Method

From a threshold to a tree

Let \(x\) be a standardized index on the time–latitude–longitude grid (optionally smoothed with a Gaussian of width \(\sigma\)) and \(d=-x\) the deficit. Labelling the connected cells with \(d \ge \lambda\) gives a drought catalogue for one threshold \(\lambda\). The max-tree stores the connected components of every upper level set \(\{d \ge \lambda\}\) at once (26-connectivity in space and time); a fixed-threshold catalogue is one horizontal cut through it.

_images/concept_tree.png

A node \(n\) is significant when its area-weighted space–time volume \(V(n) \ge v_{\min}\) and its level \(\lambda(n) \ge \lambda_{\mathrm{floor}}\). Chains of significant nodes are condensed into clusters, which branch where two or more significant children appear.

Event selection

Each cluster \(c\) has a peak deficit \(p(c)\) and a saddle \(s(c)\), the level at which it joins its siblings. Its topological persistence is

\[P(c) = p(c) - s(c).\]

A maximum created by noise rises little above the saddle that connects it to a larger structure; two droughts are separated by a deep saddle. At every split the child with the highest peak continues the parent, and every other child whose persistence is below a threshold \(\pi\) is absorbed with its subtree. The threshold adapts to the noise of the field (persistence="auto:k"):

\[\pi = k \cdot 1.4826\,\mathrm{MAD}\left[x - G_2(x)\right],\]

where \(G_2\) is a Gaussian filter of width 2. Volume weighting (pers_gamma) lets large branches survive with a smaller contrast,

\[P_V(c) = P(c)\left(\frac{V(c)}{v_{\min}}\right)^{\gamma},\]

and lineage inheritance (inherit) gives event \(j\) the identity of an older event \(i\) when at least a share \(q\) of its first footprint lay inside \(i\) in the previous month,

\[\frac{\left|\{L_{t_j-1}=i\} \cap \{L_{t_j}=j\}\right|}{\left|\{L_{t_j}=j\}\right|} \ge q .\]

Episodes and pulses

Because all events come from one tree, DRIFT returns two nested levels: episodes (selection="persistence_vl", the leaves of the simplified tree) and pulses (selection="leaf", the leaves of the condensed tree before simplification). Each pulse lies inside one episode. footprint="birth" assigns each event its component at the level where it separates from its siblings; footprint="watershed" grows the events into all cells with deficit at least ext_floor, sharing contested cells along deficit saddles.

Choosing the floor

driftdrought.select_floor() scans candidate floors and picks the one where the number of events changes least (event-count stability), excluding floors with fewer than five events or an event longer than a third of the record:

\[\lambda^{\ast}_{\mathrm{floor}} = \arg\min_{\lambda} \frac{1}{2}\left(\frac{|N(\lambda+\Delta)-N(\lambda)|}{N(\lambda)} + \frac{|N(\lambda)-N(\lambda-\Delta)|}{N(\lambda-\Delta)}\right).\]

Event confidence and significance

The confidence of an event is its mean best overlap with the events of \(K\) alternative catalogues (perturbed parameters, or another input product):

\[\mathrm{conf}(e) = \frac{1}{K}\sum_{k=1}^{K} \max_{f \in L_k} \mathrm{IoU}(e, f).\]

The significance compares the volume-weighted deficit contrast of each event, \(S_e = (\max d_e - \lambda_{\mathrm{floor}})(V_e/v_{\min})^{\gamma}\), with the events that the same configuration finds in phase-randomized surrogates of the field, which keep its space–time spectrum and marginal distribution. The test is conservative.

Limitations

  • At the smoothing that maximizes tracking accuracy, the two lobes of a splitting drought remain connected; DRIFT keeps one identity for them but does not report the split. At monthly resolution no tested description separates a real split from noise-induced fragmentation.

  • Weak events in strongly correlated noise are detected poorly by every method.

  • Pulses depend on the input product; the product confidence identifies the events that are robust to it.

  • The floor and the minimum volume remain parameters; tune them for a new region with the benchmark generator (Benchmark).