Effectiveness of Sentinel-2 Data in Quantifying Agriculture Residue Burning in Madhya Pradesh

~ By Satvik Parashar

​India is a hotspot for residue burning, which is driven by the limited time the farmers have to plant the next crop and a lack of feasible alternatives to handle the crop residue. Not only is this responsible for increased carbon emissions, but the crop residue once burnt also produces radiatively active gases and particulate matter, causing additional contributions to climate change and a public health hazard forhumans.

It is essential to map the agriculture burned areas (ABAs) to analyze and mitigate emissions in the agricultural systems. While global fire emission inventories provide valuable information on the time and location of fires, they underestimate the small and fragmented fires. The smallholder systems have smaller field sizes (< 1.5 ha) and a short duration of fire (~30 minutes). The use of sentinel-2 data, which provide improved spatial (10 m) and temporal resolution could help capture these small and short-lived fires, and the study we summarize here attempts to test its efficacy in doing so. The study was conducted in the state of Madhya Pradesh, which has the majority of its land cover under crop area (Fig. 1) and is one of the most significant wheat-producing regions in Central India.

Fig. 1. (from paper) (a) Location of the study area in India. (b) Zoom in of the study area, the entire state of Madhya Pradesh

​The Sentinel 2 satellite data product (S2A-ABAMP201921) was developed for this study. It was compared with other fire detection data products - Modis Fire burned area pixel product version 5.1 (FireCCI51) and the Modis Terra and Aqua combined version 6 burned area data product (MCD64A1). The resolution of the study’s data product (S2A-ABAMP201921) was 10m, while it was 250m and 500m for the FireCCI51 and the MCD64A1 respectively. All the products were compared from March 2019 to May 2021. The results were validated using high resolution (3m) Planetscope imagery based ground-truth data. Validation results stated that the Sentinel-2 data had the lowest errors and the highest F1 score (a measure of the model’s accuracy) of the three data products in detecting the ABAs (Table 1).

Table 1 (from paper)- Validation accuracy of S2A-ABAMP201921, MCD64A1, and FireCCI51 when compared to PlanetScope burned area maps.

​Across all years, the S2A-ABAMP201921 detected more burned area than the other two data products (Fig3 (a) to (c)). An additional 9179 km2 of burned area per month compared to MCD64A1 and 6931km2 per month compared to FireCCI51. The proportion of burned area for Sentinel-2 was 134% more than MCD64A1 and 113% more than FireCCI51 when comparing the average burned area across all years (Fig3 (d) to (f)). The S2A-ABAMP201921 data product could detect smaller burned areas (<10 ha) while the other two products could detect only larger burned areas patches (>50 ha). The model uses a fire index and proxy index to determine only the fire-affected areas which makes sure the false positive signatures are minimized. Therefore, it is safe to say that the model is not overestimating burned areas.

Fig. 5. (from paper) (a) to (c) Temporal distribution of ABA during the post winter burning season in 2019–2021. (d) to (f) Spatial distribution of S2A-ABAMP201921 showing low and high burn density areas for 2019–2021.

Emissions from agriculture residue burning were assessed by estimating agriculture dry matter, and comparing them with MCD64A1 and the global fire emission database (GFED 4.1 s). The agriculture dry matter estimates of the study's data product were four time higher than MCD64A1 and nine times higher than GFED 4.1 s. The burned area intensity matched with the time of harvest of the wheat crop (April – May) (Fig.5 (c) to (d)). The S2 data product also captured large amount of burned area in regions of high wheat yields, which were mostly northwest and central part of the state, followed by the southwest (Fig. 5 (d) to(f)). The trends in burned areas matched the trends of exponential increase in areas under wheat cultivation.

Fig. 5. (from paper) (a) to (c) Temporal distribution of ABA during the post winter burning season in 2019–2021. (d) to (f) Spatial distribution of S2A-ABAMP201921 showing low and high burn density areas for 2019–2021.

​The accuracy of the studies’ data product was found to be much higher than the traditional data products used for ABA detection. The improvement in accuracy can be attributed to two main reasons. Firstly, the high spatial resolution of the data product (10m) that could better capture small-scale and isolated fires. Secondly, it used long-lasting burned signatures to match burned areas.


Insufficient accounting of the burned areas may result in underestimating the impacts of residue burning in smallholder systems. The study highlights the importance of high-resolution imagery to map burned areas that helps in better calculating emissions from residue burning. High correlation between wheat harvest and burned areas encourage policy interventions to reduce burning of wheat residue after the winter growing season. The study uncovers the emission potential of Central India due to residue burning, which is largely underestimated and unaccounted for. The study also highlights the need of new technological interventions that can help us efficiently track agricultural emissions and help reduce their adverse effects on air quality and climate.

Picture Credits: Amrita Neelakantan, NCCI Coordinator

Categories

All
Agriculture
Community Participation
Corridors
Deforestation
Human Health
Human Wildlife Interactions
Land Use Planning
Livelihoods
LULC
Management
Mining
NCCI
Pollution
Project Update
Science Summary
Spatial Analyses
Water
Wildlife