Remote Sensing · Agricultural Monitoring

Satellite-Based Rice Paddy
Mapping at Scale

An adaptive phenological framework that maps 732,345 hectares of rice cultivation across 32 districts using Sentinel-2 imagery and district-specific calibration, achieving 93.3% accuracy in one of India's most fragmented agricultural landscapes.

Sentinel-2 MSI (10m) Google Earth Engine Telangana, India Published · Environmental Challenges 2025

01 / System Architecture

End-to-End Processing Pipeline

A cloud-native pipeline built on Google Earth Engine that ingests multi-temporal Sentinel-2 imagery, applies phenology-driven classification with district-specific calibration, and produces validated rice area maps at administrative scale.

Data Acquisition

🛰

Sentinel-2 MSI

L2A Surface Reflectance · 10m · 5-day revisit

📍

Reference Fields

953 polygons · 32 districts · Manual digitization

🗺

Auxiliary Data

JRC Water Masks · ESA WorldCover 10m

Preprocessing (GEE)

Cloud Masking & Quality Control

QA60 band filtering · 2×IQR statistical outlier removal · Radiometric normalization (0-1)

Feature Engineering

NDVI

EVI

LSWI

MNDWI

SAVI

5 spectral indices capturing vegetation vigor, water presence, and soil background

Phenological Framework

💧

Land Prep

Dec-Jan · Flooding

🌱

Vegetative

Jan-Mar · Greening

🌾

Reproductive

Mar-Apr · Peak NDVI

🍂

Ripening

Excluded · Spectral ambiguity

Phenologically-aligned compositing · District-specific temporal windows · Up to 50-day variation

Stage-Wise Classification

District-Specific
Threshold Optimization

Multi-index thresholds per stage · Spectral ratio criteria
NDVI/LSWI, EVI/LSWI, NDVI/SAVI combinations

Temporal Analytics

TSP: NDVI σ stability filtering per stage
TPA: Peak/increase/decrease pattern matching

Post-Classification Refinement

IQR Outlier Filtering

ESA WorldCover Masking

JRC Water Body Exclusion

20m Focal Mode Filter

Validation & Output

✅

Pixel-Level Validation

Stratified random sampling
4 field size categories

📊

Area Estimation

Dual-reference: GOI + Telangana
R² = 0.981 / 0.920

🌍

District-Wise Rice Maps

732,345 ha · 93.3% accuracy
32 districts · 10m resolution

02 / Methodology

Key Innovations

What makes this framework different from standard crop classification approaches.

District-Specific Calibration

Rather than applying uniform parameters, each of the 32 districts gets independently optimized thresholds. Initial cluster-based grouping showed >20% threshold divergence within clusters, proving district-level adaptation is essential for heterogeneous smallholder systems.

Phenologically-Aligned Compositing

Image composites are built around crop growth stages, not calendar dates. This normalizes the temporal dimension by biological development rather than time, handling the 50-day planting date variation across Telangana's diverse agro-climatic zones.

Dual Temporal Analytics

TSP (Temporal Stability Parameters) filters mixed pixels by flagging erratic NDVI patterns. TPA (Temporal Pattern Analysis) verifies the flood-green-senescence trajectory. Together they reduced commission errors by ~40% versus static thresholds.

Strategic Complexity

Spectral ratio criteria (NDVI/LSWI, EVI/LSWI) are applied selectively, only where basic thresholds fail. 17 of 32 districts achieved excellent results with simple thresholds alone. Complexity is added only where needed, maintaining computational efficiency.

Sub-Hectare Field Detection

Average field size is 0.16 hectares, with 79% of fields below 0.2 ha. The framework maintains 91.9% accuracy even on these tiny fields at 10m resolution, mapping fields well below 1 hectare across all 32 districts.

Operational Scalability

Built entirely on Google Earth Engine with minimal training data requirements. No deep learning, no GPU infrastructure needed. The index-based approach is interpretable, reproducible, and deployable in resource-constrained environments.

03 / Live Output

From Satellite Imagery to Crop Map

The same geographic area shown at three stages of the pipeline: raw input, classified output, and field-level delineation.

1 Satellite Input

Raw high-resolution imagery showing the fragmented agricultural landscape with a central reservoir. Field boundaries are visually indistinct at this scale.

2 Classification Output

Color-coded paddy field classification by hectare

Classified paddy fields color-coded by area (hectares). The algorithm correctly excludes the reservoir and non-agricultural areas while detecting fields across varying sizes.

3 Field Delineation

White boundary delineation of detected paddy fields

Close-up view showing individual field boundaries extracted from the classification. White outlines trace the detected paddy parcels against the satellite base layer.

04 / Results

Classification Outputs

Visual outputs from the phenology-driven classification across Telangana state.

State-Wide Rice Distribution Map

732,345 hectares of rice cultivation mapped across 32 districts. Highest concentrations in Nalgonda (86,574 ha) and Suryapet (85,754 ha).

Classification Performance

93.3% overall accuracy. 16 districts >95%, 21 districts >90%. Two districts achieved perfect 100% accuracy.

Area Estimation Validation

R² = 0.981 vs GOI statistics, R² = 0.920 vs Telangana Dept. Total mapped area within 1.4% of official figures.

High-Resolution Validation

True-color satellite, binary mask, and detection overlay for 6 representative districts across all agro-climatic zones.

Spectral-Temporal Signatures

Rice paddy's distinctive flood-green-senescence trajectory across 5 spectral indices, clearly separating from other land covers.

Field Size Analysis

6.8 percentage point accuracy decline from medium to tiny fields. Omission errors dominate in sub-0.2 ha fields.

05 / Validation

District-Level Performance

Area estimation validation against official statistics for top rice-producing districts.

District	Mapped (ha)	Official (ha)	Difference	Avg Field Size	Overall Accuracy
Nalgonda	86,574	86,191	+0.4%	0.23 ha	94.8%
Suryapet	85,754	82,472	+3.9%	0.58 ha	96.4%
Nizamabad	64,486	67,088	-3.9%	0.71 ha	95.8%
Karimnagar	46,820	48,707	-3.9%	0.19 ha	100.0%
Jagtial	44,831	—	—	—	92.2%
Khammam	33,614	31,843	+5.6%	0.12 ha	89.0%
Hanumakonda	13,635	—	—	—	99.6%
Medak	12,442	—	—	—	100.0%

Performance by Field Size

Category	Size Range	Accuracy	Kappa	Dominant Error
Tiny	< 0.2 ha	87.2%	0.721	Omission (93.1%)
Small	0.2 - 0.8 ha	91.8%	0.815	Omission (63.1%)
Medium	0.8 - 4.0 ha	94.0%	0.867	Balanced
Large	> 4.0 ha	96.8%	0.927	Minimal

Satellite-Based Rice Paddy
Mapping at Scale

End-to-End Processing Pipeline