⛰️So happy to make it without walking and 10min faster than last year: 2:34:16 in this amazing half marathon 21.1km with 1271 metres elevation
∆ Average pace 7:23 min/km, starting with 5:50 min/km first 5km.
@pointtopinnacle #P2P #kunanyi #tassie
👉🏻🐔😌👈🏻

🏌🏻‍♀️Proud to share PhD project 'Mapping Ecosystem Resilience with High-Resolution Remote Sensing Data' was selected for the
NVIDIA Academic Grant Program 👩🏻‍💻
Thanks to #NVIDIA, I’ll have 1K A100 GPU-Hours on Brev 🚀 #NVIDIAGrant #DeepLearningAlgorithms #ComputerVision #Robotics #GenerativeAI @NVIDIAAIDev

Implementation of the Monte Carlo Dropout as Uncertainty estimation in FVC segmentation mapping 🗺️#DeepLearning #RemoteSensing
[6/8]

Cm-scale FVC maps🌱from our best CNN models predictions on 3072×3072 px UAS multispectral tiles for each site (site-specific models).
[5/8]

Model generalisation & transferability 🛰️ We tested 2 setups (A & B) to see when site-specific models beat generic ones. Expanded discussion on limitations + showed how data augmentation can boost robustness as a proof-of-concept.
[4/8]

U-Net models were trained & validated with spatial block CV, including both site-specific and generic models across all sites, to produce more realistic and transferable evaluations. Random CV inflated performance estimates by up to 28%.
[3/8]

🌱📍We tested our CNN approach at 3 AusPlots sites in Calperum Station, South Australia, each representing a different National Vegetation Information System (NVIS) vegetation type: A (low), C (medium), E (dense). Together they cover 29 ha of semi-arid landscape.
[2/8]

[Session BG9.2]"Remote Sensing for forest applications" Hall X1.73| 3D voxel cubes from #UAS imagery and #LiDAR to train a 3D U-Net for dense segmentation of vegetation strata, enabling scalable ecosystem mapping across sensors and landscapes.🌿📡 #RemoteSensing #DeepLearning #drone #EGU25 @egu.eu