Calibrated Weak Supervision for Post-Harvest Burned-Cropland Mapping Under Label Scarcity

Raunak Bhagate; Maitri Polisetty; R I Minu

cs.LGOct 4, 2026

Calibrated Weak Supervision for Post-Harvest Burned-Cropland Mapping Under Label Scarcity

Authors: Raunak Bhagate, Maitri Polisetty, R I Minu

Organizations: SRM Institute of Science & Technology

Abstract

Mapping post-harvest burned cropland is difficult when fires are small and fragmented and reliable labels are scarce. We developed a calibrated weak-supervision framework for Punjab, India, using Sentinel-2 spectral change, VIIRS active-fire context, and MODIS MCD64A1 as a coarse external calibration and agreement reference. Three pseudo-label recipes, five feature representations, and linear, tree-based, boosted, and neural classifiers were evaluated using nested district-held-out cross-validation over three seeds and five folds. NBR and dNBR were excluded from classifier inputs. The best configuration used the very-strict recipe, a multilayer perceptron, and the full optical feature set (mean Cohen's kappa 0.395, AUROC 0.753, F1 0.747, balanced accuracy 0.703); Random Forest, XGBoost, and LightGBM were practically tied. Higher agreement with held-out pseudo-labels did not establish improved label correctness or independent burned-area accuracy. For deployment, a Random Forest with the very-strict recipe and full optical features was retained. An externally calibrated threshold of 0.60 yielded district-level MODIS agreement of R-squared 0.636, a mapped-to-MODIS burned-area ratio of 1.005, and Spearman correlation of 0.779 with district fire counts. Pixel-level MODIS agreement remained modest (F1 0.205, kappa 0.093). Zero-shot transfer to Haryana was promising (mean kappa 0.641), but Punjab cross-year stability was weak, and Sentinel-1/Sentinel-2 feature concatenation did not improve the optical baseline. Optical observations ended before the seasonal fire context, limiting coverage of late burns. The framework supports district-scale burden assessment and hotspot screening, with limited support for exact scar boundaries or temporally stable annual mapping.

Figures & tables

Panel of Figure 1 — Figure 1: Study area and active-fire context. (a) Punjab and Haryana in northwestern India. (b) Punjab districts and the ESA WorldCover cropland mask used to restrict analysis. (c) High-confidence VIIRS active-fire density over Punjab cropland from 1 October–29 November 2024.

Figure 2: End-to-end weak-supervision and calibration workflow. Sentinel-2 preprocessing, VIIRS-anchored pseudo-labels, matched district-held-out evaluation, statewide RF inference, external deployment-threshold calibration, and robustness testing are kept as distinct stages. The schematic’s pre/post labels denote the earlier/later optical composites; its October–November season denotes the broader reference period. NBR and dNBR are used for label construction but excluded from classifier inputs.

Panel of Figure 3 — Figure 3: Weak-label construction. Cropland and VIIRS fire context are intersected with burn-consistent spectral gates to produce weak positives, weak negatives, and unlabeled pixels. The displayed parameter block corresponds to the exact very-strict configuration.

Recipe	d NBR	Near km	$dB$ 12	HN d	HN b	Far km	NDVI
Moderate	0.18	3.0	$>200$	0.06	120	8	None
Strict	0.22	2.0	$>200$	0.05	100	8	None
Very-strict	0.28	1.5	$>250$	0.04	80	10	$>0.10$

Table 1: Weak-label recipe parameters. Near- and far-fire distances are in km; HN denotes hard-negative, with

d

and

b

denoting dNBR and

dB12

. The recipes share the same inputs and seasonal windows and differ only in spatial and spectral selection constraints.

Set	Size	Composition	Analytical purpose
A	39	Earlier, later, and delta values for 10 Sentinel-2 bands plus three indices	Full optical production representation
B	13	Delta-only optical variables	Tests whether change alone is sufficient
C	26	Earlier optical state plus delta variables	Retains baseline context with reduced redundancy
D	33	Full optical set without SWIR-related predictors	Primary circularity ablation
E	45	Set A plus Sentinel-1 VV/VH earlier, later, and delta	Simple optical–radar fusion test

Table 2: Feature-set definitions. NBR and dNBR are excluded from every classifier feature set because they are used in weak-label construction.

Rank	Recipe	Model	Set	Mean $\kappa$	SD $\kappa$	AUC	$F_{1}$	Bal. acc.
1	Very-strict	MLP	A	0.395	0.049	0.753	0.747	0.703
2	Very-strict	RF	B	0.395	0.053	0.747	0.744	0.703
3	Very-strict	XGBoost	A	0.394	0.053	0.740	0.747	0.703
4	Very-strict	RF	A	0.393	0.056	0.744	0.746	0.701
5	Very-strict	LightGBM	B	0.393	0.052	0.740	0.748	0.701

Table 3: Leading learned configurations. Values are means across 15 district-held-out evaluations (three seeds

\times

five folds).

