10 data categories & what to collect per upazila
1Geographic & Administrative Identity
Upazila name, district, division; GPS boundary coordinates; area (km²); number of unions, mouzas, villages; population (BBS census); population density; literacy rate; nearest major urban centre and distance. This data anchors every other section — confirm from DLRS mouza maps or CEGIS, not assumptions.
SpatialQuantitative
2Geology, Geomorphology & Soils
Physiographic zone (Tista Fan, Old Brahmaputra Floodplain, Sylhet Basin, Meghna Floodplain, etc.); flood plain type (active/inactive/high); dominant soil series from SRDI survey; soil texture, pH, organic matter content; land type classification on the F0–F3 flood inundation scale; elevation and slope from SRTM DEM data. Name specific soil series where known — do not generalise.
SpatialSemi-quantitative
3Land Use Classification (Current)
Zone-wise area breakdown: agricultural (irrigated/rain-fed), homestead/settlement, water bodies, wetlands, forest/plantation, social infrastructure, industrial, char land, cultural heritage zones. Use provided zoning data or RS imagery (Sentinel-2, Landsat 9). Calculate percentage of total area per zone. Critically: comment on what is absent, surprisingly small, or disproportionately large relative to the upazila's known environmental character.
SpatialQuantitative
4Hydrology & Water Bodies
Named rivers, tributaries, beels, haors, baors, canals with specific local names — never default to a regional river if a more locally accurate identification is possible. Water body area (seasonal and permanent); river discharge from BWDB gauge stations; tidal influence zone; flood frequency and inundation depth class; drainage network connectivity; groundwater depth from DPHE/BWDB records; arsenic contamination status from DPHE maps.
SpatialQuantitativeTemporal
5Disaster Vulnerability & Climate
Historical flood years and severity (FFWC); river erosion hotspot unions (BWDB erosion index); cyclone/tornado/nor'wester/hailstorm frequency and documented impact years; drought occurrence by type (pre-monsoon Rabi/Pre-Kharif; post-monsoon Kharif); Kalbaishakhi seasonal storm damage years; rainfall annual mean and distribution (BMD); temperature range; siltation as a chronic problem distinct from acute flood events; water logging duration per year. Identify which specific unions are disproportionately vulnerable to which specific hazards.
RiskTemporal
6Biodiversity & Ecology
Named fish species (resident + migratory); bird species (resident + migratory); mammals, reptiles, amphibians recorded; plant species of homestead forest; presence of IUCN Red List threatened species; documented fish production from open water capture fisheries; historical records of species now locally extinct — this last category is often obtainable only from community oral history and FGD sessions with long-resident farmers and fishers.
QualitativeSemi-quantitative
7Agricultural Profile
Cropping intensity and pattern (Boro/Aman/Aus area from DAE block data); major crops and seasonal calendar; fertiliser use per hectare; pesticide type and application rate; irrigation coverage (surface/groundwater %); yield trends from BBS crop statistics; documented fish kill events linked to agrochemical runoff. The 25% agrochemical runoff contamination rate into soil and water is a well-documented national-level figure for Bangladesh (Bangladesh Environment Conservation Act 1995 enforcement context).
Quantitative
8Environmental Pollution
Number and location of brick fields; proximity of brick fields to agricultural land; type of land affected (fertile double/triple-crop land is the critical concern); air pollution types (smoke, dust, heat); health impacts on surrounding communities (chest/lung disease); domestic solid waste management status; industrial effluent sources; water quality (DO, BOD, turbidity) from DoE/DPHE if available; enforcement records from DoE district office. Topsoil extraction for brickmaking is a distinct threat requiring separate treatment from brick field air pollution.
RiskSemi-quantitative
9Socioeconomic & Infrastructure
Total population, literacy rate, per capita income (BBS); main livelihoods and occupational breakdown; road network length (paved/earthen) from LGED; market/hat bazaar locations; school and health facility count; electricity and water supply coverage; migration trends linked to environmental stress. Critically: road construction without adequate drainage culverts is a primary driver of water logging — document specific unplanned road interventions where known.
