ITC analysis of aerial images
- Introduction
- Aerial Sensor Technology
- Image needs for computer image analysis
- Synergy with LiDAR
- BRDF Correction and Normalization of Aerial Data Area
- Tests of feature-based BRDF correction curves on two flight lines
- Conclusion
Tests of feature-based BRDF correction curves on two flight lines
Two neighbouring flight lines from a Leica ADS40-v2 sensor mission were selected to test a variety of BRDF corrections (Figure 2). Two test areas, one of softwoods and one of hardwoods, were created in the overlapping region of two flight lines (Figure 3). The differences in radiances of the ITCs within these two test areas on each image are going to be monitored for BRDF correction curves generated from a variety of increasingly precise features.
Figure 2 - Overlapping sections of adjascent ADS40v2 flight lines were selected to test feature-based BRDF correction and normalization effects. Photo: Bowater and R&B Cormier.
Figure 3 - Two test areas, one of softwoods (in white) and one of hardwoods (in green), were created in the overlapping region of two flight lines from a Leica ADS-40v2 sensor to test the correction capabilities of various BRDF curves.
Figure 4 - From the two flight lines, the two views of the softwood test area. One can see how the trees are seen leaning towards the right in the left image (from the left flight line), and vice-versa in the right side image, due to their off-nadir position in the original images. Notice also how the trees are seen slightly more front-lit in the right image than the left one.
Traditionally, BRDF correction curves are generated by accumulating the radiance values within each image column (i.e., at each pixel position), for each spectral band. The effects are generally more pronounced with the near infra-red channel. For a given spectral band, such histogram displays the illumination trends as one gets away from nadir. A curve is fitted to the histogram and the radiance at nadir is subtracted from the curve to make it relative to nadir. The inverse of the curve is applied to the image data, on a line-by-line basis, to correct its radiances for the BRDF effects.
Unfortunately, this classic approach is strongly influenced by the image content. For example, the presence of more lakes or more cut over areas on one side of the image will strongly (and adversely) influence the correction curve. A first order improvement to this approach, consist in first creating a mask of vegetative areas and then, only collecting data underneath that mask.
Another potential improvement, one easily achieved in our context, is to use the ITC mask rather than a vegetation mask. This allows the correction curve to concentrate on the effect of BRDF on trees, which may be different than that on other low lying vegetation. As trees react differently to the combination of illumination and view angles due to their different shapes and branching patterns, a further improvement could potentially be achieved by using different masks (thus, different correction curves) for hardwood and softwood trees. In addition, since the individual tree crowns generated by our crown delineation process contain the shade side of each tree, another improvement could potentially be achieved by only picking-up the lit-side of each crown. After all, the lit-side of each tree crown is what will generally be used in the ITC classification process, thus, better to optimize the correction of radiances within that context. Such feature-specific BRDF curves are shown in Figure 5.
Figure 5 - BRDF curves (near infrared band) of radiance tendencies due to the joint effects of the solar illumination angle and the view angle gathered under increasingly specific feature mask (from the left side image in Figure 3).
From the curves in Figure 5, one can see that trees (and vegetation in general) are much brighter on the left side of the image. Hardwoods are less affected by the phenomenon than softwood, the extreme case being, the lit-sides of hardwood tree crowns which appear barely affected at all (yellow line). The effects of the various feature-based correction curves are shown in Table 3.
| Cover type | Spectral band | Without BRDF | Vegetation based | ITC based | ITC(lit) based | ITC (S/H) | ITC(lit) (S/H) |
|---|---|---|---|---|---|---|---|
| Softwood | nIR | 24% | 2.1% | 1.1% | 0.5% | 2.6% | 6.9% |
| Red | 31% | 9.3% | 4.6% | 10.6% | 5.7% | 1.3% | |
| Green | 29% | 7.2% | 1.6% | 7.5% | 5.8% | 1.4% | |
| Mean | 28% | 6.2% | 2.4% | 6.2% | 4.7% | 3.2% | |
| Hardwood | nIR | 14% | 3.9% | 4.6% | 5.7% | 2.4% | 5.5% |
| Red | 6% | 22.7% | 22.3% | 16.6% | 10.8% | 2% | |
| Green | 14% | 9.2% | 9.5% | 6.5% | 7% | 3.3% | |
| Mean | 11% | 11.9% | 12.1% | 9.6% | 6.7% | 3.6% | |
| In general | Average | 20% | 8.7% | 7.3% | 7.9% | 5.7% | 3.4% |
Table 3 displays a progression towards increasingly similar spectral signatures (for both softwoods and hardwoods) as the two overlapping images are corrected and normalized with BRDF correction curves based on increasingly precise features. It shows that for two sample areas (one of softwoods, one of hardwoods) found on both images, the intensity differences between images could be brought down from 28% and 11%, respectively, to 3.2% and 3.6%, respectively. Although these results illustrate the strength of the BRDF approach, such residual differences may still be too big to ensure consistent spectral-based species classification throughout an area covered by multiple flight lines. However, one should take into consideration that these two sample areas were at 20° and 25° off-nadir, respectively. If we stay to closer to nadir, these differences may become more manageable.
For this reason, and others having to do with the quality of crown delineation, a sidelap of 50% is recommended between flight lines such that the areas being analysed on each image are within ±15° of nadir. This should allow the ITC analysis to concentrate on only the central portions of each image or flight line, while producing decent results through-out the full area of interest. It should also be noted that additional care is needed when contracting-out aerial acquisitions for digital analysis, as their acquisition, manipulation and storage criteria are not quite the same as for visual interpretation of forests (see Table 2).
Project status
- On-going