Conference Agenda

Beef cattle farming is a significant activity in Brazil, and forage quality has a direct impact on animal performance. However, traditional methods for estimating dry mass, which involve cutting, drying and weighing plant material, are slow and labor-intensive. UAVs equipped with multispectral sensors, such as the DJI Mavic 3M, offer a faster and more scalable alternative for monitoring mixed-forage pastures. This study estimates the dry mass of forage mixtures using multispectral UAV data in two scenarios: (i) using only spectral information and (ii) combining spectral data with canopy height measured in the field. Model performance was evaluated using R², RMSE, and percentage error. The multispectral-only model explained 55% of dry mass variability (720.56 kg/ha; 23.67%), while adding canopy height improved performance to 80% and reduced the error to 589.41 kg/ha (19.36%). Results show that canopy height enhances the accuracy and operational potential of UAV-based methods for estimating dry mass in mixed-forage areas.

Mining companies are continuously looking for cost efficient methods to monitor the success of their rehabilitation efforts. Although open access satellite imagery is available at regular temporal intervals, its usefulness for grassland biodiversity monitoring has been questioned due to its coarse spatial resolution with respect to the species size. To compensate for the low spatial resolution, previous studies have successfully explored the benefits of using a multitemporal set of Sentinel-2 (S2) images. However, unless the temporal patterns are studied as a whole, some of the phenological information such as growth rates are lost, and delayed snow cover may spread events like growth onset over multiple dates between plots. This study aims to explore the added value of temporal fitting of Sentinel-2 time series (ts) over existing baseline models applied using the full time series as such. Our set of temporal features included functional components, harmonic decomposition, frequency decomposition, and phenological metrics. Out of the compared models, the Random Forest regression model using a set of fitted temporal features achieved the highest holdout prediction accuracy (R2 = 0.36, RMSE = 3.87, relative RMSE = 0.20) and cross-validation accuracy similar to the baseline models. However, all the compared regression models underestimated extreme plant diversity to some extent. Future studies should account for varying vegetation cover and terrain features by incorporating auxiliary data.

Woody plant encroachment (WPE) threatens grassland ecosystems across the Canadian Prairies, causing grassland biodiversity loss with substantial economic impacts due to reduced forage production. While remote sensing offers scalable monitoring capabilities, existing approaches lack frameworks for distinguishing shrub and tree encroachment and often require extensive ground truth data. This study developed an ensemble machine learning framework integrating Sentinel-1 SAR and Sentinel-2 optical imagery with UAV-derived training data to map fractional shrub and tree cover across Saskatchewan's Aspen Parkland and Moist Mixed Grassland ecoregions, SK. A stacking ensemble combining Random Forest, Support Vector Machine, XGBoost, and Artificial Neural Network models with Ridge regression meta-learning outperformed individual algorithms, achieving mean R² values of 0.65 for shrub and 0.68 for tree cover prediction. Multi-scale training incorporating features at 10, 30, 50, and 70 m resolution improved performance by 15% for shrub and 24% for tree compared to single-scale approaches. Feature importance analysis revealed that shrub detection relied primarily on red-edge bands and moisture indices, while tree detection depended heavily on SAR backscatters. Quantile histogram matching enabled successful model transfer from Foam Lake Community pasture to Aberdeen Community Pasture, with resulting maps indicating that total WPE exceeded 50% in both study areas, with shrubs occupying 23.7% (Foam Lake) and 18.5% (Aberdeen) of both regions at rates higher than 5% shrub cover. The present framework provides a scalable, cost-effective approach for operational woody encroachment monitoring, enabling early detection and targeted functional management interventions to preserve grassland ecosystems.

Fractional vegetation cover (FVC) is a key land surface parameter describing vegetation abundance and structure, defined as the fraction of the ground area occupied by vegetation when viewed from nadir. FVC provides essential insights into ecosystem condition, productivity, and disturbance, making it a critical variable for biodiversity monitoring and habitat assessment. However, generating accurate and repeatable FVC estimates remains challenging due to scale effects, spatial resolution constraints, and inconsistencies in available validation data across time and space. This research develops a machine learning (ML) framework for large-scale FVC estimation that addresses these challenges by combining multi-sensor Earth observation data and Active Learning (AL) model refinement techniques. The ML framework is applied within key wood bison habitat in northern Alberta, focusing on mapping six vegetation components: spruce, pine, deciduous, shrub, herbaceous, and moss. The approach integrates Sentinel-1, Sentinel-2, Landsat-9, and GLO-30 data, optimized through feature selection and ensemble-based Random Forest modeling. The resulting FVC maps achieved strong predictive performance (R² = 0.50–0.88) and capture fine-scale spatial variability in vegetation composition. The ML pipeline provides a scalable and adaptive framework for FVC estimation that supports provincial landcover updates, improves understanding of wood bison habitat features, and contributes to ongoing ecosystem monitoring and conservation planning across boreal Alberta.

