Ville Heikkinen

Simulated Multispectral Imagery for Tree Species Classification Using Support Vector Machines

The information content of remotely sensed data depends primarily on the spatial and spectral properties of the imaging device. This paper focuses on the classification performance of the different spectral features (hyper- and multispectral measurements) with respect to three tree species. The Support Vector Machine was chosen as the classification algorithm for these features. A simulated optical radiation model was constructed to evaluate the identification performance of the given multispectral system for the tree species, and the effects of spectral-band selection and data preprocessing were studied in this setting. Simulations were based on the reflectance measurements of the pine (Pinus sylvestris L.), spruce [ Picea abies (L.) H. Karst.], and birch trees (Betula pubescens Ehrh. and Betula pendula Roth). Leica ADS80 airborne sensor with four spectral bands (channels) was used as a fixed multispectral sensor system that leads to response values for the at-sensor radiance signal. Results suggest that this four-band system has inadequate classification performance for the three tree species. The simulations demonstrate on average a 5-15 percentage points improvement in classification performance when the Leica system is combined with one additional spectral band. It is also demonstrated for the Leica data that feature mapping through a Mahalanobis kernel leads to a 5-10 percentage points improvement in classification performance when compared with other kernels.

Regularized learning framework in the estimation of reflectance spectra from camera responses

For digital cameras, device-dependent pixel values describe the cameras response to the incoming spectrum of light. We convert device-dependent RGB values to device- and illuminant-independent reflectance spectra. Simple regularization methods with widely used polynomial modeling provide an efficient approach for this conversion. We also introduce a more general framework for spectral estimation: regularized least-squares regression in reproducing kernel Hilbert spaces (RKHS). Obtained results show that the regularization framework provides an efficient approach for enhancing the generalization properties of the models.

Logistic Regression-Based Spectral Band Selection for Tree Species Classification: Effects of Spatial Scale and Balance in Training Samples

In this letter, we evaluated the pixel-level and plot-level tree species classification of Scots Pine, Norway Spruce, and deciduous birch in a boreal forest using 64-band AisaEAGLE II hyperspectral data in a wavelength range of 400-1000 nm. First, band selection was performed using a sparse logistic regression-based feature selection algorithm with pixel-level and plot-level data in case of balanced and imbalanced training data. This resulted in 8-11 selected hyperspectral bands, depending on the properties of the data used. We evaluated a tree species classification with 8-11 selected hyperspectral bands directly for a least squares support vector machine (LS-SVM)-based pixel-level classification with a relatively small training set size (0.5%-1.5% of the total data) and obtained an accuracy and kappa of around 93.50% and 0.90, respectively. These results are around 0.53%-0.94% points lower than those obtained using all of the hyperspectral bands. Second, one important wavelength region highlight by the selected bands was used to modify the sensor sensitivity configuration in the Leica Airborne Digital Sensor 40 (ADS40) multispectral sensor. Using a simulation model and the hyperspectral data, the modified and standard Leica ADS40 sensor responses were simulated and compared, and the modified system simulated response indicates a 3%-5% point improvement in the pixel-level and plot-level LS-SVM classification accuracy compared with the simulated responses of the standard Leica ADS40 band configuration.

Thermal and hyperspectral imaging for Norway spruce (Picea abies) seeds screening

Hyperspectral and thermal lifetime imaging were used to assess spruce seed quality.Viable, empty and infested seeds were resolved with high accuracy with both methods.400-1000nm data was not as informative as 1000-2500nm and thermal decay data.Classification of 93% accuracy was obtained using three wavelengths in SWIR range.The results suggest that high-throughput spruce seed quality testing is possible. The quality of seeds used in agriculture and forestry is tightly linked to the plant productivity. Thus, the development of high-throughput nondestructive methods to classify the seeds is of prime interest. Visible and near infrared (VNIR, 400-1000nm range) and short-wave infrared (SWIR, 1000-2500nm range) hyperspectral imaging techniques were compared to an infrared lifetime imaging technique to evaluate Norway spruce (Picea abies (L.) Karst.) seed quality. Hyperspectral image and thermal data from 1606 seeds were used to identify viable seeds, empty seeds and seeds infested by Megastigmus sp. larvae. The spectra of seeds obtained from hyperspectral imaging, especially in SWIR range and the thermal signal decay of seeds following an exposure to a short light pulse were characteristic of the seed status. Classification of the seeds to three classes was performed with a Support Vector Machine (nu-SVM) and sparse logistic regression based feature selection. Leave-One-Out classification resulted to 99% accuracy using either thermal or spectral measurements compared to radiography classification. In spectral imaging case, all important features were located in the SWIR range. Furthermore, the classification results showed that accurate (93.8%) seed sorting can be achieved with a simpler method based on information from only three hyperspectral bands at 1310nm, 1710nm and 1985nm locations, suggesting a possibility to build an inexpensive screening device. The results indicate that combined classification methods with hyperspectral imaging technique and infrared lifetime imaging technique constitute practically high performance fast and non-destructive techniques for high-throughput seed screening.

Evaluating logarithmic kernel for spectral reflectance estimation—effects on model parametrization, training set size, and number of sensor spectral channels

In this work, we evaluate the conditionally positive definite logarithmic kernel in kernel-based estimation of reflectance spectra. Reflectance spectra are estimated from responses of a 12-channel multispectral imaging system. We demonstrate the performance of the logarithmic kernel in comparison with the linear and Gaussian kernel using simulated and measured camera responses for the Pantone and HKS color charts. Especially, we focus on the estimation model evaluations in case the selection of model parameters is optimized using a cross-validation technique. In experiments, it was found that the Gaussian and logarithmic kernel outperformed the linear kernel in almost all evaluation cases (training set size, response channel number) for both sets. Furthermore, the spectral and color estimation accuracies of the Gaussian and logarithmic kernel were found to be similar in several evaluation cases for real and simulated responses. However, results suggest that for a relatively small training set size, the accuracy of the logarithmic kernel can be markedly lower when compared to the Gaussian kernel. Further it was found from our data that the parameter of the logarithmic kernel could be fixed, which simplified the use of this kernel when compared with the Gaussian kernel.

