Jan Stuckens

Integrating Contextual Information with per-Pixel Classification for Improved Land Cover Classification

Abstract A hybrid segmentation procedure to integrate contextual information with per-pixel classification in a metropolitan area land cover classification project is described and evaluated. It is presented as a flexible tool within a commercially available image processing environment, allowing components to be adapted or replaced according to the users needs, the image type, and the availability of state-of-the-art algorithms. In the case of the Twin Cities metropolitan area of Minnesota, the combination of the Shen and Castan edge detection operator with iterative centroid linkage region growing/merging based on Students t-tests proved optimal when compared to other more common contextual approaches, such as majority filtering and the Extraction and Classification of Homogeneous Objects classifier. Postclassification sorting further improved the results by reducing residual confusion between urban and bare soil categories. Overall accuracy of the optimal classification technique was 91.4% for a level II classification (10 classes) with a Ke of 90.5%. The incorporation of contextual information in the classification process improved accuracy by 5.8% and Ke by 6.5%. As expected, classification accuracy for a simplified level I classification (five classes) was higher with 95.4% and 94.3% for Ke. A second important advantage of the technique is the reduced occurrence of smaller mapping units, resulting in a more attractive classification map compared to traditional per-pixel maximum likelihood classification results.

A weighted linear spectral mixture analysis approach to address endmember variability in agricultural production systems

The least squares error (LSE) technique is frequently used to estimate abundance fractions in linear spectral mixture analysis (LSMA). The LSE is typically equally weighted for all wavebands, assuming equally important effects. This is, however, not always the case and therefore traditional LSMA often results in suboptimal fraction estimates. This study presents a weighted LSMA approach that prioritises wavebands with minor or no negative effects on fraction estimates. Synthetic mixed pixel spectra compiled from in situ measured spectra of bare soil, citrus tree and weed canopies were used for validation. The results show markedly improved fraction estimates obtained for the weighted approach, with a mean absolute gain of 0.24 in R 2 and a mean absolute reduction in fraction abundance error of 0.06.

Sampling design of ground-based lidar measurements of forest canopy structure and its effect on shadowing

This research was undertaken to study the influence of the sampling design and laser beam density of ground-based light detection and ranging (lidar) measurements of forests on the quality of the collected laser datasets in terms of shadowing. Virtual forest stands generated by stochastic L-systems as tree descriptors are used as a basis depending on the study frame and requirements. The dynamic plant modeler and plant nursery natFX (Bionatics, CIRAD, Montpellier, France) was used to simulate deciduous forest stands of three tree species (Fagus sylvatica L., Platanus acerifolia (Ait.) Willd., and Populus nigra L.) with varying structural characteristics. Hemispherical laser measurements with different laser beam densities were simulated according to three different sampling patterns (single, diamond, corners) inside these virtual forest stands using ray-tracing technology. An adjusted sampling design has proven its effectiveness, since an average shadowing decrease of 29.10% was obtained in comparison with that for a single measurement. This finding contrasts with an average decrease of 13.27% by increasing laser beam density by a factor of 25. In the next step, contact frequency values were calculated from the virtual laser datasets. These values were used to model the shadowed parts of the canopy, demonstrating the potential of ground-based laser scans to capture the three-dimensional leaf distribution inside a forest stand in terms of leaf area density (LAD). On average, the LAD estimates underestimated the true LAD by 19.55%, 12.67%, and 10.54% for the single, diamond, and corners setups, respectively. In each of the cases, the LAD values from the single design resulted in a lower accuracy compared with those for the diamond and corners setups.

3D modeling of light interception in heterogeneous forest canopies using ground-based LiDAR data

A methodology is presented that describes the direct interaction of a forest canopy with incoming radiation using terrestrial LiDAR based vegetation structure in a radiative transfer model. The proposed ‘Voxel-based Light Interception Model’ (VLIM) is designed to estimate the Percentage of Above Canopy Light (PACL) at any given point of the forest scene. First a voxel-based representation of trees is derived from terrestrial LiDAR data as structural input to model and analyze the light interception of canopies at near leaf level scale. Nine virtual forest stands of three species (beech, poplar, plantain) were generated by means of stochastic L-systems as tree descriptors. Using ray tracer technology hemispherical LiDAR measurements were simulated inside these virtual forests. The leaf area density (LAD) estimates derived from the LiDAR datasets resulted in a mean absolute error of 32.57% without correction and 16.31% when leaf/beam interactions were taken into account. Next, comparison of PACL estimates, computed with VLIM with fully rendered light distributions throughout the canopy based on the L-systems, yielded a mean absolute error of 5.78%. This work shows the potential of the VLIM to model both instantaneous light interception by a canopy as well as average light distributions for entire seasons.

The Properties of Terrestrial Laser System Intensity for Measuring Leaf Geometries: A Case Study with Conference Pear Trees (Pyrus Communis)

Light Detection and Ranging (LiDAR) technology can be a valuable tool for describing and quantifying vegetation structure. However, because of their size, extraction of leaf geometries remains complicated. In this study, the intensity data produced by the Terrestrial Laser System (TLS) FARO LS880 is corrected for the distance effect and its relationship with the angle of incidence between the laser beam and the surface of the leaf of a Conference Pear tree (Pyrus Commmunis) is established. The results demonstrate that with only intensity, this relationship has a potential for determining the angle of incidence with the leaves surface with a precision of ±5° for an angle of incidence smaller than 60°, whereas it is more variable for an angle of incidence larger than 60°. It appears that TLS beam footprint, leaf curvatures and leaf wrinkles have an impact on the relationship between intensity and angle of incidence, though, this analysis shows that the intensity of scanned leaves has a potential to eliminate ghost points and to improve their meshing.

