Dakota Agros | ML Training & Annotation Portal

Imagery Preprocessing

Select Orthomosaic Map

Agisoft 6-band format detected (includes transparency/alpha band). It will degrade ML model accuracy.

Visual Optimization

Preview Resolution

Band Composition

Brightness (Gamma) 1.0

Contrast (Percentile Stretch) 1.0

Crop Canopy Analyzer

Analysis Mode

ISODATA Clustering Parameters

Number of Classes

Processing Res

Skip Factor (Sample Interval) 5

1 (every px, slow)10 (ArcGIS default)30 (fast, coarse)

Min Class Size (pixels) 5

1 (keep all)20 (ArcGIS default)200 (large only)

Max Iterations 20

120 (ArcGIS default)100

Merge Threshold 0.50

0 (no merging)0.5 (default)2.0 (aggressive)

Plant Canopy Pre-Filter (Extract Plants First)

Vegetation Threshold (NDVI / ExG) 0.35

0.10 (loose)0.35 (recommended)0.75 (dense only)

Masks out soil, ruts & shadows so 100% of clusters isolate plant leaves and yellow blooms.

3D Elevation & Point Cloud Source

Select any .laz point cloud or DEM GeoTIFF.

Min Plant Height Threshold (CHM) 0.40m

0.15m (early bud)0.40m (mid-season)1.50m+ (mature)

Clusters 3D Canopy Height (CHM) and Apical Prominence to isolate plant tops for stand counting.

Canopy Peak Extractor Post-Processing

Isolates search to NDVI plant canopy. Treats the flattest part as the local canopy floor to normalize field slope, then extracts all prominent peaks as true apical tops. Min Peak Prominence Above Canopy +0.07m

+0.02m (early bud)+0.07m (closed canopy)+0.15m (isolated dome)

Local Canopy Floor Window (Smoothing) 0.45m

0.20m (tight)0.45m (sunflower crown)0.80m (wide)

Min Plant Spacing (Merge Double Counts) 0.35m

0.15m (tight)0.35m (row spacing)0.60m+ (mature)

Tagging Shape Type:

Band Standardization Spatial Denoiser (Disable to keep small blooms)

Feature Selection RGB Channels (Color) NDVI (Veg Density) ExG (Excess Green) Yellowness Index (Blooms) NDRE (Red Edge Veg)

Manual Digitization

Target Crop

Attribute Type

Tip: Right-click a shape on the map to delete/remove it.

Shapes: 0

ISO object extraction limits

Applied when you click Use for Chips. Include small components to inspect them with Select by attributes.

Minimum component area (source pixels) Maximum component area (source pixels)

Shapes File Manager

Shapefile Export Name

De-noise deleted shapes (penalize if incorrectly classified)

Load Existing Shapefile

ML Model Training

RTX 4090 Ready

Target Model Pipeline 1-Class Model

3D Height & Point Cloud Source Active

Dense LAZ Point Clouds (D:\19AgisoftExports\PAG2020\laz) are rasterized pixel-for-pixel to orthomosaic grid with DTM ground subtraction for CHM (0–2.5m) and TPI (±0.3m apical convexity dome).

Multi-Layer Input Tensor Breakdown

0: Red 1: Green 2: Blue 3: RedEdge 4: NIR 5: NDVI 6: NDYI 7: ExG 8: CHM (LAZ Height) 9: TPI (Apexes)

Multi-Temporal ISO Ground Truth Scanning datasets...

20_805
LAZ+GT

21_718
LAZ+GT

21_730
LAZ+GT

21_805
LAZ+GT

21_811
LAZ+GT

Epochs

Patience

Batch Size

Resume from Checkpoint (if exists) Backbone: YOLOv8l-seg

Select Training Shapefiles Select All

Training Job Type

Base Weights to Fine-Tune

Learning Rate

Frozen Layers

Use Merge Strategy (Prevent Forgetting)

Capacity

Image Size

Epochs

Spacing

Model Name

Real-Time Pipeline Logs

Console idle...