Knock Hill Farm

• Involves the use of geospatial technologies to improve agricultural management
• Utilising remotely sensed imagery, such as Earth observation satellite images and unmanned aerial vehicle (UAV) photographs to generate information about the land
• This is in contrast with field- and ground-based techniques, such as using the global positioning system (GPS)

To utilise geospatial and remote sensing data, namely Earth observation and UAV imagery, to generate insights about Knock Farm’s land cover and topography

• Sentinel-2 10 m image, with four spectral bands: blue (B), green (G), red (R), and near-infrared (NIR)
• Sentinel-2 imagery has high spatial, temporal, and spectral resolutions, suitable for land cover mapping
• Different spectral band orders produce different composites
• True colour (B-G-R) and false colour (NIR-R-G) composites are shown in the left and right, respectively

A 6-class land cover classification scheme for Knock Farm based on the National Land Use Database version 4.4:

• Cropland (e.g. field crops, horticulture, orchards)
• Grassland (e.g. permanent grass)
• Woodland (e.g. mixed and conifer woodlands)
• Shrubland
• Bare land (e.g. natural bare surfaces, harvested fields)
• Buildings or rocky surface (e.g. infrastructure, hilltops)

• Similar land cover units will have similar spectral profiles
• Points with high intensity of greenness will have a steep curve between NIR and R
• NIR has high reflectance and R has low reflectance of green vegetation, respectively
• The ratio of NIR to R is known as normalized difference vegetation index (NDVI), a form of unsupervised classification to identify different vegetation units
• Unsupervised classification, which is different from NDVI, clusters similar-coloured pixels together and assigns them to classes, the amount of which is user-defined
• Both classifications are done for the Knock Farm imagery, and class names and colours are assigned afterwards

• In contrast with unsupervised classification, this requires more user input
• Six areas of interest polygons are constructed on the satellite imagery to represent each land cover class
• Band combinations are checked to determine correlation
• This can help contrast the different classes better
• NIR-R-G image was chosen as the composite for classification, due to low correlation between NIR-R
• The supervised classifier will be trained to detect and classify pixels similar to the pixels for each class covered by the areas of interest polygons

• Assessed after it produces an output
• 256 random control points are generated on the output raster, which is then manually compared with the NIR-R-G composite to assign reference classes
• Once assigned, an accuracy report can be generated, which has detailed statistics

• Aerial images from UAVs can be used to generate orthomosaics and digital elevation models (DEMs) of an area
• This is known as orthomapping and softcopy photogrammetry
• 15 drone imagery captured over a section of Knock Farm was made available
• The images are first orthorectified using ArcGIS Pro’s ortho mapping tools to perform block adjustments and reduce uncertainties caused by the UAV and camera for more accurate outputs

• The NDVI map (left) has five land cover classes and pseudocolours, which have been manually assigned for the row numbers denoted
• The unsupervised classifier was told to produce six classes (right)
• Both classifiers interpret woodland and bare land differently

• The six land use classes have been assigned manually
• Compared to the unsupervised classifiers, the pixel clusters are much larger, notably around the foot of Knock Hill towards the north-west

• Grasslands are the most misclassified, with 23/39 being incorrectly classed as buildings or rocky surfaces and only 14/39 being correct
• Misclassification may be attributed to insufficient contrast in the imagery or human error, which may be rectified using additional image composites
• The kappa statistics show that the classifier and user were in agreement the most for croplands, followed by woodland and shrubland
• Overall, the classifier and user agreed on the control point’s classes 53 % of the time

• This exercise demonstrated the use of remotely sensed data to generate land cover maps and ortho maps for precision agriculture
• Varying accuracies were found in both classification tasks and the orthomapping
• These accuracies may be improved further by obtaining satellite imagery from different time periods, using more band orders and combinations, and using additional sources of imagery, as well as ground-truthed data