Wildlife monitoring from aerial imagery often requires precise animal localization under practical constraints where only object counts are needed. Traditional detection methods rely on bounding-box annotations, introducing unnecessary cognitive load for small objects spanning only a few dozen pixels. This work introduces DINOKey, a modified DINO transformer-based detector adapted to operate natively on point annotations rather than bounding boxes. Key contributions include: (1) architectural modifications to the DINO decoder, detection head, and denoising queries to directly predict 2D keypoints; (2) a combined loss function integrating L1 regression, focal loss, and average Hausdorff distance, with ablations validating each component; (3) open-source implementation within an existing detection framework; and (4) demonstration of improved small-object localization and reduced false positives on an aerial elephant dataset compared to box-supervised baselines. Ablation studies show that the Hausdorff distance term provides the largest accuracy gain by effectively reducing false positives, while focal loss improves stability in densely clustered regions. The proposed method achieves 0.786 mAP and accurately localizes animal centers across diverse environmental conditions, offering a practical solution for conservation practitioners working under tight logistical constraints.

Fog-dependent ecosystems in the Atacama Desert host highly specialized vegetation, yet monitoring their functional traits remains challenging due to remoteness and limited spectral detectability. The bromeliad Tillandsia landbeckii exhibits extremely low reflectance in the VIS/NIR range, rendering conventional multispectral approaches ineffective.

This study evaluates the potential of a novel UAV-based VNIR/SWIR multi-camera system (camSWIR) for estimating canopy water content (CWC) in Tillandsia landbeckii. A UAV survey conducted in northern Chile acquired high-resolution (≈3 cm GSD) SWIR imagery across four operational bands (1100–1650 nm). Field-based destructive sampling (n = 20) provided reference CWC measurements, and a statistically rigorous workflow was applied to mitigate overfitting in a high-dimensional predictor space.

Results show that the spectral slope between 1200 and 1510 nm is the most informative predictor of CWC, with cross-validated performance indicating moderate predictive skill (LOOCV R² ≈ 0.52), but reduced stability under nested validation. The repeated selection of predictors within this wavelength region confirms a physically meaningful relationship with liquid water absorption.

Despite limitations due to a small sample size and species-specific optical properties, particularly the dense trichome layer that affects light interactions, the study demonstrates the feasibility of SWIR-based, non-destructive CWC estimation in hyper-arid ecosystems. These findings provide a proof of concept for future upscaling, highlighting the need for larger calibration datasets and improved modelling to enable reliable spatial mapping of plant water status.

Aerial imagery provides a powerful avenue for monitoring wildlife populations, yet automated detection remains challenging. Animals typically occupy only a tiny fraction of large-scale aerial imagery, may be partially obscured, and appear against highly diverse Arctic and sub-Arctic backgrounds. Suppervised deep-learning detectors also depend on large, fully annotated datasets, making broad ecological surveys labor-intensive and slow to scale.

This study explores an alternative perspective: viewing wildlife as rare events within mostly background imagery. Instead of training on annotated animal samples, an anomaly-detection framework learns the visual patterns of normal landscapes and identifies deviations from these patterns as potential animal locations. To guide the model without costly labels, simple animal-like shapes are inserted into background patches during training, encouraging the network to recognise features associated with real targets while avoiding the need for detailed masks or bounding boxes.

The approach generates two outputs: patch-level predictions distinguishing empty from potentially occupied areas, and pixel-level anomaly maps highlighting likely target locations. When evaluated on a highly varied Arctic dataset, the method remains reliable despite major shifts in terrain, surface texture, animal distributions and postures, and pronounced class imbalance that often degrade supervised models. Unlike distribution-based anomaly approaches that rely on stable normal-feature statistics and frequently misinterpret natural texture variability as anomalies, this method handles heterogeneous environments more effectively.

Overall, the study shows that anomaly-oriented frameworks, typically used in industrial and medical settings, have strong potential to ease annotation demands and support scalable, automated wildlife detection in complex remote-sensing environments.