Getting the Most Out of the Image-Level Labels: (Un)Supervised Learning for Extracting Soil Parameters From Hyperspectral Images

Technological advancements are expanding the potential of hyperspectral image (HSI) analysis for Earth observation. Extracting insights from high-dimensional images led to various supervised artificial intelligence (AI) approaches. However, the world is not labeled, and acquiring ground truth (GT) is expensive. For certain tasks, such as estimating soil parameters, only coarse-grained, field-level measurements are available. We tackle the challenge of building (un)supervised AI models from weakly-labeled sets with image-level labels. We propose a comprehensive framework for estimating soil parameters from HSIs, and a spectrally- and spatially-informed algorithm for generating pseudolabels based on the original GT. We analyze the spatial variations of spectral pixel characteristics within the parcels (images) using superpixels, which group neighboring pixels based on their spectral features—superpixels are determined using unsupervised clustering. Then, the superpixels are clustered (based on the feature vectors calculated for the superpixels). The resulting clusters are either used to estimate soil parameters for the incoming unseen samples (pixels, superpixels, or images) in a fully unsupervised fashion, or are exploited to elaborate pseudolabels (at the superpixel level) which can be used to train supervised models. The experiments not only showed that our (un)supervised methods outperform the supervised state-of-the-art models on the HYPERVIEW benchmark, collocating HSIs and in-situ measurements of soil parameters, but also indicated that pseudolabels make the modeling process much easier and generalizable, and may be considered as de-noising of the originally acquired data. We release our models and pseudolabels to ensure reproducibility of our study.