QuantitativeSemi-quantitative
10Legal & ECA/ESA Status
Whether any part of the upazila falls within a DoE-designated Ecologically Critical Area (ECA). If no official designation exists: identify the most ecologically significant local wetland, beel, or river system by name, treat it as a de facto ecologically sensitive area, describe its biodiversity and threat profile in detail, and argue for formal recognition under existing law. Key applicable laws: Bangladesh Environment Conservation Act 1995; Bangladesh Environment Conservation Rules 1997; ECA Notification SRO No. 210, 1999; Wildlife (Conservation and Security) Act 2012; Wetland Policy 1998.
Legal
Primary institutional sources — Bangladesh context
How to analyse each data type — methods & outputs
ALand Use Change Analysis
Overlay two-date LULC maps (e.g., SPARRSO 2010 vs Sentinel-2 2024). Calculate net change per zone in hectares and percentage. Use QGIS or ArcGIS. Key metrics: agricultural land loss rate per year, water body area reduction, settlement expansion index. A transition matrix shows which zones converted to what — for example, how much agricultural land became settlement or brick field between two dates. This is the single most important analysis for a land zoning report.
GIS — QGIS / ArcGISQuantitative output
BFlood Vulnerability Analysis
Cross-reference FFWC inundation data with BWDB land type classification (F0=non-flooded, F1=shallow, F2=medium, F3=deeply flooded). Map union-level flood frequency from historical records. Overlay with population density and crop land to produce a socio-ecological vulnerability score per union. Identify the worst-affected unions by name — this is what distinguishes a specific, credible report from a generic one.
GIS overlayRisk scoring
CWater Body Depletion Assessment
Use NDWI (Normalized Difference Water Index = (Green − NIR) / (Green + NIR)) from Landsat/Sentinel time series to track water body area change over 10–20 years. Run this in Google Earth Engine (free) using the upazila boundary polygon as the region of interest. Cross-validate with BWDB beel and haor records. Calculate the annual rate of wetland loss. Identify siltation-affected khals and beels by location. This analysis directly supports the ECA/ESA argument in the report.
Remote sensingTime-series
DSoil Quality & Agricultural Land Profiling
Map SRDI soil series for the upazila. Cross with land type classification and crop intensity data (double/triple-crop areas). Identify prime agricultural soils and flag areas at conversion risk. Agrochemical load is estimated from DAE fertiliser use data and converted to kg/ha per season. Cross this with DoF fish kill event records and beel/river locations to establish proximity between agrochemical use and aquatic mortality events.
QuantitativeCross-source
EBiodiversity Status Assessment
Compile a species inventory from secondary literature, DoF records, IUCN Bangladesh Red List (Volumes 1–7), and primary FGD sessions with local fishers, farmers, and UP members. Classify species by IUCN status: Least Concern, Near Threatened, Vulnerable, Endangered, Critically Endangered. Document locally extinct species based on community oral history — this is often the only source for historically present species. No statistical modelling required; systematic documentation and classification is the output.
Qualitative inventoryIUCN classification
FECA / ESA Suitability Analysis
Apply DoE ECA criteria from the 1999 SRO No. 210 notification: biodiversity significance, migratory bird presence, fish spawning and nursery grounds, threatened ecosystem type, ecological connectivity value. Score major wetlands and beels against these criteria. If a specific beel or haor meets the threshold but has no formal designation, document the scoring evidence and argue explicitly for designation under the Bangladesh Environment Conservation Rules 1997. This is a formal legal argument, not a vague recommendation.
Legal analysisCriteria scoring
GBrick Field Impact Analysis
Map all brick field locations from satellite imagery or field survey. Draw 500 m and 1 km buffer zones around each. Calculate agricultural land area within buffers that is affected by topsoil extraction and air pollution fallout. Cross with SRDI fertile soil zones to identify the quality of land being destroyed. This produces a localised pressure index on prime agricultural land. Also cross-reference with school, health facility, and settlement locations within the air pollution buffer for health impact assessment.
