Nrdly
Get Nrdly Free Trial Built with Nrdly

Mapping Fire Severity in the Miombo Woodlands

Introduction

Fire is a common disturbance that affects ecosystem processes, vegetation structures, and atmospheric carbon balances globally. Researchers report that forest fires have increased intensity and frequency due to climate change and human activities. Over the past decades, most burnt areas have been located in Africa. Generally, most forest fires occur in tropical and subtropical savanna regions like the Miombo woodlands.

The Miombo woodlands are located in the savanna ecosystems. These woodlands cover about 2.7 million km2 in Central, Eastern, and Southern Africa. Tree species such as Brachystegia and Julbernardia and a continuous grass layer dominate the Miombo woodlands. The grass layer provides fuel during the dry season, exposing the Miombo woodlands to fire. Consequently, fire significantly impacts above-ground woody biomass and vegetation structure in the Miombo ecosystem, resulting in low tree density. Therefore, there is a need for a continuous data- and evidence-driven fire management system to mitigate severe wildfires.

Earth Observation (EO) Data and Machine Learning for Fire Severity Mapping

Fire severity refers to the loss of above- and below-ground organic matter. National forestry agencies in the Miombo woodland ecosystem require accurate and consistent national- and subnational-scale fire severity maps because fire severity is an essential component of the fire regime. Fire severity influences the post-fire response of vegetation, increases erosion, and changes water quality. Therefore, forest researchers and policy-makers need fire severity maps to: (i) assess trends in fire regimes, (ii) understand and gain insights into fire effects on the ecosystem, and (iii) prepare effective fire management strategies. While researchers in advanced economies have used Earth observation (EO) data to map fire severity, mapping is still rare in developing regions. National forestry agencies can use EO data and machine learning approaches to map fire severity and improve fire management in the Miombo woodlands.

During the past decades, Satellite-based EO sensors provided vast volumes of remotely-sensed data. Remote sensing researchers have used optical and SAR sensors to map fire severity. For example, researchers have used coarse spatial resolution satellites such as the Moderate Resolution Imaging Spectrometer (MODIS) and medium resolution satellites such as Landsat and Sentinel-1. In addition, researchers have used process-based, statistical, and machine learning-based modeling approaches to map fire severity. While machine learning algorithms have effectively predicted fire severity, they are complex and challenging to interpret. Hence the need for data-centric explainable machine learning methods.

Next Steps

This blog post will use Fire Information for Resource Management System (FIRMS) data, Sentinel-2 spectral indices, Sentinel-2 differenced Normalized Burn Ratio (dNBR), Sentinel 1, forest height, and digital elevation model derivatives (e.g., slope). We will also use the LIME method to gain insights into a random forest model. Readers can access the blog tutorial and download data sets in the links below.

Access the tutorial here

Download data here

There are many resources to learn about data-centric and explainable machine learning. You can also check my book to learn about data-centric explainable machine learning methods.