Maciej Wielgosz
Forsker
Biografi
Min ekspertise ligger innenfor feltet maskinlæring; jeg jobber med segmenteringsmodeller for punktskyer. Mitt hovedfokus er å håndtere utfordringene som stilles av sparsomme deler av punktskyer, spesielt de som er avgjørende for skogbruksapplikasjoner, som seksjonene nær bunnen av trestammer. Mens data fra droner og fly er lett tilgjengelige, kan det å sikre høy semantisk nøyaktighet under behandling være ganske innviklet. Derfor er det behov for nye metoder i instans- og semantisk segmentering av punktskyer.
Sammendrag
Time and motion studies in forest operations benefit from video-based analysis, but manual annotation is time consuming. This pilot study aims to reduce analysis time by developing a deep-learning framework that classifies dashcam video into four work elements: crane out, cutting and processing, driving, and processing. Using a 3D ResNet-50 (PyTorchVideo) trained on manually annotated clips, the model achieved validation F1 = 0.88 and precision = 0.90, showing that spatiotemporal CNNs can capture rele-vant motion and appearance cues in forest environments. Overfitting indicates that more diverse data and better class balance are needed, but the approach shows clear potential to scale automated work-element monitoring and efficiency analysis.
Forfattere
Stefano Puliti Binbin Xiang Maciej Wielgosz Eivind Handegard Nicolas Cattaneo Marta Vergarechea Terje Gobakken Juha Hyyppä Erik Næsset Mikko Vastaranta Tuomas Yrttimaa Rasmus AstrupSammendrag
Accurately determining the age of individual trees is important for understanding forest dynamics, tree growth, site productivity and describing ecological processes. Traditional methods, such as dendrochronological coring, are invasive, labor-intensive, and costly. This study investigates the use of deep learning (DL) to predict tree age from high-density laser scanning data as a scalable, non-invasive alternative. The dataset includes approximately 1700 tree point clouds from approx. 1 K trees across Norway, Sweden, and Finland, encompassing Norway spruce (Picea abies) and Scots pine (Pinus sylvestris) and a broad range of tree age and developmental stages, from young seedlings (1 year) to old trees (∼350 years). Data were collected using terrestrial, mobile, and high-density airborne laser scanning platforms, enabling the development of sensor-agnostic models. We evaluated multiple modelling approaches, from linear regression to transformer architectures, using both training-from-scratch and fine-tuning strategies. Models fine-tuned starting from pre-trained weights from ForestFormer3D's U-Net as well as the transformer architecture (PointTransformerV3) trained from scratch, proved effective for age regression (RMSE ≤23 years). Although our analysis was limited to two tree species, we demonstrated that a single joint age-estimation model can be successfully trained for both species. We demonstrate that models trained on high-resolution data can generalize to lower-resolution, less costly inputs, provided that data augmentations that mimic reduced resolutions are included during training. This study presents a data-driven framework for estimating tree age without destructive sampling. The findings support the potential for AI-based methods to complement or replace traditional age estimation techniques in forest inventory and monitoring.
Sammendrag
Det er ikke registrert sammendrag
