Image Capture and Initial Data Processing
Drone field mapping begins with the drone following a predetermined flight path and capturing overlapping images. The drone records image files and associated metadata: camera position (latitude, longitude, altitude), camera orientation (roll, pitch, yaw), and timestamp. This metadata is essential—it provides initial estimates of where each image was taken and what it shows.
For multispectral drones, raw imagery from multiple spectral bands (red, green, blue, red edge, near-infrared) is captured separately. Radiometric calibration is the first processing step, converting sensor values (which depend on lighting, atmospheric conditions, and sensor gain) to standardized reflectance values (which represent the actual light reflected by landscape features). This calibration uses reference images or field-measured values to normalize sensor output. Radiometric calibration is essential for consistent NDVI across different flights and lighting conditions—without it, NDVI values would drift based on time of day or weather.
Feature Detection and Image Alignment
Once images are radiometrically calibrated, photogrammetry software identifies distinctive features in each image—sharp edges, corners, textured areas—and matches corresponding features across overlapping images. This process, called feature matching or image alignment, determines how much each pair of adjacent images overlaps and from what angle they were captured.
For example, if two adjacent images both contain a distinctive tree, the software identifies that tree in both images and calculates that the tree appears in slightly different positions between images. This positional shift reveals the relative positions and orientations of the two camera positions when the images were captured. With thousands of matched features across dozens or hundreds of images, the software constructs a mathematical model of the 3D landscape and the camera positions throughout the flight.
This process, called structure-from-motion (SfM), reconstructs not just camera positions but also the 3D coordinates of matched features. The result is a sparse 3D point cloud—thousands of 3D points representing features in the landscape, positioned relative to camera locations. This point cloud forms the geometric foundation for subsequent processing.
Georeferencing and Coordinate System Registration
The point cloud reconstructed from image matching is geometrically accurate but not absolutely positioned—it exists in an arbitrary coordinate system. Georeferencing anchors this point cloud to real-world coordinates (latitude, longitude, elevation), enabling the orthomosaic to represent actual field locations.
Georeference data comes from two sources: GNSS positioning during flight (recorded in image metadata) and ground control points (surveyed field locations). RTK-equipped drones provide centimeter-level GNSS accuracy throughout the flight, which DroneField uses to directly georeference the point cloud with minimal additional processing. Drones without RTK rely on standard GNSS (meter-level accuracy initially) plus ground control points for refinement. Ground control points are surveyed field locations whose images appear in drone photos; by measuring how these surveyed points appear in imagery, the software refines the coordinate transformation, improving absolute accuracy.
This refinement process, sometimes called aerotriangulation or bundle adjustment, solves for the optimal coordinate transformation that places surveyed ground control points at their correct measured positions while maintaining internal geometric consistency. The result is a georeferenced point cloud whose coordinates correspond to real-world locations.
Orthomosaic Generation and Blending
With a georeferenced 3D point cloud established, generating an orthomosaic is the next step. The software projects each image onto a horizontal plane (viewing straight down), removing perspective distortion. This orthogonal projection transforms the tilted, perspective-distorted raw images into vertically oriented, geometrically corrected views aligned with a coordinate grid.
Once all images are projected orthogonally, they are stitched together into a seamless composite mosaic. Color blending is critical: without it, seams between adjacent images would be visible due to lighting variations and exposure differences. The blending algorithm smooths color transitions along image boundaries, creating a seamless composite that appears uniformly lit. The final result is a single large image (often hundreds of megapixels for a full field) where every pixel corresponds to a known real-world location.
The orthomosaic can be exported in various formats and resolutions. GeoTIFF format preserves georeference information and is compatible with GIS software. Smaller web-viewable formats (PNG, JPEG) enable easy sharing and visualization. Ground resolution (how many centimeters each pixel represents) depends on flight altitude and camera specifications: lower flights or cameras with longer focal lengths produce finer resolution; higher flights produce coarser resolution.
Spectral Analysis and NDVI Computation
For multispectral drones, spectral analysis proceeds in parallel with orthomosaic generation. Once radiometrically calibrated multispectral images are orthorectified (projected to horizontal plane and georeferenced), spectral indices like NDVI are computed.
NDVI is calculated pixel-by-pixel from orthorectified near-infrared and red band images. For each pixel, the formula NDVI = (NIR - Red) / (NIR + Red) is applied, producing an NDVI value between -1 and +1. The result is an NDVI orthomosaic—a georeferenced grid where each pixel represents the vegetation vigor at that real-world location. NDVI values are often visualized using color ramps: red tones represent low NDVI (sparse or stressed vegetation), green tones represent high NDVI (dense, healthy vegetation).
Other vegetation indices (Red Edge Index, Green Normalized Difference Index, chlorophyll estimates) can be computed similarly from appropriate spectral bands. All these indices are delivered as georeferenced products in GIS-compatible formats, enabling spatial analysis and integration with farm management software.
Quality Control and Accuracy Assessment
Throughout the processing pipeline, quality control checks ensure output reliability. Photogrammetry software validates internal consistency: do feature matches remain accurate when the entire point cloud is adjusted for georeferencing? Are there regions where image coverage is sparse or misaligned? Automated checks flag potential problems; operators review flagged areas and reprocess if necessary.
Accuracy is validated through comparison with known surveyed points. If ground control points were used, residuals (differences between predicted and actual GCP positions) indicate georeferencing quality. Accuracy assessment also considers orthomosaic visual quality: are seams obvious? Are features crisp or blurred? These visual indicators suggest whether processing parameters (feature matching sensitivity, blending algorithm) are appropriate.
DroneField provides accuracy estimates with each orthomosaic, typically 5-30 cm absolute accuracy depending on GCP availability and RTK precision. For most agricultural applications, this accuracy is adequate. Documentation of accuracy and processing parameters enables users to understand confidence levels and appropriate use cases for their maps.