The science under Understory

Pulling a native plant by mistake is the costliest error a volunteer can make, and American bittersweet, staghorn sumac and black walnut all have invasive look-alikes. So Understory’s classifier gives one of four decisions. PULL means a core invasive target, DON’T PULL means a native or look-alike is most likely, and OTHER INVASIVE means a secondary invasive such as barberry or buckthorn. Otherwise it says INSPECT and asks for the view that settles it, such as where the fruit grows on a bittersweet vine.

A BioCLIP 2 foundation model, adapted with LoRA, is the teacher. It is distilled into BioCLIP’s much smaller ViT-B/16, which runs offline on an iPhone as a 57.9 MB Core ML model.

The hardest case is highbush blueberry with red leaves in October, which looks like burning bush. Patch extent is measured without LiDAR: ARKit visual-inertial odometry plus Depth Anything V2 monocular depth, scale-aligned to ARKit’s feature points.

Each surveyed patch spreads in a stage-structured stochastic integrodifference model: plants grow in place, then seeds, fruit and rhizome fragments travel by each species’ dispersal kernel, with rare long jumps, and settle where the habitat allows. A Monte Carlo ensemble turns that into a probability of invasion for every 2 m cell, year by year, and into the area that clearing each patch would prevent. Every one of the model’s numbers cites its source in the literature.

The simulator is C++. The API runs it natively; the map runs the same code compiled to WebAssembly, so the impact of skipping a patch is simulated in your browser.

Understory’s integrodifference model fits in one line. Next year’s density of plants $n_{t+1}$ at a point $x$ adds up what this year’s plants $f(n_t)$ produce at every point $y$, carried to $x$ by the dispersal kernel $k$, and keeps what the habitat $H(x)$ lets settle there.

A workday is a budget of volunteer-minutes. Each surveyed patch becomes a task worth the spread that clearing it prevents over five years, estimated by the simulator, and costing the walk in and out plus the person-minutes of work, matched to the skill it needs. Bittersweet is the slow one, done in three parts as on a real workday: the brush around the vines is cleared, each vine cut free and its roots pulled, and the cut vines piled beside the patch; at the end of the day everyone carries the piles to the trailhead; then they load the dumpster.

Volunteers work in groups of two to five, as real workdays do, never alone. Timefold Solver tries every way of splitting the crew, and for each it routes the groups along the property’s trails, solving a team orienteering problem with time windows in which a group shares the work at each patch. After the most prevented spread, it minimises everyone’s time on site, so groups finish together and, when the work worth doing runs out, the day ends early instead of leaving people standing around. The API simulates the chosen patches together, so overlapping spread isn’t counted twice, and compares the plan with three rules of thumb for the same groups, replayed exactly: largest patch first, nearest patch first, and random order.

The plan is the set of patches $S$ that prevents the most spread. Here $A_5(S)$ is the area invaded in year 5 when the patches in $S$ are cleared now, and $\mathcal{R}(B)$ holds every set the crew can walk to and clear within its minutes $B$. Today’s planner adds up single-patch values $F(\{p\})$ instead.