Effect of UAV Flight Altitude on Wheat Field Classification Accuracy and Operational Efficiency

Flight altitude ties the quality of unmanned aerial vehicle (UAV) imagery to the cost of acquiring it, yet operational guidance on selecting it for crop mapping remains largely qualitative. This study quantifies the trade-off using a repeat-altitude survey of one winter wheat field in Kayseri, Türkiye. A 20 MP RGB camera on a fixed-nadir gimbal acquired 300 frames over the same ground at six altitudes between 10 m and 120 m (0.227–2.725 cm/pixel). After removing 25 frames whose gimbal yaw deviated from the flight heading, imagery was resampled onto a common 5 m × 5 m ground cell grid so that every altitude represents identical ground at an identical input size, making ground sampling distance (GSD) the only variable. A total of 183 colour, colour index and texture features were extracted from 11 251 cell tiles and classified with Random Forest under five-fold spatial block cross-validation with a 10 m buffer; the cell (n = 654) rather than the tile was the statistical unit. For three-class land cover over a 2× GSD range (60–120 m, n = 654), overall accuracy declined monotonically from 0.804 to 0.734 (McNemar p < 0.0001 for 60 m versus both coarser altitudes). For binary wheat extent mapping over a 6× GSD range (20–120 m, n = 186), accuracy was statistically flat from 20 m to 90 m (0.903, 0.909, 0.914; McNemar p = 0.75–1.00) and fell only at 120 m (0.866; p = 0.039 and 0.049); because the 20 m mission itself carried motion blur, this plateau compares effective resolution and not nominal GSD. Under a fixed 75 %/65 % overlap plan, 90 m required 17× fewer images, 19× less storage and 4–17× less flight time than 20 m for equal accuracy, identifying 90 m as the operational optimum for this task and sensor. Processing time did not decrease with altitude, because it scales with ground area and not with image count; a pipeline that classifies each ground cell once instead of once per overlapping frame ran 6–12× faster. A 40 m mission flown two hours later in weaker light required a 1/41 s exposure, which introduced approximately 12 pixels of motion blur and reduced effective sharpness to 34 % of the 60 m mission despite a finer nominal GSD, showing that nominal GSD is not effective resolution.