GIS buffer analysisPressure index
HClimate Risk Narrative Synthesis
Use BMD rainfall trend data, FFWC historical flood records, and BBS crop loss statistics together. Identify years with compound disasters — flood compounded by storm, or drought following flood damage. Write a qualitative narrative identifying which unions face the highest multi-hazard exposure. Also incorporate climate change projections for the district from BCAS or CEGIS reports. This does not require modelling — documentary analysis of existing records suffices for an upazila-level report.
Temporal synthesisMulti-hazard
Step-by-step report production workflow — one upazila
1
Define upazila scope & confirm identity
Collect administrative identity: district, division, union list, total area, population. Confirm boundary from DLRS mouza maps or CEGIS polygon data. Get GPS centroid coordinates. Critically: confirm the primary river(s), beels, and haors by cross-referencing BWDB records and SPARRSO maps — do not assume the district-level river is the primary local river. This verification step prevents the most common geographic error in upazila reports.
2
Compile secondary data from institutions
Pull BBS census data, SRDI soil reports, BMD climate normals, DAE crop area, DoF fisheries data, BWDB flood and erosion records, DPHE groundwater maps. Create a structured data file per upazila (JSON or spreadsheet) using a consistent schema with defined field names and values. Flag every field where data is unavailable or estimated — these become your "আনুমানিক তথ্য অনুযায়ী" notes in the final report.
3
Process land use / zoning data
Input provided zoning data (zone name, area in acres, percentage of total). Calculate totals. Identify dominant zones, critically small zones, and zones that are absent entirely. Calculate ratios: agriculture vs settlement; water body vs upazila total; forest vs settled area. Draft analytical commentary — this critical interpretation is what distinguishes a substantive report from a simple data recitation and is required by the report prompt.
4
GIS analysis (where maps are available)
Using QGIS: overlay land use zones, SRDI soil map, BWDB flood map, and brick field locations. Produce buffer analysis around water bodies and brick fields. Calculate NDWI change from Sentinel-2 imagery using Google Earth Engine (free). Run the upazila boundary as the region of interest, extract water body area for two time periods (e.g. 2010 and 2024), and calculate the change. Output summary statistics per zone. This step can be skipped if maps are genuinely unavailable — note the limitation in the report.
5
Primary data collection — field / consultation
FGD (Focus Group Discussion) with farmers, fishers, and UP members at union level. Document local knowledge on species loss, flood years, water logging duration, brick field locations and impacts. Record oral history of environmental change. Use a structured questionnaire with consistent fields across upazilas for comparability. This is the only source for locally extinct species, recent fish kill events, and the lived impact of water logging on agriculture. Cannot be replaced by secondary data.
6
Build the structured data file (JSON / YAML)
Compile all collected data into a single JSON or YAML file per upazila using the master schema. Every field has a defined key. Fields with no data have a null value and a reason string in a notes section. This file becomes the single source of truth for report generation — if a number is not in this file, it cannot appear in the generated report. This schema also serves as the audit trail for QC verification.
7
Generate Bangla report (AI-assisted via Claude API)
Feed the structured data file and land use data into the report prompt template. Use the Claude API (claude-sonnet-4) with the formal Bangla environmental report prompt. The prompt injects all location-specific variables from the data file. Review the output immediately: any number in the report that does not appear in the data file is a potential hallucination and must be flagged before the QC step.
8
Automated QC check
Run the QC checker script against the generated report and its source data file. The checker verifies: word count (1,500–2,000), all numbers traceable to the data file, no unexpected English words, no headings/bullets/tables in the prose, no excessive placeholder citations, and upazila/district name consistency. Fix every flagged issue before proceeding to expert review. A report that fails QC must not go to the expert — that wastes expert time on avoidable errors.
