Hyperspectral Remote Sensing of Soil Properties Using Machine Learning
Kerri Lu, Massachusetts Institute of Technology
Soil composition significantly impacts agricultural development and crop yields as well as soil greenhouse gas emissions. However, spatial mapping of soil properties has historically been limited and coarse-grained because the laboratory processes for chemically analyzing soil samples are often costly. Hyperspectral satellite imaging, which captures a target area’s reflectance of radiation at many wavelengths across the electromagnetic spectrum, is an efficient method for remote sensing of physical and chemical characteristics of a region at a finer-grained spatial resolution. In this work, we train supervised machine learning algorithms to predict soil nutrient concentrations from hyperspectral reflectance profiles, using ground-truth soil nutrient survey data and hyperspectral satellite imagery collected from locations in Mexico, India, and Kenya. We empirically identify the spectral features most useful for predicting each nutrient and evaluate generalizability of learned features across the three locations, toward the goal of developing a spatially robust predictive algorithm for global soil nutrient mapping.
Abstract Author(s): Kerri Lu