Species-Level Salt Marsh Intelligence
Automated UAV vegetation mapping powered by deep learning and AWS. Designed for ecologists requiring high-fidelity geographic and statistical precision.
Scientific Precision at Scale
Engineered for ecologists, our platform translates complex deep learning outputs into actionable, verifiable environmental data.
Species-Level Segmentation
Our two-stage pipeline classifies vegetation down to the species level with high accuracy, distinguishing morphologically similar halophytic grasses across heterogeneous salt marsh canopies.
Quantified Confidence
ecoVision 2.0 applies conformal prediction via MAPIE to every dominance estimate, producing statistically guaranteed coverage intervals rather than uncalibrated softmax scores. Researchers receive not just a prediction, but a defensible uncertainty bound, critical for ecological reporting and regulatory use cases.
Cloud-Native Scalability
Process gigabyte-scale drone imagery in minutes, not hours. Built on AWS infrastructure, the platform dynamically scales GPU instances to handle massive parallel inference tasks without bottlenecking your research.
1. High-Res Ingestion
UAV imagery resized to 512×512 px with minimal normalisation to preserve ecologically relevant spectral variation.
2. Sequential Inference
SegFormer-B5 segments vegetation at the pixel level; connected component analysis extracts discrete vegetation patches; ConvNeXt-Base classifies each patch to species level on AWS g4dn instances.
3. Confidence Filtering
Only classifications exceeding a defined confidence threshold are forwarded, low-confidence predictions are withheld from downstream aggregation.
4. Geo-Spatial Assembly
Classified patches are assigned to 2×2 m grid cells; species dominance is computed as proportional areal coverage and exported as georeferenced maps and colour-scaled dominance heatmaps.
Advanced Model Architecture
ecoVision's pipeline treats vegetation mapping as a hierarchical perception problem, moving progressively from pixel-level spatial structure to object-level species identity to ecologically interpretable dominance scores.
At the segmentation stage, a fine-tuned SegFormer-B5 model, built on the hierarchical Mix Transformer (MiT-B5) encoder with a lightweight MLP decoder and passed to a fine-tuned ConvNeXt-Base classifier, which assigns a species-level label with a confidence score. Only classifications exceeding a defined confidence threshold proceed further. This combination excels at delineating boundary regions where vegetation classes intermix, a common challenge in dynamic salt marsh environments.
Read the Technical WhitepaperDeveloped in Partnership With