Forest Stand Delineation Using a Hybrid Segmentation Approach Based on Airborne Laser Scanning Data

Forest stand delineation is an important task for forest management. Traditional manual stand delineation based on aerial color-infrared images is a labor intensive process and its results are partially subjective. These images are also highly affected by weather conditions and imaging parameters. In this work, we applied a hybrid segmentation approach on Airborne Laser Scanning (ALS) data to delineate forest stands. The ALS data was firstly pre-processed to extract a three band feature image, containing tree height, density, and species information, respectively. Then the image was segmented by the mean shift algorithm to generate raw stands, which were refined by the Spectral Clustering (SC) algorithm in the following stage. In the SC algorithm, we also estimated the number of stands based on eigengap heuristics. We tested our method on real ALS data acquired at Juuka in Finland, and compared the results with the manually delineated result visually and numerically, as well as results based on previous methods. The experimental results showed that our method worked well for the forest stand delineation based on ALS data, and return better results in most cases when compared to previous methods.

Spectral imaging using consumer-level devices and kernel-based regression

Hyperspectral reflectance factor image estimations were performed in the 400-700 nm wavelength range using a portable consumer-level laptop display as an adjustable light source for a trichromatic camera. Targets of interest were ColorChecker Classic samples, Munsell Matte samples, geometrically challenging tempera icon paintings from the turn of the 20th century, and human hands. Measurements and simulations were performed using Nikon D80 RGB camera and Dell Vostro 2520 laptop screen as a light source. Estimations were performed without spectral characteristics of the devices and by emphasizing simplicity for training sets and estimation model optimization. Spectral and color error images are shown for the estimations using line-scanned hyperspectral images as the ground truth. Estimations were performed using kernel-based regression models via a first-degree inhomogeneous polynomial kernel and a Matérn kernel, where in the latter case the median heuristic approach for model optimization and link function for bounded estimation were evaluated. Results suggest modest requirements for a training set and show that all estimation models have markedly improved accuracy with respect to the DE00 color distance (up to 99% for paintings and hands) and the Pearson distance (up to 98% for paintings and 99% for hands) from a weak training set (Digital ColorChecker SG) case when small representative training data were used in the estimation.

ALS data based forest stand delineation with a coarse-to-fine segmentation approach

Forest stands are important forest management units for many forestry applications. Airborne Laser Scanning (ALS), also known as Light Detection and Ranging (LiDAR), provides abundant information on vertical structures of forests, and it is more stable than the conventional aerial or satellite spectral imaging in various conditions. In order to obtain accurate Forest Stand Delineations (FSD), a method using a coarse-to-fine segmentation approach solely based on ALS data is proposed in this work. Firstly, a three-band feature image is extracted from ALS point clouds. Secondly, from the feature image, coarse forest stands are generated with the mean-shift algorithm. Then each coarse stand is refined by Simple Linear Iterative Clustering superpixels based Seeded Region Growing (SLIC-SRG) to generate final forest stands. Also we process separately the forest and non-forest areas to increase the delineation accuracy. Moreover, some parameters of our method have physical meanings in FSD, easing the parameter setting process. Our method was tested on real ALS data and compared with other methods. Experimental results show that our method outperforms previous methods in terms of accuracies for ALS data based FSD.

Color and Image Characterization of a Three CCD Seven Band Spectral Camera

In this study spectral and spatial characterization of a seven channel FluxData 1665 MS7 three-CCD spectral video camera was performed in terms of the sensor spectral sensitivity, linearity, spatial uniformity, noise and spatial alignment. The results indicate small deviation from ideal linear sensor response. Also, a small spatial non-uniformity and systematic shift exists between the channel images. However, images were observed to have high quality in term of noise. Spectral characterization showed that the sensor has good response in the 380-910 nm region with only some sensitivity limitations in the 715-740 nm range. We also evaluated the spectral reflectance estimation in 400-700 nm range using empirical regression methods and the Digital ColorChecker SG and ColorChecker charts. These experiments resulted in average ΔE00 color accuracy of 1.6 – 2.4 units, depending on the illuminant and estimation method.

Segmentation of Skin Spectral Images Using Simulated Illuminations

Spectral imaging has been drawing attentions in skin segmentation related medical and computer vision applications. Usually spectral images contain radiance spectra or approximated reflectance spectra of objects. The reflectance spectrum is independent of illuminations, whereas the radiance spectrum is dependent on the illumination that was used during the measurement. Recently, the general illumination is using increasingly LED lights that have different spectral characteristics when compared to incandescent or fluorescent lights. In this paper, we studied effects of illuminations in unsupervised pixel-based skin segmentation with the Spectral Clustering (SC) algorithm based on two sets of skin spectral images, human hand and face. We adopted the eigengap heuristic to select the illumination in skin segmentation with the SC algorithm. Spectral radiance images under three CIE standard illuminants (A, D65, and F11), and two white LED illuminations, were respectively simulated from corresponding spectral reflectance images. The experimental results showed that CIE D65 illuminant return the highest average segmentation accuracies with 2.1% point and 3.4% point higher than the lowest results of illuminant CIE A in hand images and illuminant CIE F11 in face images, respectively. The accuracies of two LED illuminations are also high with around 1.5% point and 1.0% point decreases in average from the accuracy of illuminant CIE D65.