Vock, Clarissa; Spyrakos, Evangelos; Neder, Camila; Weith, Friederike; Gomez Farfan, Rosby R; Stark, Juliane V; Bracher, Astrid; Richter, Claudio; Schloss, Irene R; Barrera, Facundo; Jerosch, Kerstin: Unmanned Aerial Vehicle-based Suspended Particulate Matter Concentrations, King George Island, Antarctica (ANT-Land-SIGMA-I, January - March 2023 and ANT-Land-SIGMA-II, December 2024 - February 2025) [dataset]. PANGAEA, https://doi.pangaea.de/10.1594/PANGAEA.997987 (dataset in review)

This dataset contains raster data with retrieved Suspended Particulate Matter (SPM) concentrations from mosaicked Unmanned Aerial Vehicle-based data from a DJI Phantom 4 Multispectral quadcopter UAV, coupled to a GNSS reference mobile station DJI D-RTK 2. The dataset was collected from January until March 2023 in Maxwell Bay (Potter Cove and Marian Cove), and Admiralty Bay (Ezcurra Inlet and Lussich Cove) in the ANT-Land-SIGMA-I campaign, and during the austral summer of 2024/25 in Admiralty Bay (Mackellar Inlet and Lussich Cove, December 2024), and Maxwell Bay, including Potter Cove (January, February 2025), Marian Cove (January 2025), and Collins Harbour (February 2025), as well as Stranger Point (February 2025) within the ANT-Land-SIGMA-II campaign, using the DJI GS Pro App (V. 2.0.17), which automatically adjusts the flight path to minimize battery consumption and achieve the desired capture overlap of 30% or a front overlap of 90% and a side overlap of 60%. The captures were taken at an altitude of 100m when possible, but the altitude was chosen lower during unfavourable weather conditions. UAV flights were conducted regardless of cloud cover, which was accounted for by the incident light sensor at the top of the UAV, solar angle, and the tidal state, but only at wind speeds below 10 m/s. Image post-processing, i.e., band alignment, vignetting, distortion correction, radiometric adjustment, and direct georeferencing, was performed using the DroneWQ library (https://github.com/aewindle110/DroneWQ), which was specifically designed for multispectral mapping of aquatic surfaces. This was adapted to be compatible with DJI P4M metadata and to follow the DJI processing guide (DJI 2020). Ground calibration panels were unavailable during fieldwork. For practicality, the conversion factor ρ_NIR between the camera and the sunlight sensor signal values was assumed to be 1. The obtained reflectance is therefore only proportional to a physically meaningful reflectance. Data collected during the second campaign contained strong sensor noise. To mitigate salt noise, affected images were filtered using a localised median filter (11x11 kernel). Georeferenced UAV captures were mosaicked using the "Mosaic to New Raster" tool in ArcGIS Pro (3.6.2), whereby the mean value of overlapping pixels was chosen. A random calibration-validation split (80, 20) of the matchup in situ dataset (n=162), which contains extracted measured surface SPM concentrations and R_D, extracted at each in situ water sampling location as the median within a buffer of 0.5 m radius around the sample point, was created for model comparison. An empirical model employing red-blue and blue-green band ratios (modified after Siswanto et al. (2011)) was fitted using a non-linear least squares regression. The resulting model is of the form 10^((c_0+c_1 ((R_D (650))/(R_D (450))) + c_2 ((R_D (450))/(R_D (560))))), with c_0=1.43, c_1=0.49, and c_2= -0.40. The model's performance, given as RMSE [mg/L], R^2, and SSPB(%), is 5.85, 0.60, 3.83, and 7.88, 0.65, 5.92, on the calibration and validation datasets, respectively.