Results

I split my findings into two parts. First, real-world evidence from public data that I can stand behind right now. Second, the analysis framework I built and validated for the field data I still have to collect.

Part 1 · Real-world evidence real data

Where the species actually occur

I pulled 1,402 georeferenced records of my four species from GBIF across northwest and central Connecticut, from the Litchfield hills east to the Farmington Valley. The records cluster around towns, roads, and trails.

real data Interactive map of 1,402 GBIF records by species (points cluster when zoomed out). Click any point for species, year, and source; toggle the density heatmap or switch to satellite. Open the full map (all layers) →

Real clusters (Getis-Ord Gi*)

When I grid the records into 1 km cells and run Getis-Ord Gi*, each species shows statistically significant clusters. Japanese knotweed forms a clear hotspot along the developed Torrington corridor. I also include the all-species map as an honest control for where people tend to look.

real data Interactive Getis-Ord Gi* hotspots (red cells) with a monitoring-priority layer. Turn the species points on or switch the basemap. Hover a cell for its record count. Full map →

Local patterns

Visible reporting clusters appear in several developed-edge landscapes, including parts of the Farmington Valley, Greater Hartford, and northwest Connecticut. These clusters should be interpreted carefully because public records can reflect where people observe and report plants.

Still, the patterns help identify useful places for local monitoring, education, and field validation.

Most invaders sit on disturbed land more than chance

This is my central real-data result. I sampled CT ECO 2015 land cover at each occurrence and at random background points. The land itself is only 29.8% developed/disturbed, and three of the four species turn up there far more often. Japanese barberry is the exception, for a reason I return to below:

Species	Records	On disturbed land	Selection ratio	p-value
Japanese knotweed	196	58.7%	1.97×	2 × 10⁻¹¹
Oriental bittersweet	209	56.9%	1.91×	1 × 10⁻¹⁰
Garlic mustard	523	51.6%	1.74×	4 × 10⁻¹¹
Japanese barberry	269	30.5%	1.02×	0.91 (n.s.)

Land cover disturbance test — real data Left: percent of occurrences on developed/disturbed land vs. the background availability (dashed line); *** = p < 0.001. Right: full land-cover composition by species and background. Knotweed, bittersweet, and garlic mustard far exceed the background; barberry sits right at it.

What I conclude from the real data. Independent of my own field survey, three of the four invaders are strongly over-represented on human-disturbed land. Observer bias remains a limitation of public occurrence data, but the consistency of the land-cover and distance-to-development patterns suggests the disturbance association is unlikely to be only a reporting artifact. The fourth, barberry, needs a different test, which comes next.

How close to development? real data

Barberry looked harmless in that first test only because it is a forest-understory shrub: at any given point it usually sits under tree canopy, not on pavement, so a land-cover-at-the-point test cannot see its link to disturbance. A cross-check with 10 m Google Dynamic World (2025 to 2026) said the same, with knotweed, bittersweet, and garlic mustard well above the landscape average (2.5×, 1.9×, 1.9×) and barberry not.

So I asked a sharper question: not whether a plant is on developed land, but how close it is. I measured each occurrence's distance to the nearest built pixel and compared it to random background points. Now all four species, barberry included, are significantly closer to development than random:

Species	Median distance to built	Within 100 m	p (closer than random)
Japanese knotweed	54 m	62%	4 × 10⁻²⁶
Garlic mustard	80 m	55%	9 × 10⁻³⁸
Oriental bittersweet	125 m	46%	2 × 10⁻¹⁵
Japanese barberry	198 m	32%	6 × 10⁻⁸
Background (random)	385 m	21%	n/a

Distance to built land by species — real data Distance to the nearest built land (Dynamic World 2025-26). Left: every species sits far closer to development than random (*** = p < 0.001). Right: cumulative share within a given distance, where all four species curves rise above the background.

The conclusion. All four invaders are edge- and disturbance-associated. What differs is how far each reaches into intact habitat: knotweed hugs roadsides and fill (median 54 m from development), garlic mustard follows trails and edges (80 m), Oriental bittersweet reaches deeper as a canopy vine (125 m), and Japanese barberry penetrates furthest of all into the forest understory (198 m). The lesson: for understory and canopy invaders, distance to disturbance is a more sensitive test than land cover at a single point.

Predicting relative invasion risk real data

Finally, I pulled the pieces together into a model. I trained a random forest to tell invasive-occurrence locations apart from the average landscape, using six landscape features, and had it map relative invasion risk across the region. Scored honestly with 5-fold cross-validation, the pooled model reaches an AUC of 0.79 (0.60 to 0.70 per species), so it discriminates reasonably well.

Predicted relative invasion risk map with record clusters — real data Predicted relative invasion risk across NW and central Connecticut, with the real record clusters overlaid. Risk runs highest along the developed eastern edge (the Farmington Valley and Greater Hartford lowlands) and lowest in the forested interior, and the hot eastern zone lines up with the biggest clusters.

