| Summary: | Land use activities such as agriculture and forestry have a significant impact on greenhouse gas emissions, but these impacts are often underestimated. This is because agriculture and forestry sectors cover a large land area, making on-ground surveys difficult to scale, expensive, and time-consuming. While there have been localized examinations of urban and industrial domains, it remains unclear how agriculture and forestry sectors would fare in a global examination using modern tools such as hyperspectral cameras.
Hyperspectral cameras can substantially bridge the gap between uncertainties in observations and offer little room for disparities in measured readings between countries, resulting in an unbiased survey. However, there are spatial (resolution) and temporal (date availability for long dates) gaps in hyperspectral satellite data. These gaps can be filled using machine learning algorithms.
The algorithm utilised paired data from the Sentinel 2 (RGB) images and the corresponding nitrous oxide reading from the Sentinel 5P satellite to obtain seasonal reading levels in 24 different locations spanning 8 different soil varieties. The data was analysed to obtain valuable insights on the variances between readings for a range of time between September 2018 to August 2022 (4 years). This data can be used to develop more effective policies to reduce emissions and mitigate climate change.
A Spatio Temporal Neural network was created that can ingress an RGB image and its associated date and obtain the associated nitrous oxide reading accurate within 12% (Mean Absolute Percentage Error) of the actual value. This model can be utilized as a prototype to fill in gaps in existing hyperspectral images with reasonable compromise in accuracy.
|
|---|