"""Audit official CCAV V4 sample provenance and eligibility for local validation. No model accuracy is estimated: sample train/test membership is unavailable. """ import argparse import hashlib import json from pathlib import Path import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import numpy as np import pandas as pd from pyproj import Transformer from scipy.spatial import cKDTree BOUNDS = (117.404195622047, 23.9134768474132, 117.432930102047, 23.9374309174132) LABELS = ['S. alterniflora', 'Suaeda spp.', 'P. australis', 'T. chinensis', 'S. mariqueter', 'mangroves'] COLORS = ['#c44e52','#ddad39','#4c956c','#927653','#8172b3','#178a9b'] def main(): ap=argparse.ArgumentParser(description=__doc__) ap.add_argument('--source',type=Path,required=True) ap.add_argument('--listing',type=Path,required=True) ap.add_argument('--legacy',type=Path) ap.add_argument('--output',type=Path,required=True) args=ap.parse_args();args.output.mkdir(parents=True,exist_ok=True) payload=args.source.read_bytes(); md5=hashlib.md5(payload).hexdigest() records=json.loads(args.listing.read_text())['data'] official=next(x for x in records if x['fileName']==args.source.name) assert md5==official['md5'] and len(payload)==official['size'] d=pd.read_excel(args.source) assert d.columns.tolist()==['ID','Name','Time','Longitude','Latitude'] assert d[['Longitude','Latitude','Time','Name']].notna().all().all() d['class']=d.Name.str.strip() assert set(d['class'])==set(LABELS) assert d.ID.is_unique key=['Longitude','Latitude','Time','class'] unique=d.drop_duplicates(key).copy() unique['source_multiplicity']=unique.set_index(key).index.map(d.groupby(key).size()) x0,y0,x1,y1=BOUNDS d['inside_image']=d.Longitude.between(x0,x1)&d.Latitude.between(y0,y1) unique['inside_image']=unique.Longitude.between(x0,x1)&unique.Latitude.between(y0,y1) groups=d.groupby(['Longitude','Latitude','Time'])['class'].nunique() counts=pd.DataFrame({'raw_rows':d.groupby('class').size(), 'unique_coordinate_year_class':unique.groupby('class').size()}).reindex(LABELS) counts['excess_duplicate_rows']=counts.raw_rows-counts.unique_coordinate_year_class counts.to_csv(args.output/'sample_class_counts.csv') d.groupby(['Time','class']).size().rename('rows').to_csv(args.output/'sample_year_class_counts.csv') local=unique.loc[unique.inside_image].copy() transform=Transformer.from_crs(4326,32650,always_xy=True) if len(local)>1: xy=np.column_stack(transform.transform(local.Longitude.to_numpy(),local.Latitude.to_numpy())) nn=cKDTree(xy).query(xy,k=2)[0][:,1] local['nearest_distinct_record_m']=nn local.to_csv(args.output/'local_unique_samples.csv',index=False) unique.to_csv(args.output/'samples_unique_coordinate_year_class.csv',index=False) report={'dataset_doi':'10.57760/sciencedb.31077','version':'V4','license':'CC-BY-4.0', 'attribution':'Yuying Li et al., CCAV-10m (Science Data Bank, V4, 2026)', 'provider_md5_verified':True,'md5':md5,'sha256':hashlib.sha256(payload).hexdigest(), 'bytes':len(payload),'raw_rows':len(d),'unique_coordinate_year_class':len(unique), 'excess_duplicate_rows':len(d)-len(unique),'unique_coordinates':len(d.drop_duplicates(['Longitude','Latitude'])), 'conflicting_class_coordinate_year_groups':int((groups>1).sum()), 'image_bounds_wgs84':BOUNDS,'local_raw_rows':int(d.inside_image.sum()), 'local_unique_coordinate_year_class':len(local), 'local_unique_coordinates':len(local.drop_duplicates(['Longitude','Latitude'])), 'local_classes_raw':d.loc[d.inside_image,'class'].value_counts().to_dict(), 'local_classes_unique':local['class'].value_counts().to_dict(), 'local_nearest_distinct_record_m_median':float(np.median(nn)) if len(local)>1 else None, 'legacy_byte_identical':args.legacy.read_bytes()==payload if args.legacy else None, 'limitations':['Coordinates have no explicit CRS field; WGS84 assumed for overlay.', 'Exact duplicate records collapsed for audit, not proof of independent plots.', 'Year-repeat coordinates remain dependent; nearest-distance is diagnostic.', 'Published 15558 total includes nonpublic samples; public sheet has no split membership.', 'No independent accuracy metric can be computed from these records.']} (args.output/'sample_audit.json').write_text(json.dumps(report,indent=2)+'\n') plt.rcParams.update({'font.size':9,'axes.spines.top':False,'axes.spines.right':False,'svg.fonttype':'none'}) fig,axs=plt.subplots(1,3,figsize=(12,4),layout='constrained') for lab,col in zip(LABELS,COLORS): sel=unique[unique['class']==lab] axs[0].scatter(sel.Longitude,sel.Latitude,s=3,c=col,label=lab,alpha=.55,rasterized=True) axs[0].scatter([(x0+x1)/2],[(y0+y1)/2],marker='*',s=90,facecolor='none',edgecolor='black') axs[0].set(xlabel='Longitude (°E)',ylabel='Latitude (°N)',title='a Public sample coordinates') axs[0].legend(fontsize=6,markerscale=2,frameon=False) ypos=np.arange(6) # Counts shown explicitly because most exact duplicate differences are small. axs[1].barh(ypos,counts.unique_coordinate_year_class,color=COLORS,label='Unique coordinate–year–class') axs[1].set(yticks=ypos,yticklabels=LABELS,xlabel='Records',title='b Exact duplication audit') axs[1].set_xlim(0,3800) for i,row in enumerate(counts.itertuples()): axs[1].text(row.unique_coordinate_year_class+35,i,f'{row.unique_coordinate_year_class} (−{row.excess_duplicate_rows})',va='center',fontsize=7) axs[1].set_xlabel('Unique records (excess duplicates removed)') axs[2].scatter(local.Longitude,local.Latitude,s=25,c=COLORS[-1],alpha=.7) axs[2].set(xlim=(x0,x1),ylim=(y0,y1),xlabel='Longitude (°E)',ylabel='Latitude (°N)',title='c Image footprint: mangrove labels only') axs[2].ticklabel_format(useOffset=False,axis='both');axs[2].tick_params(axis='x',rotation=30) axs[2].text(.03,.97,f'{len(local)} coordinate–year records\n{int(d.inside_image.sum())} original rows\n0 Spartina labels',transform=axs[2].transAxes,va='top',fontsize=8) for ext in ['png','pdf','svg']:fig.savefig(args.output/f'ccav_sample_audit.{ext}',dpi=220) print(json.dumps(report,indent=2)) if __name__=='__main__':main()