Patent Application
20110150290
In forest management, it is important to know information about the trees in a forest area. Such information can include the species of trees in the forest, their spacing, age, diameter, health, etc. This information is useful for revenue prediction, active management planning (such as selective thinning, fertilizing etc.), determining where to transport logs or how to equip a sawmill to process the logs and for other uses. While it is possible to inventory a forest area using statistical surveying techniques, it is becoming increasingly cost prohibitive to send survey crews into remote forest areas to obtain the survey data. As a result, remote sensing is becoming increasingly used as a substitute for physically surveying a forest area. Remote sensing typically involves the use of aerial photography or satellite imagery to produce images of the forest. The images are then analyzed by hand or with a computer to obtain information about the trees in the forest.
The most common way of analyzing an image of the forest in order to identify a particular species of tree is to analyze the brightness of the leaves or needles of the trees in one or more ranges of wavelengths or spectral bands. Certain species of trees have a characteristic spectral reflectivity that can be used to differentiate one species from another. While this method can work to distinguish between broad classes of trees such as between hardwoods and conifers, the technique often cannot make finer distinctions. For example, spectral reflectance alone is not very accurate in distinguishing between different types of conifers such as Western Hemlock and Douglas Fir. Given these limitations, there is a need for an improved technique of analyzing images of forest lands to predict information about the trees in the images.
The technology disclosed herein relates to a method of predicting information about trees based on a spatial variation of pixel intensities within an image of the forest where the area imaged by each pixel is less than the expected crown size of the trees in the forest. In one embodiment, a number of training images of forest areas are obtained for which ground truth data for one or more measurement metrics of the trees in the forest are known. The training images of the forest area are analyzed to determine a measure of the spatial variation in the intensity of the pixel data in one or more spectral bands for the images. The determined spatial variations are correlated with the verified metrics for the trees in the training images to determine a relationship between the spatial variations and the particular metric. Once a relationship has been determined, the relationship is used to predict values of the metric for trees in other forest areas.
In one embodiment, the spatial variation of the pixel intensities is determined by analyzing pixel intensity data in a frequency domain. In one embodiment, a two-dimensional fast Fourier transform (FFT) is computed on the pixel intensity data for an area of an image. Parameters from an FFT output matrix are used to quantify the spatial variation of the pixel intensities and to predict a value for the correlated metric for the trees in the image using a relationship determined from the ground truth data.
In one embodiment, the average power of the frequency components and the standard deviation of the powers of the frequency components in rings of cells surrounding an average pixel intensity value in the FFT output matrix are used to quantify the spatial variation in pixel intensities.
This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This summary is not intended to identify key features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.
The foregoing aspects and many of the attendant advantages of this invention will become more readily appreciated as the same become better understood by reference to the following detailed description, when taken in conjunction with the accompanying drawings, wherein:
As indicate above, the technology disclosed herein relates to a method of operating a computer system to predict a metric for trees in a forest area from a corresponding image of the trees. In one disclosed embodiment, the metric to be determined is the percentage of a particular species of tree in a forest area. However, the metric may be other information such the number of trees of a particular species in the forest area, the average age of the trees, the average diameter of the trees or other information that is capable of being verified with ground truth data.
As indicated above, the disclosed technology analyzes a spatial variation in pixel intensities within an image of a forest to predict a metric for the trees in the image. The spatial variation captures the higher intensity pixels caused by brighter reflections from the leaves or needles in the tree canopy as well as the darker spots where there are no leaves or needles or where the leaves and needles are in shadow. The spatial pattern of lighter and darker areas in the canopy provide information that is related to the metric being predicted.
