognition, object detection, semantic segmentation) thanks to a large repository of annotated image data. Large labeled datasets for other sensor modalities, e.g., multispectral imagery (MSI), are not available due to the large cost and manpower required. In this paper, we adapt state-of-the-art DCNN frameworks in computer vision for semantic segmentation for MSI imagery. To overcome label scarcity for MSI data, we substitute real MSI for generated synthetic MSI in order to initialize a DCNN framework. We evaluate our network initialization scheme on the new RIT-18 dataset that we present in this paper. This dataset contains very-high resolution MSI collected by an unmanned aircraft system. The models initialized with synthetic imagery were less prone to over-fitting and provide a state-of-the-art baseline for future work.

When classifying point clouds, a large amount of time is devoted to the process of engineering a reliable set of features which are then passed to a classifier of choice. Generally, such features -usually derived from the 3Dcovariance matrix -are computed using the surrounding neighborhood of points. While these features capture local information, the process is usually time-consuming, and requires the application at multiple scales combined with contextual methods in order to adequately describe the diversity of objects within a scene. In this paper we present a 1D-fully convolutional network that consumes terrain-normalized points directly with the corresponding spectral data, if available, to generate point-wise labeling while implicitly learning contextual features in an end-to-end fashion. Our method uses only the 3D-coordinates and three corresponding spectral features for each point. Spectral features may either be extracted from 2D-georeferenced images, as shown here for Light Detection and Ranging (LiDAR) point clouds, or extracted directly for passive-derived point clouds, i.e. from muliple-view imagery. We train our network by splitting the data into square regions, and use a pooling layer that respects the permutation-invariance of the input points. Evaluated using the ISPRS 3D Semantic Labeling Contest, our method scored second place with an overall accuracy of 81.6%. We ranked third place with a mean F1-score of 63.32%, surpassing the F1-score of the method with highest accuracy by 1.69%. In addition to labeling 3D-point clouds, we also show that our method can be easily extended to 2D-semantic segmentation tasks, with promising initial results.

Abstract-A fast, completely automated method to create 3D watertight building models from airborne LiDAR point clouds is presented. The proposed method analyzes the scene content and produces multi-layer rooftops with complex boundaries and vertical walls that connect rooftops to the ground. A graph cuts based method is used to segment vegetative areas from the rest of scene content. The ground terrain and building rooftop patches are then extracted utilizing our technique, the hierarchical Euclidean clustering. Our method adopts a "divide-and-conquer" strategy. Once potential points on rooftops are segmented from terrain and vegetative areas, the whole scene is divided into individual pendent processing units which represent potential building footprints. For each individual building region, significant features on the rooftop are further detected using a specifically designed region growing algorithm with smoothness constraint. Boundaries for all of these features are refined in order to produce strict description. After this refinement, mesh models could be generated using an existing robust dual contouring method.

No abstract

This paper describes a proof-of-concept implementation that uses a high dynamic range CMOS video camera to integrate daylight harvesting and occupancy sensing functionalities. It has been demonstrated that the proposed concept not only circumvents several drawbacks of conventional lighting control sensors, but also offers functionalities that are not currently achievable by these sensors. The prototype involves three algorithms, daylight estimation, occupancy detection and lighting control. The calibrated system directly estimates luminance from digital images of the occupied room for use in the daylight estimation algorithm. A novel occupancy detection algorithm involving color processing in YCC space has been developed. Our lighting control algorithm is based on the least squares technique. Results of a daylong pilot test show that the system i) can meet different target light-level requirements for different task areas within the field-of-view of the sensor, ii) is unaffected by direct sunlight or a direct view of a light source, iii) detects very small movements within the room, and iv) allows real-time energy monitoring and performance analysis. A discussion of the drawbacks of the current prototype is included along with the technological challenges that will be addressed in the next phase of our research.

This paper focuses on the calibration of multispectral sensors typically used for remote sensing. These systems are often provided with “factory” radiometric calibration and vignette correction parameters. These parameters, which are assumed to be accurate when the sensor is new, may change as the camera is utilized in real-world conditions. As a result, regular calibration and characterization of any sensor should be conducted. An end-user laboratory method for computing both the vignette correction and radiometric calibration function is discussed in this paper. As an exemplar, this method for radiance computation is compared to the method provided by MicaSense for their RedEdge series of sensors. The proposed method and the method provided by MicaSense for radiance computation are applied to a variety of images captured in the laboratory using a traceable source. In addition, a complete error propagation is conducted to quantify the error produced when images are converted from digital counts to radiance. The proposed methodology was shown to produce lower errors in radiance imagery. The average percent error in radiance was −10.98%, −0.43%, 3.59%, 32.81% and −17.08% using the MicaSense provided method and their “factory” parameters, while the proposed method produced errors of 3.44%, 2.93%, 2.93%, 3.70% and 0.72% for the blue, green, red, near infrared and red edge bands, respectively. To further quantify the error in terms commonly used in remote sensing applications, the error in radiance was propagated to a reflectance error and additionally used to compute errors in two widely used parameters for assessing vegetation health, NDVI and NDRE. For the NDVI example, the ground reference was computed to be 0.899 ± 0.006, while the provided MicaSense method produced a value of 0.876 ± 0.005 and the proposed method produced a value of 0.897 ± 0.007. For NDRE, the ground reference was 0.455 ± 0.028, MicaSense method produced 0.239 ± 0.026 and the proposed method produced 0.435 ± 0.038.

Many targets that remote sensing scientists encounter when conducting their research experiments do not lend themselves to laboratory measurement of their surface optical properties. Removal of these targets from the field can change their biotic condition, disturb the surface composition, and change the moisture content of the sample. These parameters, as well as numerous others, have a marked influence on surface optical properties such as spectral and bi-directional emissivity. This necessitates the collection of emissivity spectra in the field. The propagation of numerous devices for the measurement of midwave and longwave emissivity in the field has occurred in recent years. How good a re these devices and how does the accuracy of the spectra they produce compare to the "tried and true" laboratory devices that have been around for decades? A number of temperature/emissivity separation algorithms will be demonstrated on data collected with a field portable Fourier transform infrared (FTIR) spectrometer and the merits and resulting accuracy compared to laboratory spectra made of these identical samples. A brief look at off-nadir view geometries will also be presented to alert scientists to the possible sources of error in these spectra that may result when using sensing systems that do not look straight down on targets or when their nadir looking sensor is looking at a tilted target.