"""EMSR838 overview figure from figs/mosaic_emsr838.npz (made on the server by make_figures.py) + results/pop_*.csv. python3 pakflood/plot_overview.py -> figs/fig_emsr838_overview.png Left: where the activation's AOIs are (official flood outline over WorldPop density). Right: zoom on the two most telling AOIs, official outline vs hold-out vs few-shot vs full fine-tune, with the affected population per AOI. """ from pathlib import Path import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import numpy as np import pandas as pd import scipy.ndimage as ndi from matplotlib.patches import Patch, Rectangle HERE = Path(__file__).resolve().parent z = np.load(HERE / "figs" / "mosaic_emsr838.npz", allow_pickle=True) W, S, E, N = z["bounds"] res = float(z["res"]) mapped, people = z["mapped"], z["people"] H, Wd = mapped.shape layers = {"label": z["label"], "sa_holdout": z["sa_holdout"], "sa_fewshot": z["sa_fewshot"], "full_finetune": z["full_finetune"]} NAMES = {"label": "CEMS official outline", "sa_holdout": "LoRA hold-out\n(never saw South Asia)", "sa_fewshot": "LoRA few-shot\n(+26 SA tiles)", "full_finetune": "Full fine-tune\n(trained on PK)"} pops = {} for k, f in [("label", "pop_sa_holdout"), ("sa_holdout", "pop_sa_holdout"), ("sa_fewshot", "pop_sa_fewshot"), ("full_finetune", "pop_full_finetune")]: d = pd.read_csv(HERE / "results" / f"{f}.csv").groupby("aoi")[["pop_pred", "pop_label"]].sum() pops[k] = d.pop_label if k == "label" else d.pop_pred INK, INK2 = "#0b0b0b", "#52514e" plt.rcParams.update({"font.size": 9, "axes.titlesize": 9, "figure.facecolor": "#fcfcfb", "axes.facecolor": "#fcfcfb", "savefig.facecolor": "#fcfcfb"}) dens = np.log10(1 + people) vmax = np.percentile(dens[dens > 0], 99.5) # AOI bounding boxes from connected components of the mapped mask around each AOI centroid comp, n = ndi.label(mapped > 0.05) boxes = {} for aoi, (x, y) in zip(z["aois"], z["aoi_xy"]): c, r = int((x - W) / res), int((N - y) / res) cid = comp[r, c] if cid == 0: # centroid not on mapped pixel: nearest mapped pixel rr, cc = np.nonzero(mapped > 0.05) i = np.argmin((rr - r) ** 2 + (cc - c) ** 2) cid = comp[rr[i], cc[i]] rows, cols = np.nonzero(comp == cid) boxes[str(aoi)] = (W + cols.min() * res, N - (rows.max() + 1) * res, W + (cols.max() + 1) * res, N - rows.min() * res) def draw(ax, key, ext=None, alpha_map=0.25): w, s, e, n = ext or (W, S, E, N) c0, c1 = int((w - W) / res), int(np.ceil((e - W) / res)) r0, r1 = int((N - n) / res), int(np.ceil((N - s) / res)) sl = (slice(r0, r1), slice(c0, c1)) ex = (W + c0 * res, W + c1 * res, N - r1 * res, N - r0 * res) ax.imshow(dens[sl], cmap="Greys", vmin=0, vmax=vmax, extent=ex, interpolation="nearest") rgba = np.zeros((r1 - r0, c1 - c0, 4), np.float32) rgba[..., :3] = (0.12, 0.47, 0.71) rgba[..., 3] = alpha_map * (mapped[sl] > 0.05) ax.imshow(rgba, extent=ex, interpolation="nearest") rgba = np.zeros((r1 - r0, c1 - c0, 4), np.float32) rgba[..., :3] = (0.84, 0.15, 0.16) rgba[..., 3] = np.clip(layers[key][sl] * 1.4, 0, 1) ax.imshow(rgba, extent=ex, interpolation="nearest") ax.set_xlim(ex[0], ex[1]) ax.set_ylim(ex[2], ex[3]) fig = plt.figure(figsize=(15, 6.4)) gs = fig.add_gridspec(2, 5, width_ratios=[1.35, 1, 1, 1, 1], wspace=0.12, hspace=0.3, left=0.04, right=0.99, top=0.86, bottom=0.1) ax = fig.add_subplot(gs[:, 0]) draw(ax, "label", alpha_map=0.35) offs = {"EMSR838-10": (-0.25, 0.05), "EMSR838-11": (0.2, 0.05), "EMSR838-8": (-0.3, 0.0), "EMSR838-1": (0.35, 0.05)} for aoi, (x, y) in zip(z["aois"], z["aoi_xy"]): dx, dy = offs.get(str(aoi), (0.25, 0.0)) ax.annotate(str(aoi).replace("EMSR838-", "AOI "), (x, y), (x + dx, y + dy), fontsize=7.5, color=INK, ha="center", arrowprops=dict(arrowstyle="-", color=INK2, lw=0.6)) for key in ["EMSR838-1", "EMSR838-7"]: w, s, e, n = boxes[key] ax.add_patch(Rectangle((w - 0.03, s - 0.03), e - w + 0.06, n - s + 0.06, fill=False, edgecolor=INK, lw=1.0, ls="--")) ax.set_xlabel("longitude") ax.set_ylabel("latitude") ax.set_title("EMSR838 activation: 9 mapped areas in Punjab\nofficial flood outline (red) over WorldPop density", fontsize=9) keys = ["label", "sa_holdout", "sa_fewshot", "full_finetune"] GROUPS = [("EMSR838-1", ["EMSR838-1", "EMSR838-8"], "AOI 1 + 8 (Chenab)"), ("EMSR838-7", ["EMSR838-3", "EMSR838-7"], "AOI 3 + 7 (Sutlej)")] for row, (aoi, members, rowname) in enumerate(GROUPS): w, s, e, n = boxes[aoi] pad = 0.03 for col, key in enumerate(keys): ax = fig.add_subplot(gs[row, col + 1]) draw(ax, key, (w - pad, s - pad, e + pad, n + pad)) ax.set_xticks([]); ax.set_yticks([]) p = sum(pops[key][m] for m in members) off = sum(pops["label"][m] for m in members) t = f"{p:,.0f} people" if key == "label" else f"{p:,.0f} people ({100 * (p / off - 1):+.0f}%)" ax.set_title((NAMES[key] + "\n" if row == 0 else "") + t, fontsize=8.5) if col == 0: ax.set_ylabel(rowname, fontsize=9.5, fontweight="bold") fig.legend(handles=[Patch(color="#d62728", label="flooded (fraction of ~200 m cell)"), Patch(color="#1f77b4", alpha=0.3, label="mapped area (AOI)"), Patch(color="grey", label="WorldPop 2020 population density (log)")], loc="lower center", ncol=3, frameon=False, fontsize=9, bbox_to_anchor=(0.6, 0.0)) fig.suptitle("Pakistan floods, August 2025 (EMSR838, Sentinel-1 at 10 m): flood extent and affected population per model", fontsize=11, fontweight="bold") fig.savefig(HERE / "figs" / "fig_emsr838_overview.png", dpi=170) print("wrote figs/fig_emsr838_overview.png; boxes:", {k: tuple(round(v, 2) for v in b) for k, b in boxes.items()})