EL PORTAL TREE CANOPY MAP — ANALYSIS METHODS Analysis: 3 October 2026. Description updated: 5 October 2026. PURPOSE AND LAYERS This map supports a community inventory of El Portal's urban forest. Two distinct layers represent continuous canopy cover and proposed individual crown shapes. The canopy layer supplies the coverage estimate. Crown records can be identified, reshaped, split or supplemented through resident observations. An outline may represent several trees, and overlapping outlines are not a reliable tree count. IMAGERY AND STUDY AREA Miami-Dade County MapCache/MDCImagery MapServer, layer 29: https://gisweb.miamidade.gov/arcgis/rest/services/MapCache/MDCImagery/MapServer Nominal ground sampling: 0.0762 metres per pixel (3 inches). The source comprises 55 georeferenced PNG exports. The service description refers to 2025 while source filenames refer to 2026; the acquisition date remains unconfirmed. The study area follows the county's municipal boundary: https://gisweb.miamidade.gov/ArcGIS/rest/services/MD_Communications/MapServer/0 Analysis includes 30 metres of surrounding context so boundary-crossing crowns can be retained in full. A crown must overlap the village to enter the inventory. Canopy coverage is measured only within the village. MODEL ROLES AND CHOICE Restor TCD (SegFormer B3) supplies a tree/non-tree canopy classification. Its role is to estimate cover across connected foliage without first separating every tree. This fits the coverage task and its 10 cm aerial-image input scale. Model documentation: https://huggingface.co/restor/tcd-segformer-mit-b3 Detectree2 (250312_flexi, Mask R-CNN R101 FPN) supplies individual crown masks. Its emphasis on crown delineation in dense forest makes it a relevant starting point for outlining trees in joined canopy. Suitability for the task is not a claim of measured accuracy in El Portal. Both models use pretrained weights; neither was retrained on a labeled El Portal dataset for this analysis. Project: https://github.com/PatBall1/detectree2 Model archive: https://zenodo.org/records/15863800 CANOPY PROCESSING Source imagery is resampled to 10 cm for model input. Predictions from 180 overlapping 150-metre windows are blended with reduced weight at their edges. The tree-class probability threshold is 0.50. A 3-by-3-cell morphological closing joins small gaps, and connected patches below about 43 sq ft are removed. The final mask has 25 cm cells. Coverage is the count of canopy cells inside the rasterized village divided by its total cell count: approximately 41.87%, or 111.54 acres, for this analysis. This is an experimental photo-derived estimate. Restor model revision: d1ab39591103c514c38fa6e80061f95d9c3a026e. CROWN PROCESSING Detectree2 receives 100-metre windows at 12.5 cm per pixel with a 70-metre stride (323 windows). Proposals require confidence of at least 0.25, an area of at least about 32 sq ft and at least 55% agreement with the canopy classification. Internal tile-edge fragments are rejected in favor of overlapping-window results. Proposals are considered in confidence order; one is omitted if more than 55% of its area is already covered by retained proposals. This reduces duplicates but does not guarantee that the remaining crowns are distinct trees. IDENTIFICATIONS AND FIELD OBSERVATIONS The models propose canopy and crown locations, not species, palm identities, health or trunk diameter. Additional photo-interpreted palm records are provisional. Resident identifications and drawn outlines form part of the map. Tree type determines color; only a selected species makes an active outline solid. A solid line records an identification, not independent expert validation. Removed trees retain their history and are shown with red hatching. DBH records trunk diameter for a single tree. Measurement year and measurer are recorded separately from the editor and edit time. A displayed trunk is a circle of the measured diameter at the estimated canopy centroid, visible only when large enough to resolve at the current zoom. This point is not a surveyed trunk location and may fall outside an irregular or multipart crown. INTERPRETATION AND LIMITS Visual checks cover residential streets, the school, joined canopy, Sherwood Forest and Little Farm. They are qualitative checks, not a local validation dataset or a measured precision/recall assessment. Dense foliage, deep shadow, palms, small trees and shrubs can cause omissions or incorrect outlines. Fine image pixels and processing grids do not establish positional accuracy. The coverage mask is independent of the editable crown inventory. A saved crown edit or removal does not recalculate the photo-derived coverage percentage. Species summaries describe identified records, not the species composition of all village trees. Crown areas and inferred diameters are plan-view estimates. GIS DATA AND REPRODUCIBILITY The map's Download GIS file action exports the current browser workspace as linked crown polygons and measured-trunk points, joined on tree_id. Use this export to include resident edits; linked base files do not include later browser changes. GeoJSON coordinates are WGS84 longitude/latitude (EPSG:4326). The crown_diameter_m attribute is sqrt(4 * area_m2 / pi), the diameter of an equal-area circle. It is not a field-measured crown spread. dbh_cm holds canonical trunk diameter; trunk_diameter_m supports actual-size GIS symbology. Reproducible local records in this folder: source/rescan.py — imagery preparation, canopy and crown inference. source/reconcile.py — integration of resident and photo-interpreted records. source/quality_review.py — visual comparison and geometry checks. data/canopy_summary.json and data/crown_summary.json — settings and results. data/source_provenance.json — source image extents and checksums. ../data/imagery_high_resolution/manifest.json — display imagery manifest. The separate SOURCES_AND_METHODS.txt retains implementation and revision notes.