High-spatial resolution UAV multispectral data complementing satellite imagery to characterize a chinstrap penguin colony ecosystem on deception island (Antarctica)

Figures & data

Table 1. Comparative table of the characteristics of satellite and UAV imagery.

Satellite Imagery	UAV Imagery
Global and synoptic vision of the Earth’s surface	Higher spatial resolution (centimeter/pixel)
Complete spatial coverage of whole regions Frequent repeat coverage	Higher control of temporal resolution, although it does not cover areas as large as the satellite
Mostly publicly and freely available accessible data	Low cost equipment compared to other remote sensing tools
High dependence on the weather in data collection	Data capture below the clouds

Figure 1. Maps showing the location of (a) South Shetland Islands and most concretely, (b) Deception Island (Antarctica) and (c) the orthomosaic of the Vapor Col Chinstrap penguin colony generated with the UAV flight (RGB composite with bands Red-668 nm, Green-560 nm and Blue-475 nm, Micasense RedEdge-MX dual sensor achieving 6.04 cm/pixel size). (d-e) Pictures taken in the colony during the XXXIV Spanish Antarctic Campaign (February 2021) are also shown.

Figure 2. (a) Condor hexacopter used in this research. (b) Micasense RedEdge-MX dual multispectral camera. (c) Sentinel-2, Landsat 8 and Micasense RedEdge-MX spectral bands comparison.

Table 2. Input variables selected for each image classification algorithm in SAGA GIS software.

SVM	RF	MLC	SAC
Model Source: Create from polygons	Model Source: Create from polygons	Model Source: Create from polygons	Model Source: Create from polygons
Type: C-SVC	Number of Samples: 5000	Probability Threshold: 0	Spectral Angle Threshold: 0
Kernel: radial basis function	Number of trees: 100	Probability Reference: relative
Degree: 3	Minimum Sample Count: 2
nu-SVR: 0.5	Maximum Categories: 15
SVR Epsilon: 0.1	Regression Accuracy: 0.01
Cache Size: 100
Epsilon: 0.001

Figure 3. Zoom of the Vapor Col penguin colony. MLC, SAC, RFC and SVM algorithms were applied to make the supervised classifications showed in the figure.

Table 3. Accuracy Assessment of thematic maps generated from UAV imagery using the four algorithms tested in this study (MLC, SAC, SVM and RFC), including user accuracy (“U-acc”), producer accuracy (“P-acc”), the OA (%) and the Cohen’s Kappa index.

Class	SVM		RFC		MLC		SAC
	U-acc	P-acc	U-acc	P-acc	U-acc	P-acc	U-acc	P-acc
Soil	94.71	97.84	90.35	95.52	99.27	88.89	95.02	85.93
Snow	98.92	82.94	98.89	68.78	87.15	99.95	95.06	45.98
Guano	78.86	99.96	79.51	99.96	77.22	98.89	76.38	98.46
Moss specie 1	86.73	78.67	95.12	29.47	82.31	40.23	76.24	60.35
Moss specie 2	89.69	49.61	94.81	34.27	77.44	77.44	75.01	63.21
Moss specie 3	92.68	27.50	92.47	97.96	48.36	48.37	49.73	78.68
Overal Accuracy (%)	93.19		90.83		89.07		82.01
Cohen’s Kappa	0.88		0.85		0.77		0.62

Figure 4. SVM thematic maps obtained from the UAV, S2 (L1 and L2) and L8L2 imagery, with three highlighted regions of interest. In S2 (L1 and L2) and L8L2 RGB composite images, the area covered by the UAV in Vapor Col is indicated in yellow. Dates: UAV (8 February 2021), S2 (2 February 2021) and L8L2 (12 January 2021).

Figure 5. Land cover classes reflectance spectrum for each sensor (UAV, S2 (L1 and L2) and L8L2). It should be noted that in the case of the UAV, the vegetation substrate has three different classes. Wavelength (nm) is represented in the x-axis and reflectance (dimensionless) is represented in the y-axis.

Figure 6. Pixel-by-pixel and band-by-band comparison of UAV surface reflectance data with S2 Level 2 atmospherically corrected product data (triangles) and with S2 Level 1 orthoimage product data (circles). Twenty pixels (n = 20) were selected in four different types of substrate: snow (gray color), guano (brownish color), soil (dark brown color) and vegetation (green color).

Table 4. Statistics calculated (R², RMSE, MAE and bias) for the linear regression models applied to the relationships between the multispectral sensor equipped in the UAV and S2 (L1 vs L2) for comparison.

		R^2	RMSE	MAE	bias
Blue-475 nm Band	L1	0.07	0.17	0.11	−0.03
	L2	0.79	0.09	0.06	−0.04
Green-560 nm Band	L1	0.06	0.20	0.12	−0.09
	L2	0.80	0.09	0.07	−0.06
Red-668 nm Band	L1	0.01	0.24	0.15	−0.14
	L2	0.74	0.14	0.10	−0.10
RedEdge-705 nm Band	L1	0.01	0.24	0.17	−0.16
	L2	0.64	0.19	0.14	−0.14
RedEdge-740 nm Band	L1	0.03	0.25	0.19	−0.18
	L2	0.55	0.19	0.15	−0.14
NIR-840 nm Band	L1	0.04	0.25	0.20	−0.19
	L2	0.63	0.13	0.10	−0.10

Figure 7. Sentinel-2 L2 true color composition on 2 February 2021, in which sediment plumes have been pointed out with red arrows.

Data Availability Statement

The authors confirm that the data supporting the findings of this study are available within the article [and/or] its supplementary materials.