Assessment of the impact of the forest fire pollutants on vegetation and crop health in Northeast India

Wildfires are increasing tremendously in Northeast India mostly due to anthropogenic intervention. Indian states are prone to high-intensity fire events resulting in long-term impact on the forest ecosystem such as changes in the vegetation pattern and life cycle of species, decrease in certain species population, and quality of vegetation. Also, forest fires are known to influence the air pollution rate, and the pollutants generated through forest fire are way more complex when compared with urban air pollutants as the composition of forest fire pollutants (FFP) depend on the type of vegetation in the region. Additionally, volatile organic compounds (VOC’s), soot, ozone and black carbon which are few of the products from forest fire has the ability to travel far away from the source fire affecting different ecosystem at various ways and intensity.

As FFP can cause a decline in local, regional and global terrestrial productivity, a deep learning model will be useful in understanding and assessing the impact on vegetation health. Consequently, it is necessary to model the effects of the forest fires, so that their effects in both nearby and far off areas is understood.  In order to accomplish this, we employ Bi-directional Long Short-Term Memory (BiLSTM). The advantages of using a LSTM model are its ability to learn from Spatio-temporal series of data, avoid vanishing and exploding gradient problem, while being tractable to train. The inputs consist of concentrations of, aerosol, carbon monoxide, ozone, black carbon density, evapotranspiration, leaf area index, soil moisture, temperature and relative humidity obtained from MERRA – 2, MODIS and Sentinel – 5P satellite datasets accessed using google earth engine portal. Utilizing these datasets, the normalized vegetation index can be predicted. Standard techniques (mean squared error, root mean squared error, mean absolute error and mean absolute percentage error) are employed to determine the performance of the algorithm.

Keywords: forest fire, pollutants, BiLSTM, google earth engine, deep learning