Evaluation of Transformers and Convolutional Neural Networks for High-Dimensional Hyperspectral Soil Texture Classification

Soil texture is an important parameter influencing a multitude of ecosystem services. However, its determination in the laboratory is complex, time-consuming, and only reveals soil texture at a specific sampling point. Therefore, topsoil soil texture determined from space-borne remote sensing data offers advantages (areal and temporal availability, expanding possibilities with upcoming hyperspectral satellite systems).
Since no hyperspectral satellite data are available, we use hyperspectral reflectance data provided in the Land Use/Land Cover Area Frame Survey (LUCAS) dataset by the European Soil Data Centre. We resample the provided 4200 bands to the Environmental Mapping and Analysis Program (EnMAP) Resolution of 222 bands. Hereafter, we classify soil texture as sandy, silty, clayey, and loamy from these by applying distinct Transformer architecture as well as a one-dimensional convolutional neural network. Our best models multitemporal SimpleVIT and an ensemble approach score 65.89% and 67.62% overall accuracy, respectively.