One of the lesser known sources of multispectal data is the National Agricultural Imagery Program, administered by the US Department of Agriculture. The intended purpose of the data is to aid public officials in conducting crop assessments. To facilitate this, the imagery is collected during the "leaf on" season with virtually no cloud cover. The resolution is high: 1m per pixel. The imagery is orthorectified and is accurate to within 6m of horizontal displacement from known control points.
The data is collected over the continental United States only. Before 2007, only three bands corresponding to visible blue, green and red wavelengths were collected using analog cameras with filters. After 2007 the imagery has been acquired digitally and four bands are available with a near infrared (NIR) band added to the visible bands. The most common use of the additional band has been in color infrared (CIR) images produced by inputting the green, red and NIR bands into the three channels of an image color model. CIR images can highlight and discriminate different types of vegetation better than standard RGB images in many cases. Although limited to four spectral bands, the high resolution of this data compared to other imagery makes this an interesting alternative.
Another and possibly more effective use of the data is the identification of ground features using spectral signature matching using PANCROMA™ Point Spectrum Generator and Spectral Analysis utilities. In order to do this, start by downloading NAIP data. Four band NAIP is available from the USGS Data Gateway website. Follow the instructions on the site by first selecting orthoimagery for your area of interest, then selecting one of the NAIP data sets. Since NAIP data is acquired cyclically every three years, there are usually a few to choose from. Make sure that you select GeoTiff format as the data is also offered in JPEG, JPEG 2000 and IMG format.
The four bands are bundled into a single Planar Configuration=2, 8-bit GeoTiff file. After downloading the single file, the first step for using them in PANCROMA™ is as usual to unbundle them four individual 8-bit data files. To do this select 'Decompose Bundles' | 'Four File NAIP Data' from the Main Menu. When the file opens, you will be prompted for a base file name. Input and save as GeoTiff. The band number designations will be atomically appended to your selected base file name. The image files are typically about 30MB per band. Row and column configurations vary, with some images being rather narrow and long. A subset from a typical image covering an area in northern Vermont USA where I live is shown below. It was clearly acquired during the summer months and the highly reflective gravel roads characteristic of the rural areas of the state or strikingly visible.
Multispectral analysis using NAIP data is a bit different from that using lower resolution data like Landsat. As a result of the 30m Landsat resolution, many subtle features are lost. This can actually make multispectral analysis easier. High resolution images like NAIP introduce several challenges. The tremendous detail that they capture can make features that look uniform from a distance appear distressingly variable. The mown alfalfa fields in the example image. We generally get a lot of rain in Vermont, and the fields can stay quite wet will into the summer. Variations in ground moisture as a result of surface gradient can make obvious differences in reflectivity for fields that are all planted with the same crop. The imagery is so detailed that features left by hay gathering. When the mower and hay rake pass over the field, they lay down the mown grass in opposite directions as they pass over the field. This causes significant differences in reflectivity. The furrows in the plowed fields are another example of variability in an essentially homogeneous target.
Another important consideration is elevation effects. In low resolution imagery, the heights of buildings, trees, etc. are foreshortened and rendered unimportant with respect to surface reflectivity. In high resolution images, these objects can cast long shadows that significantly affect apparent reflectivity. Shadows cast by tree lines over planted fields will alter the reflectivity of the crops near the edges. This effect is so pronounced that the identification of tree species by signature matching can be quite problematic as each individual tree shows areas of high reflectivity and high shadow. It is very important to use images that were acquired around mid day when the sun is at its zenith as opposed to later in the day when long shadows are cast.
Lets now make an attempt at the identification of targets using NAIP imagery. We will try a fairly easy target first: farm ponds. There are a lot of these on Vermont dairy ponds. Open the four band files in the order blue, green, red and NIR in that order. The select 'Spectral Analysis' | 'NAIP Point Spectrum Generator' | 'Four 8-Bit DN Files'. Using your mouse, click on some points on one of the ponds that you can identify visually. Select 'OK' on the Point Spectrum Form and the spectrum for the target will be computed. The one for my target is shown below.
Next select 'Close Graphic Image and Reset'. Then re open the four band images again by selecting 'File' | 'Open' as usual. Now use the NAIP Spectral Analyzer tool to identify all ground features matching your target spectrum. To do so select 'Spectral Analysis' | 'SPOT Spectral Analyzer' | 'Four DN Bands' | 'Euclidean Distance'. When the Spectral Criteria Form becomes visible, it is generally a good idea to check the 'Distance Fraction' check box and the '11%' radio button. This will only highlight the areas of closest match and is usually necessary for these highly detailed NAIP images. Select 'OK' and the match will be computed. The images below (depicted at half scale) show the result.
This run was pretty successful. Many ponds and wetland areas were highlighted successfully and there were relatively few false positives. Note that you can switch layers by selecting the RGB and band 1 images using the menu at the top of the image display window. A similar attempt to discriminate between gravel roads and macadam roads was similarly successful, as the image below demonstrates.
Now we will attempt a more challenging target, a single tree species. Vermont lowland forests are characterized by a mix of oak, beech and of course maple among other species. We will start by selecting a single tree. The image below shows one standing relatively alone and easily identifiable.
The species is unknown but it could be easily verified from the ground, as it is right next to a road. Using the Point Spectrum Generator, I selected several pixels on that specific tree only. The resulting spectrum is shown in the image below.
Using the Spectral Analyzer with the Euclidean Distance Fraction set as described above, I generated the plot shown below.
Inspection of the plot indicates that it identified other trees in the forested areas proximal to the target. However there were considerable false positives as some areas of the hayfields showed similar reflectivities to the target. There are not enough bands. However the analysis is arguably superior to visual inspection of the CIR images, and yields a quantitative rather than qualitative result.
This example shows that NAIP data can provide very interesting high resolution data for multispectral analysis. It provides a method for target identification for applications where high ground resolution is essential.