Figure 4: Quantity–selectivity pattern across pseudo-label recipes. The very-strict recipe produced fewer positive weak labels and higher mean district-held-out agreement with its own pseudo-labels. Evaluation labels and sampled populations varied across recipes, so this pattern does not establish improved label correctness or independent burned-area accuracy.

Item	Value
Production classifier	Random Forest (300 trees)
Pseudo-label recipe	Very-strict
Feature set	A (39 optical variables)
Median internal CV threshold	0.30
Externally calibrated threshold	0.60
District $R^{2}$ vs. MODIS	0.636

Table 4: Production configuration and selected-threshold calibration/agreement statistics.

Figure 5: External deployment-threshold calibration for RF/very-strict/Set A. The internal threshold (0.30) and calibrated threshold (0.60) are compared across district MODIS fit, aggregate area ratio, fire-count rank correlation, and pixel MODIS $F_{1}$ . The selected cutoff balances the predefined criteria rather than maximizing a single metric.

Experiment	Configuration/protocol	Mean $\kappa$	Interpretation
Haryana zero-shot	Punjab-trained RF/very-strict/A; thresholds 0.30, 0.40, 0.30; no Haryana fitting or calibration	0.641	Promising neighboring-state transfer
Punjab 2023	Same seasonal framework	0.662	Stronger year-specific performance
Punjab 2024	Primary experiment	0.393	Weaker year-specific performance
S1 + S2 fusion	Set E minus optical Set A	$-0.015\,\Delta\kappa$	Simple concatenation did not improve performance

Table 5: Robustness and transfer summary. Cross-year reporting is limited to Punjab 2023 and 2024.

Explore similar work

A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery
Sep 20, 2026Alisson Cleiton de Oliveira, Thales Sehn KörtingRandom Forest Multiple Satellite Systems
Earth Observation and Geoinformatics Division (DIOTG), National Institute for Space Research (INPE), São José dos Campos, SP, Brazil
Low-Rank Adaptation of Geospatial Foundation Models for Wildfire Mapping Using Sentinel-2 Data
May 6, 2026Ali Shibli, Andrea Nascetti, Yifang BanGeospatial Foundation Models Wildfire Prediction
KTH Royal Institute of Technology Stockholm, Sweden
PhenoStitch: Training-Free Panoptic Crop Mapping from Satellite Image Time Series
Aug 1, 2026Xuechen LiAgricultural Monitoring Panoptic Segmentation
University of Minnesota Minneapolis, MN, USA

Sep 20, 2026, cs.LGJ/K move · Enter open · S save

A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery

Alisson Cleiton de Oliveira, Thales Sehn Körting

Earth Observation and Geoinformatics Division (DIOTG), National Institute for Space Research (INPE), São José dos Campos, SP, Brazil

Random Forest Multiple Satellite Systems Portuguese Time Series

Open paper AskPDF arXiv

This paper introduces a multi-temporal tabular dataset derived from satellite images to map burned areas in the Chapada dos Veadeiros National Park, in Goiás, Brazil, covering the years 2020 to 2022. The dataset contains blue, green, red, and near-infrared bands, as well as the BAI, EVI, GEMI, NDVI, and NDWI spectral indices from the WFI sensor on the CBERS-4A, CBERS-4, and AMAZONIA-1 satellites, organized into a regular grid. We applied the Random Forest classifier to develop and validate models based on samples labeled as totally burned, partially burned, and non-burned. Two classification approaches were tested: one combining burned and non-burned areas into binary classes and another distinguishing between totally burned (TB), partially burned (PB), and non-burned (NB) classes. Seven validation approaches assessed different post-classification combinations, focusing on accuracy, precision, recall, and intersection over union (IoU) metrics. Results showed higher IoU when TB, PB, and NB were used as individual classes and TB was reclassified as burned area (BA) while PB and NB were grouped as non-burned. Comparing the annual results of this approach to the MCD64A1 product, the errors of omission for the BA class were 22% in 2020, 28% in 2021 and 59% in 2022, while the errors of commission were 46%, 43% and 46%, respectively. The study highlights the utility of the WFI sensor for burned area mapping without inter-satellite spectral calibration and suggests further exploration with other machine learning algorithms to evaluate the dataset potential and limitations.

Abstract

Explore similar work

A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery

Low-Rank Adaptation of Geospatial Foundation Models for Wildfire Mapping Using Sentinel-2 Data

PhenoStitch: Training-Free Panoptic Crop Mapping from Satellite Image Time Series