Deep Learning-Based Adaptive Downsampling of Hyperspectral Bands for Soil Organic Carbon Estimation

Accurate estimation of soil organic carbon (SOC) is critical for assessing soil health and guiding sustainable land management. Hyperspectral sensing has emerged as an approach for SOC analysis due to its ability to capture detailed spectral signatures of soil properties. However, hyperspectral data suffer from redundancy and noise, which can degrade predictive accuracy if not properly addressed. Existing dimensionality reduction techniques, such as fixed-interval downsampling and autoencoders, either risk discarding informative bands or disrupt spectral continuity, limiting their effectiveness for models like one-dimensional convolutional neural networks (1D-CNNs) that rely on local spectral patterns. The recent supervised band selection algorithm BSDR improves accuracy but does not retain spectral continuity. This study introduces a deep learning model, AD-CNN, which employs a novel adaptive downsampling technique that jointly prioritizes band relevance and continuity for SOC estimation. Unlike prior techniques, the proposed technique is supervised, physically interpretable, and fully differentiable, enabling end-to-end learning with minimal parameter overhead. It preserves original spectral bands for scientific traceability and does not require extensive manual feature engineering. Experimental results on a large-scale soil spectral dataset show that AD-CNN improves <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> by 17.95% and 19.48% over BSDR and a recent autoencoder-based baseline, respectively. The selected bands also exhibit automatic exclusion of noisy wavelengths, confirming the model’s ability to learn scientifically aligned features.