Assessment of Light Environment Variability in Broadleaved Forest Canopies Using Terrestrial Laser Scanning

Light availability inside a forest canopy is of key importance to many ecosystem processes, such as photosynthesis and transpiration. Assessment of light availability and within-canopy light variability enables a more detailed understanding of these biophysical processes. The changing light-vegetation interaction in a homogeneous oak (Quercus robur L.) stand was studied at different moments during the growth season using terrestrial laser scanning datasets and ray tracing technology. Three field campaigns were organized at regular time intervals (24 April 2008; 07 May 2008; 23 May 2008) to monitor the increase of foliage material. The laser scanning data was used to generate 3D representations of the forest stands, enabling structure feature extraction and light interception modeling, using the Voxel-Based Light Interception Model (VLIM). The VLIM is capable of estimating the relative light intensity or Percentage of Above Canopy Light (PACL) at any arbitrary point in the modeled crown space. This resulted in a detailed description of the dynamic light environments inside the canopy. Mean vertical light extinction profiles were calculated for the three time frames, showing significant differences in light attenuation by the canopy between April 24 on the one hand, and May 7 and May 23 on the other hand. The proposed methodology created the opportunity to link these within-canopy light distributions to the increasing amount of photosynthetically active leaf material and its distribution in the considered 3D space.

Webcams for Bird Detection and Monitoring: A Demonstration Study

Better insights into bird migration can be a tool for assessing the spread of avian borne infections or ecological/climatologic issues reflected in deviating migration patterns. This paper evaluates whether low budget permanent cameras such as webcams can offer a valuable contribution to the reporting of migratory birds. An experimental design was set up to study the detection capability using objects of different size, color and velocity. The results of the experiment revealed the minimum size, maximum velocity and contrast of the objects required for detection by a standard webcam. Furthermore, a modular processing scheme was proposed to track and follow migratory birds in webcam recordings. Techniques such as motion detection by background subtraction, stereo vision and lens distortion were combined to form the foundation of the bird tracking algorithm. Additional research to integrate webcam networks, however, is needed and future research should enforce the potential of the processing scheme by exploring and testing alternatives of each individual module or processing step.

RPV model parameters based on hyperspectral bidirectional reflectance measurements of Fagus sylvatica L. leaves

The bidirectional reflectance parametric and semi-empirical Rahman-Pinty-Verstraete (RPV) model was inverted based on Bidirectional Reflectance Factor (BRF) measurements of 60 Fagus sylvatica L. leaves in the optical domain between 400 nm and 2,500 nm. This was accomplished using data retrieved from the Compact Laboratory Spectro-Goniometer (CLabSpeG) with an azimuth and zenith angular step of 30 and 15 degrees, respectively. Wavelength depended RPV parameters describing the leaf reflectance shape (rho0), the curve convexity (k) and the dominant forward scattering (Θ) were derived using the RPVinversion-2 software (Joint Research Centre) package with Correlation Coefficient values between modelled and measured data varying between 0.71 and 0.99 for all wavelengths, azimuth and zenith positions. The RPV model parameters were compared with a set of leaves not participating in the inversion procedure and presented Correlation Coefficient values ranging between 0.64 and 0.94 suggesting that RPV could be also used for simulating single canopy elements such as leaves.

Normalization of illumination conditions for ground based hyperspectral measurements using dual field of view spectroradiometers and BRDF corrections

BRDF effects present in dual field-of-view spectroscopy datasets were investigated. A data-driven normalization procedure was developed by decomposing the target BRDF into a target specific Lambertian component and a bi-directional component characterizing a group of similar targets,. The normalization method was used to convert reflectance factors obtained under cloud obscured conditions into clear sky conditions. An evaluation on four targets measured under different illumination conditions suggests that the normalization can reduce relative reflectance errors between 400 and 1800 nm from 15% to less than 5% even under full cloud obscuration. At higher wavelengths a decreased signal-to-noise ratio increases the error level.

Detecting citrus tree water status by integrating hyperspectral remote sensing and physiological data in a water flow-storage model

This study presents a method for determining the stem/xylem water potential of non — stressed citrus trees from leaf spectra, taking into account the hydraulic properties of the trees. Existing spectral indexes can only accurately predict the water content of plants but not the water potential, thus limiting their usefulness for irrigation management. In the proposed method, leaf reflectance data was used to derive the water content of non-stressed Satsuma mandarin leaves using an inversion of the PROSPECT model. Crown water content, obtained from up-scaled leaf water content, was then used as an input to a dynamic water flow-storage model. Model validation with the non-stressed dataset gave an R2 between the measured and modeled stem water potential of 0.85, with a slope and intercept of 1.03 and 0.31, respectively. These results suggest that integrating hyperspectral and in situ data potentially yields accurate estimates of the plant water potential.