EvalCommunity Academy Practical Tutorial

AI-Powered Image Classification for M&E: Validate Field Data & Create New Indicators

A practical, no-code tutorial for Monitoring & Evaluation professionals. Learn to use AI to classify field photographs, validate survey data, and construct additional covariates and outcome measures for agriculture, aquaculture, and natural resource interventions.

What this tutorial helps you do

AI-based image classification makes it faster and more consistent to identify and measure farm, fishpond, or infrastructure characteristics at scale — whether from satellite imagery or ground-level field photographs. This tutorial adapts real-world applications from 3ie’s impact evaluation work (source: Making AI work: Practical applications in impact evaluation).

Specifically, you will learn to use AI for two critical M&E tasks:

• Validate or ground truth intervention design — Confirm whether a reported pond, garden, or structure actually exists and appears active.
• Construct additional covariates and outcome measures — Generate structured indicators (e.g., water presence, pond structure, vegetation cover) from images to enrich your analysis.

This approach helps evaluations go beyond survey responses, using AI to systematically verify program implementation and generate richer, more objective measures across many sites.

Learning objectives

Real-world application: Fishpond evaluation (3ie)

In a real impact evaluation, 3ie needed reliable baseline statistics and outcome measures for an aquaculture intervention. The team required detailed information on fishpond characteristics: whether ponds exist, contain water, are formally constructed, and appear in use. Manually reviewing thousands of field photographs would be time-consuming and inconsistent.

The same method can be applied to kitchen gardens, irrigation infrastructure, livestock shelters, reforestation plots, or any visible intervention feature.

Step-by-step AI image classification workflow for M&E

This workflow turns a folder of field photos into structured validation data and new M&E indicators.

# Install and import libraries
!pip install openai pandas -q
import openai
import pandas as pd
import base64
import os
from google.colab import files, userdata
# SECURITY: Use Colab Secrets (Recommended)
# 1. Click on the 🔑 "Secrets" tab in the left panel
# 2. Add a new secret named "OPENAI_API_KEY" and paste your key
# 3. Uncomment the line below and comment out the insecure method
try:
    openai.api_key = userdata.get('OPENAI_API_KEY')
    print("✅ API key loaded from Colab Secrets.")
except:
    # INSECURE FALLBACK - Remove this after testing
    openai.api_key = "your-api-key-here"  # Replace, but for testing only
    print("⚠️ WARNING: Using hardcoded API key. Do not share this notebook.")
# Initialize the OpenAI client (new v1.0+ syntax)
client = openai.OpenAI(api_key=openai.api_key)
# Upload images
print("📤 Please upload your image files (JPEG or PNG):")
uploaded = files.upload()
# Define the enhanced prompt for M&E classification
prompt_text = """
You are an M&E assistant. Analyze the fishpond image carefully.
First, provide a 1-sentence plain-language description of what you see.
Then, classify using these exact categories. Return ONLY valid JSON with no extra text.
{
    "image_description": "Your 1-sentence description here.",
    "pond_exists": "yes/no",
    "water_visible": "yes/no",
    "water_size": "small/medium/large/not_applicable",
    "water_color": "clear/green/brown/murky/not_applicable",
    "pond_structure": "formal/informal/not_visible",
    "surrounding_land": "vegetation/bare_rock/mixed/not_visible"
}
"""
# Function to encode image
def encode_image(image_path):
    with open(image_path, "rb") as img_file:
        return base64.b64encode(img_file.read()).decode('utf-8')
# Process each image with error handling
results = []
for image_name in uploaded.keys():
    print(f"🖼️ Processing: {image_name}")
    try:
        base64_image = encode_image(image_name)
        response = client.chat.completions.create(
            model="gpt-4-turbo",  # or "gpt-4o" for newer model
            messages=[
                {"role": "system", "content": "You are an expert M&E assistant. Respond only with valid JSON."},
                {"role": "user", "content": [
                    {"type": "text", "text": prompt_text},
                    {"type": "image_url", "image_url": f"data:image/jpeg;base64,{base64_image}"}
                ]}
            ],
            max_tokens=500,
            temperature=0.2  # Lower temperature for more consistent, deterministic outputs
        )
        ai_output = response.choices[0].message.content
        results.append({
            "image_id": image_name.split('.')[0],  # Remove file extension
            "ai_classification_json": ai_output
        })
        print(f"   ✅ Success")
    except Exception as e:
        print(f"   ❌ Error: {e}")
        results.append({
            "image_id": image_name.split('.')[0],
            "ai_classification_json": f'{{"error": "{str(e)}"}}'
        })
# Save results
df = pd.DataFrame(results)
csv_file = "ai_classifications.csv"
df.to_csv(csv_file, index=False)
print(f"\n🎉 Done! Download '{csv_file}' from the file browser on the left.")

Critical quality checks (from 3ie’s experience)

As highlighted in the original 3ie blog, AI outputs are not deterministic and mistakes will occur. Build in these safeguards:

• Human review sample: Review 10–15% of AI classifications manually to assess measurement error.
• Ask AI for a plain-language description: In addition to classification, ask the model to describe the image. This helps catch edge cases (e.g., photos taken in the dark classified as “black” water). (Note: Our template above already includes this!)
• Establish ground truth benchmarks: For a small subset (e.g., 50 images), have two human coders classify independently and compare with AI.
• Document everything: Keep records of prompts, model versions, and error rates.

“The lesson across all these use cases is the same: AI can dramatically expand what’s possible in evaluation, but it demands the same rigor, and arguably more critical thinking than any other method.” – 3ie

Prompt pack for image classification in M&E

Adapt these prompts to your own intervention context.

Recommended tools for AI image classification

Related EvalCommunity Academy tutorials

Final checklist before using AI image classification in M&E

• ✅ Have you defined the specific classification variables needed for validation or outcomes?
• ✅ Are your images clearly named to match survey IDs?
• ✅ Have you tested the prompt on a small sample (20–30 images) and reviewed outputs?
• ✅ Have you built in human review of at least 10–15% of classifications?
• ✅ Have you calculated validation metrics (agreement, false positive rate, false negative rate)?
• ✅ Are you storing raw AI outputs separately from cleaned data?
• ✅ Have you documented your prompt, model version, and error checks?
• ✅ Have you planned how to use the new covariates in your analysis?

Turn field photos into structured M&E evidence

EvalCommunity Academy helps M&E professionals use AI tools with structure, transparency, and methodological rigor.

Explore EvalCommunity Academy