In excess of 12,000 oblique geo-referenced images of agricultural fields were captured over three counties in southern Ontario on two dates in May and one in November of 2016. Data collection for this project was performed using a single, continuously moving, multi-camera mobile imaging survey vehicle designed for the sole purpose of reducing both environmental and financial costs related to the standardized sampling practices. These oblique images were evaluated against a sample of 155 static nadir images captured in-situ from OMAFRA research field plots in effort to expand upon both the sampling size, and classification accuracy in the quantification of post-harvest tillage practices employed in the region. 

Initial analysis of the mobile roadside data corresponds with a very high degree of confidence to the methods currently in use, which require multiple vehicles, dozens of people, and physical trespass for data collection. The benefits in having significantly greater coverage than afforded by physical field measures, at a reduced cost, both in personnel, equipment, and time are significant. Rather than making decisions based on extrapolating a sample of the field tillage practices for each of the three counties, every field (between 3,000-3,500 per county) was captured. 

To expedite the manual task of sorting and classifying each image, a machine learning semantic segmentation protocol was developed which has not only proven to be reliable in the classification of each field with a very high level of agreement to those gathered in-situ, but was shown to be transferrable to different locations under different atmospheric conditions (clear, scattered cloud, and overcast). Current field sampling methods, while acceptable for aiding in validation of orbital remote sensing data (e.g. Landsat) classification are contingent on clear-sky conditions, whereas the mobile roadside method is not. Such mobile ground-based survey vehicles, as described in this research may also be deployed when required, as opposed to field survey scheduling being reliant on windows of satellite overpass. 

The cost savings, and confidence level of both county and province-wide post-harvest tillage assessment is determined to be significantly increased using the methods presented here. The classification of such high volumes of data addressed via semantic segmentation machine-learning processes, effectively permits up to 40 times the data, at one-tenth the cost (ecological and financial), in one half the time. 

As the geo-referenced images may be batch imported into a GIS database, and/or directly overlain upon basemaps or remote sensing imagery, rapid evaluation of large area land-use classification, and/or change detection studies may be performed in an expeditious manner. Such data is currently being used for calibration and validation of fused Landsat and Radarsat image classification, optimally to aid in policy decisions at both the Provincial and Federal level.