The model also ranks which landscape features matter most:

Landscape feature	Importance (pooled)
Elevation	0.27
Distance to development	0.17
Distance to water	0.15
Slope	0.14
Tree cover nearby	0.14
Developed land nearby	0.13

Variable importance by species — real data How predictive each landscape feature is, pooled and per species. Longer bars mean more predictive.

What drives the risk. Relative invasion risk is highest in low, developed valley landscapes near roads and water, and lowest in the high, intact forest interior. Elevation ranks first, but mostly as a stand-in for that valley-developed versus upland-forest split. Distance to development is the direct signal, with distance to water close behind, and no single feature dominates.

How to read this map. Predictions represent relative invasion potential, not confirmed future occurrences: higher predicted risk reflects greater landscape suitability, not certainty that invasion will occur. The model uses a random background, so it partly reflects where people record plants. When I instead compared invasive locations to where people record plants in general, the landscape signal nearly vanished, so a good part of the pattern is observation effort. I treat it as a monitoring-priority screening surface, association not cause.

A caution about where we look. There really are more invasives along edges, and my data support that. But part of the pattern is simply that edges and disturbed places are where people can get to, so we go looking there and we tend to find what we are looking for in the places that are easiest to search. This means the places with the most recorded invasives are not automatically the places where invasion matters most, and some high-stakes spots may be the ones we rarely visit. This is why I treat the map as a guide for where to look, not a ranking of where the problem is worst.

Why it matters. Because early detection is substantially less costly than managing established infestations, a screening surface like this can help conservation organizations and land managers allocate limited monitoring resources more efficiently.

Field check: does the map hold up on the ground? real data

Everything above uses public data. To see whether it holds up, I went into the field and recorded 64 of my own occurrences, each with a photo and a GPS point (collected through two of my own submission forms). This is a real ground-truth of the model, focused on the Farmington Valley.

Field-confirmed occurrences by species — real data My 64 field-confirmed occurrences, colored by species, around Avon and Simsbury. Each one is a plant I found and photographed on the ground.

Field observations on a topographic basemap showing roads, water, and parks — real data The same points on a topographic basemap. My detections line up with the developed Route 10 (Hopmeadow Street) corridor and the Farmington River, while the forested Talcott Mountain and Metacomet ridge to the east, including Talcott Mountain State Park, stay nearly empty. This is the disturbed-edge pattern on the ground, next to roads, rivers, and recreation land.

Then the key question: does the model predict higher invasion risk where I actually found invasives? It does, clearly.

Predicted risk at field points vs random background — real data The model predicts a median relative risk of 0.56 at my field points versus 0.20 at random locations, and separates the two well (validation AUC = 0.95).

What this shows. On an independent set of 64 real, ground-confirmed points that the model never saw, it assigns much higher risk where invasives actually are (AUC 0.95). My field points also sit much closer to developed land than random (median 176 m vs 391 m, p < 0.001), matching the regional pattern. This is genuine validation of the map, not just an internal check.

This validates the model. The other half of my plan, a belt-transect test of how cover falls with distance from an edge, needs quadrat measurements I have not entered yet, so the figures in the next section are still a simulated framework.

Part 2 · Analysis framework illustrative, simulated

These figures use simulated data. My field validation above uses real, ground-collected data, but my belt-transect quadrat measurements (cover class at set distances from an edge) are not entered yet. So the edge-gradient figures here are built on realistic simulated data to prove the analysis runs end to end. When I enter the real quadrat data, the same scripts produce the real figures with no other changes.

The edge gradient is nonlinear and species-specific

GAM nonlinear decline — illustrative My generalized additive model recovers a nonlinear decline in cover with distance from the edge, with the shape differing by species.

Different species cluster in different places

Per-species hotspots at seven sites — illustrative Across all seven field sites, each species' hotspots (colored by species) fall in different places, so the invaders rarely cluster in the same spots (garlic mustard vs. knotweed Jaccard ≈ 0.04). This points to species-specific management.

From hotspots to a priority map

Priority zones — illustrative Crossing invasive hotspots with conservation value yields four management zones: Remove/Treat, Restore native, Monitor, and Low priority.

Conservation priorities

Putting both parts together, and taking into account expert feedback on this project, my recommendations for the study area are:

Watch the edges and act early. Roads, trails, and field margins are where new patches turn up first, so early detection there is cheap and effective.
Make protecting intact forest the priority. The interior, like Great Mountain Forest, shows lower invasion. Keeping it that way by stopping spread from the edges is often more valuable than fighting the biggest existing stands.
The worst infestations are not always the best places to treat. Where invasives have already outcompeted the natives, removing them, even with replanting, often just lets similar or worse invaders move back in, unless there are resources for several years of repeat treatment. In those places effort usually goes further on preventing spread into cleaner ground and on easing the conditions (disturbance, canopy gaps, altered soil) that let invasives win in the first place.
Target control species by species. Because the invaders occupy different parts of the landscape, a single "invaded zone" approach would miss most of the problem.

Why prevention beats cleanup. Invasives take over where the conditions let them outcompete natives, so pulling them from a badly disturbed site can just open the door for the next one. Replanting rarely works unless you also fix what caused the loss in the first place.

See my data sources & tools →