In one embodiment of the disclosed technology, the spatial variations in pixel intensities within an image are measured by converting the pixel intensities of the image into a corresponding frequency domain. In one particular embodiment, the pixels are converted into the frequency domain using a two-dimensional FFT or wavelet analysis. To convert the pixel intensities into the frequency domain, a pixel block from the image is selected. Preferably the pixel block is square with a number of pixels that is evenly divisible by 2 e.g. 16×16, 32×32, 64×64 etc. The area imaged by each pixel and the number of pixels in the pixel block is a selected to be able to detect small variations within the canopy while not requiring too long to analyze all the pixels within the images of the forest. In one embodiment, each pixel images an area of approximately 1 meter square and the pixel block has 32 by 32 pixels.
In the example shown, the FFT output matrix 200 is calculated from a 16×16 pixel block and has 8 rings surrounding the center cell 250. The average power of the frequency components in the cells of each ring are calculated as P1-P8. That is, P1 is the average power of the frequency components in the ring 252. P2 is the average power of the frequency components in the cells of the ring 254. P3 is the average power of the frequency components in the cells of the ring 256 etc. The standard deviations for the powers of the frequency components in the cells of each ring are calculated as SD1-SD8 in a similar manner i.e. SD1 is the standard deviation of the powers in the cells of ring 252, SD2 is the standard deviation of the powers in the cells of ring 254 etc. In this embodiment, each FFT output matrix is used to calculate 16 variables that vary with the spatial variation of the pixel intensities of the corresponding pixel block.
At 308, the computer system performs a statistical correlation between the measure of the spatial variation in pixel intensity values as determined by the quantities P1-P8 and SD1-SD8 and measurements taken from the trees that are imaged by each pixel block. For example, a correlation can be made between the values P1-P8 and SD1-SD8 computed from the FFT output matrix for each pixel block and the measured percentage of a particular species of tree in the areas corresponding to each pixel block.
In one embodiment, the correlation is made by computing a least squares linear regression of the measured ground truth metrics from the areas corresponding to the pixel blocks in each of the training images and the 16 variables determined from the FFT output matrices that quantify the spatial variations in pixel intensities from the pixel blocks. As will be understood by those of skill in the art, the result of the linear regression is a set of 16 coefficients, each of which corresponds to one of the 16 variables that quantify the spatial variation in pixel intensity values. The sum of the 16 variables and their corresponding coefficients determined from the regression predict a value for a metric for the trees in the image.
In one embodiment, each training image has pixel data for a number of spectral bands e.g. green, red, infrared etc. The spatial variation in pixel intensities for each spectral band is analyzed and used to compute a set of corresponding coefficients using a regression analysis. At 310, an error, such as a least squares error, can be computed for the coefficients determined for each spectral band in order to select which spectral band correlates best with the particular metric in question. As will be appreciated, some metrics (e.g. tree species) may be better predicted using pixel intensities in one spectral band while other metrics (e.g. tree age) may be better predicted using pixel intensities in another spectral band. In another embodiment, the variables from two or more spectral bands may be used in determining the relationship between the measurement metric and the variation in pixel intensities from the images. For example, if two more spectral bands are used, then the linear regression analysis can be performed with the variables determined from the FFT's computed from the images in each spectral band.
As shown in
To predict a metric for trees in an area of a forest, an image of the forest area is obtained at 402. The image is divided into one or more pixel blocks at 404 and the spatial variation of the pixel intensities using the spectral band or bands that best correlated with the metric to be predicted is determined at 406. At 408, a predicted value for a metric (species, age, diameter etc.) for the trees imaged by the pixel block is predicted using the relationship previously determined from the training images.
While illustrative embodiments have been illustrated and described, it will be appreciated that various changes can be made therein without departing from the scope of the invention. For example, other techniques besides a two-dimensional Fourier transform could be used to quantify the spatial variation in pixel intensities. Furthermore, pattern analyses such as cluster analyses or other two-dimensional image processing techniques could be used to quantify the spatial variation in the pixel intensities in an image. Similarly, other measurements from the FFT output matrix such as the standard deviation alone or the average power alone could be used in the correlation. Therefore, the scope of the invention is to be determined from the following claims and equivalents thereof.
