Our * goal :
Support diary farmers to improve their management of cattle feed production by integrating farmers expert / historical knowledge with state of the art use of platforms such as drones, imaging sensors, image analysis and GIS.
Borre brs ref. dronexpert.nl ,are integrating our SpectroCam UAV version with their DJI configurable drones. They also set up a grassland test field near their location in Bentelo.
Through the growing season, about weekly at least fly the sensor at 20 m or 40 m. At Z= 20 m the spectrometer has a projected FOV width of 2 m across the direction of movement. At Z= 40 m the width is 4m. The forward speed of the drone in combination with spectrum integration time ensures spectral FOV_ground samples of 2mx2m or 4mx4m . Each spectral samples corresponds to about 32×32 pixels in the coaxial RGB image. After correction of all images for radiometric errors (vignetting) the images or transformed into an (ortho)photo mosaic. The interface between SpecroCam and DJI M100 provides Latitude, Longitude, Height and Quaternion position and attitude info.
As all data are referenced to the same UTM grid, we can and should add layers of relevant data including time series. The above Google picture hints at how the local brook was instrumental in forming the landscape. Geomorphology provides info on what to expect in terms of soil patterns and soil properties.
The test field was prepared starting as a bare field with 2 types of grass mixtures sowed. The planned 3 differences in manure application will provide a total of 6 main variants .
What can be learned from the first “bare soil”data set ?
RGB corrected > orthomosaic : this is used for testing the assumption of little variation in soil spectral reflectance. Spatial variation is estimated from likelihood for clusters in RGB space: the next figure shows a subset of the Mosaic with segment boundaries derived from k-means clustering with 5 classes.
Figure 2. k= 5, clusters in RGB space based on k-Means clustering.
Soil01 cluster has mean-RGB = [ 176 134 138 ]
Soil02 cluster has mean-RGB = [ 199 133 138].
Conclusion: the bare soil segments are quite homogeneous and only vary in the reflected photon flux in the red band. So the size and position of the spectral sample’s FOV on the ground is not critical.
How do clusters in RGB relate to clusters in the SpectroCam’s 1024 spectral channels ?
Figure 3. Clustering on model: if bare soil then error in polyfit 2nd order polynomial < 0.007, else likely mix of soil and vegetation reflection / absorption.