2026-10-13T14:52:38Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33937988 2026-09-18T17:32:40Z category_272 category_39 category_734 portal_5 item_type_3 month_year_09_2026

<p>Experimental design of the ablation study. The three segmentation architectures and four training strategy axes are summarised. For each axis, one hyperparameter family is varied while all others remain at the reference configuration (grey rows). Fixed shared parameters: patch size  px, batch size 8, learning rate 10⁻⁴, maximum 75 training epochs with early-stopping patience of 20 epochs. All configurations were evaluated by 5-fold cross-validation on 810 patches (CIR + DSM, 4 channels, 20 cm GSD), yielding training runs.</p> Adrien Le Guillou (11387999) Manon Brehier (25073240) Jérôme Ammann (25073243) Xavier Dauvergne (10888650) Pierre Stéphan (7319153) Environmental Sciences not elsewhere classified Ecology Biological Sciences not elsewhere classified systematically detrimental across systematic ablation study operational coastal monitoring limited spatial extent likely reflects limitations div >< p %, enabling wall small geospatial datasets genuinely new sites three segmentation architectures removing elevation costs 810 patches derived 5 – 79 trained encoder rather three architectures sites resembling labelled datasets encoder freezing worst strategy within 2 variable estimate tversky loss target species structurally constrained strategy axes still needed spartina </ relative influence predictions agree photographically pre overlap accuracy manual photo invasive cordgrass heterogeneous modality fold estimate fold cross elevation data ecological irrelevance aerial colour 7 pts 3 ha 10 pts <p>Experimental design of the ablation study. The three segmentation architectures and four training strategy axes are summarised. For each axis, one hyperparameter family is varied while all others remain at the reference configuration (grey rows). Fixed shared parameters: patch size  px, batch size 8, learning rate 10⁻⁴, maximum 75 training epochs with early-stopping patience of 20 epochs. All configurations were evaluated by 5-fold cross-validation on 810 patches (CIR + DSM, 4 channels, 20 cm GSD), yielding training runs.</p> 2026-09-18T17:32:28Z Dataset Dataset 10.1371/journal.pone.0358464.t003 https://figshare.com/articles/dataset/_p_Experimental_design_of_the_ablation_study_The_three_segmentation_architectures_and_four_training_strategy_axes_are_summarised_For_each_axis_one_hyperparameter_family_is_varied_while_all_others_remain_at_the_reference_configuration_grey_r/33937988 CC BY 4.0