Overall, the model did a poor job of identifying tilled land within the study area. As outlined in objective 2, the model was designed around utilizing WhiteboxTools' MaxAnisotropyDev tool. An assumption was made that the tool would recognize tilled land due to its strong anisotropic geometry while ignoring the rest of the data within the raster due to being generally isotropic. Once the model had been run, the results from the raster images invalidated this assumption. In some fields, the furrow/ridge systems were visually identifiable based upon their shape. However, the unitless values of anisotropy that were assigned to cells within the tilled fields were typically very low, with shape being the sole indicator of tillage. A sample of the final raster for each grid resolution can be seen in Figures 4 through 7 below. Additionally, the MaxAnisotropyDev tool identified areas without tillage as being strongly anisotropic. Areas that had previously been covered in forest prior to groundpoint filtering were identified as being more strongly anisotropic, although often without any identifiable shape. In places where off-terrain features had been removed, the MaxAnisotropy tool identified these areas as being highly anisotropic. Areas where features such as roads, gravel piles, and building footprints had once been all had very high anisotropic values associated with them as well. An example of this misidentification can be seen in the northwest quadrants of Figures 4 through 7; the red polygon-like shapes are all building footprints.
Figure 4 - MaxAnistrophyDev tool raster image on a 1.0 m raster image
Figure 5 - MaxAnistrophyDev tool raster image on a 0.75 m raster image
Figure 6 - MaxAnistrophyDev tool raster image on a 0.50 m raster image
Figure 7 - MaxAnistrophyDev tool raster image on a 0.25 m raster image
In an attempt to try and extract tillage features within the anisotropic raster based on potential difference in statistics, the StandardDeviation tool within WhiteboxTools was run using a filter size of 11.0 metres. Unfortunately, the deviation between tillled and non-tilled areas is effectively the same.
The reason why the MaxAnisotropyDev tool failed to function as expected remains unclear, but there are potential reasons. During discussion with the creator of the tool, Daniel Newman, it was revealed that the tool does not function as intended when a small kernel size is used, as the results are statistically too varied to mean anything. This might explain why the results were the opposite of what was expected, where areas without any significant anisotropic features being identified as such. Another reason why the tool may have failed to work has to do with the data being input into the tool, specifically the raster image created from the Subtract tool. Discussion concerning cellsize follows below.
Differences Between Gridsize Rasters
As per objective 2, it was deemed necessary to test the process on multiple raster images that had varying grid resolutions in order to address the range of widths that furrows can be created at. It was expected that decreasing the gridcell size of a raster would increase the accuracy of the results, or at least increase the anisotropic value associated with cells in a furrow, since the size of the features would increase and become more defined. As noted above, the process did not work as intended. Regardless, there was still a noticeable effect when the grid cell size was decreased, and it had the opposite effect of what was predicted. When the difference rasters were being inspected after being processed through the subtraction step, it was noted that the amount of noise within each image was increasing as the grid cell size decreased. The increase in image noise is most evident when looking at Figures 8 through 11. There are some furrows that lose all shape in the 0.25 meter raster images that were previously distinguishable in coarser rasters.
Figure 8 - Difference in elevation between Smoothed and Regular 1.0 m DEMs
Figure 9 - Difference in elevation between Smoothed and Regular 0.75 m DEMs
Figure 10 - Difference in elevation between Smoothed and Regular 0.50 m DEMs
Figure 11 - Difference in elevation between Smoothed and Regular 0.25 m DEMs
Validation of Results and Evaluation
Due to the fact that the results of the model were difficult to interpret, there would be no spatial data of tilled and non-tilled land to compare to a pre-existing data set. Furthermore, no pre-existing spatial data could be obtained that was collected on a field-by-field basis. The smallest unit within the Census of Agriculture is the census subdivision. The Trent Hills census subdivision is larger area than the study area in Northumberland county, so there is already uncertainty in comparing the results. Additionally, since the census is not mandatory, not all farmland tilling practices are reported back to Statistics Canada. For example, for the census subdivision of Trent Hills, a total of 13 587 hectares were reported where 51.1 percent of the land was not tilled and 48.9 percent was tilled. However, when looking at the Agricultural Resources Inventory in Trent Hills, the area of potentially tilled land is 31 361 hectares. Only 43.3 percent of the potentially tilled farms within the study area reported their activities regarding tillage. Since the identification process was going to be identified via individual farms in the study area, significant asusmptions would have to be to use this aggregate data as a means of evaluation. Also, the Census of Agriculture occurred in 2016 while the LiDAR acquisition happened in 2017. There is a possibility that some farms could have changed their farming practises the next year if crop rotation was being used. This further complicated any validation that could be made.
If the identification process was successful but the only pre-existing data that is present was the Census of Agriculture could be used for evaluation, two key assumptions would need to be made to make a correct aggregate comparison. First of all, the results obtained for the Trent Hills subdivision would need to be normally distributed spatially inside the study area. Secondly, the tilling practises of all farms in Trent hills subdivision could not have changed between the year 2016 to 2017. With those assumptions, an aggregate comparison of the percentage of land in hectares reported as till and no-till operations could be used as a reference of validation to the percentage of hectares reported in the identification process of tillage to the study area of Northumberland county.
Since the identification model proved to be ineffective, there was no purpose in determining the limits of the model with coarser resolution DEMs such as those that could be obtained from the South Central Ontario Orthophotography Project (SCOOP). Should a different approach succede in identifying tillage from a DEM, it would be worthwhile to evaluate the model using coarser DEMs and differing interpolation methods.
Considering the poor performance of the MaxAnisotropyDev tool, it is not the appropriate tool to be used to identify tillage. However, parts of the process can be considered a success, and should continue to be built upon. All processing prior to the application of the MaxAnisotropyDev tool performed satisfactorily. Rather than trying to identify tillage using values based upon the output of the MaxAnisotropyDev tool, the identification process could shift to focus on object identification. The LaplacianFilter tool within WhiteboxTools can be applied to the difference rasters in order to reduce image noise and sharpen the tillage features. A boolean image can be created from the laplacian filtered difference raster, with the threshold values being determined by visual inspection. The Clump tool within WhiteboxTools is then used, which assigns each distinct boolean object a unique ID number.The ElongationRatio tool within WhiteboxGAT can then be applied to these objects, which assigns values based upon the ratio of the object's perimeter to its area. Since the furrows are linear, they should have a high elongation ratio, and should be filtered by using the raster calculator once more. The specific threshold values used for filtering will need to be determined, and additional filtering may be required before the first raster calculator can be applied. Additionally, further research need not concern itself with processing data at a grid resolution less than 0.50 metres as the noise introduced at such fine resolutions muddles any data that may be present.
Additionally, future research needs to be concerned with validation of their model with real world results. Coordinating with farmers and a LiDAR acquisition company will allow for the results to be more intensely scrutinized. In-field test farms would require inclusion of a specific farm that would report their tillage practices to a geographical spatial boundary that can be mapped inside the study area. This way, results from the identification process could be directly compared to the in-field test farms. If the identification process proved to map the exact same amount of tilled and non-tilled land on the test site then such a model would be proven to be accurate. Including more test farms would statistically validate the identification process results with greater significance.