9
Expert review, citation check & final submission
A subject-matter expert familiar with that district reviews the QC-passed draft. They verify qualitative accuracy: species names, soil series names, river names, and historical flood events specific to that upazila. All citations are verified against the actual source documents. Unverifiable citations are replaced with [তথ্যসূত্র যাচাই প্রয়োজন]. Final Bangla text is formatted in Word/PDF with government report conventions. Archive the data JSON file for future update cycles.
Estimated time per report
Secondary data compilation
4–6 hrs
GIS analysis (if maps available)
3–5 hrs
Field consultation (FGD)
1–2 days
Data file assembly
1–2 hrs
AI report generation
15–30 min
QC check (automated)
5–10 min
Expert review + edit
2–4 hrs
Total per upazila
2–4 days
Scale note: A district with 12 upazilas can be processed in 4–6 weeks if secondary data compilation is parallelised across team members and field visits are batched by geography. The AI generation and QC steps are effectively instantaneous at scale once the data files are ready.
Automation strategy for batch upazila report production
Core principle: Environmental reports at upazila level share ~80% of their structure — only the specific local data changes. The right investment is not in writing each report faster, but in building a data pipeline once that makes every subsequent report faster, more consistent, and more verifiable.
AStep A — Build the master data schema (JSON)
Design a single JSON schema that covers all data fields needed across any upazila in Bangladesh. Each upazila gets its own
.json file using this schema. Fields with no data are set to null with a reason string in a notes section. This schema is the backbone of the entire system — never skip defining it properly.{
"administrative": { "upazila": "Bijoynagar", "district": "Brahmanbaria", ... },
"hydrology": { "primary_river": "Titas", "named_beels": [...], ... },
"disasters": { "flood": { "major_flood_years": [1974,1988,1998,2004] }, ... },
"land_use": { "data_year": 2026, "zones": [...] },
"eca_status": { "officially_designated_eca": false, ... },
"citations": { "bbs": "BBS 2022...", "srdi": null },
"_notes": { "null_reasons": { "srdi": "Not yet retrieved" } }
}
BStep B — Build the master prompt template
Create a prompt template where all location-specific variables are placeholders:
{UPAZILA_NAME}, {DISTRICT}, {PRIMARY_RIVER}, {FLOOD_YEARS}, {LAND_USE_DATA}, etc. The template stays fixed across all upazilas and districts. Only the data injected into it changes per report. Version-control the template — any change to the prompt affects every future report.CStep C — Python script: data file → filled prompt → API call → output
A Python script reads each upazila
.json file, fills in the template placeholders, sends the complete prompt to the Claude API, and saves the output as a .txt and .docx file. Run in a loop for all upazilas in a district with rate-limit delay between calls.# Install: pip install anthropic python-docx
import anthropic, json
client = anthropic.Anthropic(api_key="your-key")
# Load upazila data
with open("bijoynagar.json") as f:
data = json.load(f)
# Build and send prompt
prompt = TEMPLATE.format(
LOCATION = data["administrative"]["upazila"],
DISTRICT = data["administrative"]["district"],
LANDUSE = json.dumps(data["land_use"]["zones"], ensure_ascii=False),
RIVERS = data["hydrology"]["primary_river"],
FLOOD_YRS = ", ".join(str(y) for y in
data["disasters"]["flood"]["major_flood_years"])
)
message = client.messages.create(
model="claude-sonnet-4-20250514",
max_tokens=4096,
messages=[{"role": "user", "content": prompt}]
)
# Save output
report = message.content[0].text
open(f"{data['administrative']['upazila']}_report.txt","w",
encoding="utf-8").write(report)
DStep D — GIS automation (QGIS Python / Google Earth Engine)
Use PyQGIS (the Python API inside QGIS) or Google Earth Engine Python API to auto-generate standard analysis outputs for each upazila by bounding box: NDWI water body area for two dates, NDVI vegetation cover, LULC classification, flood zone overlay. GEE is free and requires only a Google account. A single GEE script can run for all upazilas in a district sequentially, outputting a CSV of summary statistics per upazila that feeds directly into the data schema files.
