Y.Remote Sens. 2021, 13, 4181 Remote Sens. 2021, 13, x FOR PEER REVIEW4 of 18 4 of(a)(b)(c)(d)(e)Figure 1. A comparison of a element in the utilised education information to evaluate which from the visualisation algorithms had been improved Figure 1. A comparison of a component with the utilised coaching data to evaluate which from the visualisation algorithms were better suited for the detection of burial mounds: (a) satellite view of that area together with the identified tumuli marked; (b) DTM; (c) suited for the detection of burial mounds: (a) satellite view of that location with all the identified tumuli marked; (b) DTM; (c) MSRM; MSRM; (d) slope gradient; (e) SLRM. (d) slope gradient; (e) SLRM.two.2. Deep Learning Shape Detection 2.2. Deep Understanding Shape Detection For the DTM-based shape detection we made use of YOLO [29], an R-CNN-based algorithm For the DTM-based shape detection we applied YOLO [29], an R-CNN-based algorithm previously employed within the field of archaeology for the detection of inscriptions in oracle previously employed in the field of archaeology for the detection of inscriptions in oracle bones [30]. The YOLOv3 algorithm is faster than other R-CNN strategies like Faster Rbones [30]. The YOLOv3 algorithm is faster than other R-CNN strategies like More quickly RCNN. Its backbone, Darknet-53, is 1.five times faster than ResNet-101, functioning at 78 frames CNN. Its backbone, Darknet-53, is 1.5 instances more quickly than ResNet-101, working at 78 frames per second [29,31]. YOLOv3 predicts atat 3 various scales, YQ456 Formula whichsimilar to what the per second [29,31]. YOLOv3 predicts 3 distinctive scales, that is is similar to what feature pyramid network does [29,32]. This structure allows enables the detection of tiny the function pyramid network does [29,32]. This structure the detection of little objects. The bounding boxes are predicted by the anchor boxes generated working with k-meansk-means objects. The bounding boxes are predicted by the anchor boxes generated employing clustering with an Intersection over Union (IoU) threshold of 0.five. The 0.5. The class prediction clustering with an Intersection over Union (IoU) threshold of class prediction is created employing binary cross-entropy loss and independent logistic classifiers, the latter thefacilitate is created using binary cross-entropy loss and independent logistic classifiers, to latter to multilabel classification [29]. facilitate multilabel classification [29].Remote Sens. 2021, 13,5 ofRemote Sens. 2021, 13, x FOR PEER REVIEWAn Nvidia Titan XP graphics processing unit (GPU) with 12 GB of RAM hosted in the five of 18 Pc Vision Center (CVC) on the Autonomous University of Barcelona (UAB) was applied to run the DL algorithms. The selected perform environment was the parallel computing platform CUDA 11.two, the ML library Tensorflow two.1.0, the DL library cuDNN 8.1.1, the An improvement tool CMake 3.20.2 and unit (GPU) with 12 GB of RAM hosted at software program Nvidia Titan XP graphics processing the CV library OpenCV four.5.2 as advised the YOLOv3 [33]. Personal computer Vision Center (CVC) on the Autonomous University of Barcelona (UAB) for was employed to run the DL algorithms. The we made use of the environment wasburial mounds data As instruction and validation data, chosen operate present recognized the parallel computing platformthe research led ML M. Carrero-Pazos two.1.0, B. Vilas [16,34] in Xestospongin C supplier Galicia and obtained from CUDA 11.2, the by library Tensorflow as well as the DL library cuDNN 8.1.1, the software improvement tool CMake three.20.two as well as the CV library OpenCV four.five.two as recomJ. Fonte in the location of Northern Portugal (Figure 1). T.