# Google Earth Engine — NDWI water body change (example)
import ee
ee.Initialize()
upazila_boundary = ee.FeatureCollection("your-asset-path")
s2_2010 = ee.ImageCollection("LANDSAT/LT05/C02/T1_L2") \
.filterBounds(upazila_boundary) \
.filterDate("2010-01-01","2010-12-31") \
.median()
s2_2024 = ee.ImageCollection("COPERNICUS/S2_SR") \
.filterBounds(upazila_boundary) \
.filterDate("2024-01-01","2024-12-31") \
.median()
# NDWI = (Green - NIR) / (Green + NIR)
ndwi_2010 = s2_2010.normalizedDifference(["SR_B2","SR_B5"])
ndwi_2024 = s2_2024.normalizedDifference(["B3","B8"])
water_2010 = ndwi_2010.gt(0).selfMask()
water_2024 = ndwi_2024.gt(0).selfMask()
# Then compute area statistics per upazila...
EStep E — Automated Word document assembly (python-docx)
Use the
python-docx library to auto-insert the generated Bangla text into a pre-designed Word template with the government report header, proper page margins (1.2 in top/bottom, 1.3 in left/right), Bangla-compatible font (SutonnyMJ or Kalpurush), page numbering, and section spacing. Export to PDF using LibreOffice headless mode. No manual formatting needed per report.# Install: pip install python-docx
from docx import Document
from docx.shared import Pt, Inches
doc = Document("govt_letterhead_template.docx")
doc.add_heading(f"{upazila} উপজেলা — পরিবেশ প্রতিবেদন", level=1)
for paragraph in report_text.split("\n\n"):
p = doc.add_paragraph(paragraph.strip())
p.style.font.size = Pt(12)
doc.save(f"{upazila}_env_report.docx")
# Convert to PDF (requires LibreOffice installed)
import subprocess
subprocess.run(["libreoffice","--headless","--convert-to","pdf",
f"{upazila}_env_report.docx"])
FStep F — Automated QC checker
A Python QC script reads the generated report and its source data file and runs seven checks: (1) word count 1,500–2,000; (2) all numbers in report traceable to data file — any number not in the data file is flagged as a potential hallucination; (3) no unexpected English words in Bangla text; (4) no headings/bullets/tables in prose; (5) count of [তথ্যসূত্র যাচাই প্রয়োজন] placeholders; (6) upazila and district names present; (7) null fields noted in data file. Outputs a machine-readable QC JSON report plus console summary.
# Run QC on a single report
python report_qc_checker.py \
--report ./reports/Bijoynagar_report.txt \
--data ./data/bijoynagar.json \
--save-json
# Run QC on an entire district batch
python report_qc_checker.py \
--reports-dir ./reports \
--data-dir ./data \
--save-json
# Output example:
# ✓ [word_count] 1,847 words — OK
# ✗ [hallucination_guard] 3 number(s) not in source data
# ✓ [forbidden_structure] prose-only format maintained
# ✓ [name_consistency] OK
Critical risk: Claude may still produce plausible-sounding but unverifiable qualitative statements even when all numbers are correct. The QC checker catches numerical hallucinations — but expert review (Workflow Step 9) remains essential for qualitative accuracy: species names, river names, soil series, and historical events specific to that upazila.
Full automation tools summary
Realistic scale estimate: Once the master JSON schema and prompt template are built (one-time investment of ~2–3 days), and secondary data for an upazila is compiled into the schema (~4–6 hours per upazila), the AI generation + QC step adds only ~30 minutes per upazila. A team of three people can produce verified, QC-passed drafts for all upazilas in a district (typically 8–14 upazilas) within 3–4 weeks, leaving expert review as the primary time constraint.