DSeP-xNet: A Feature Optimized Ensemble Framework for Hyperspectral-Based Soil Organic Carbon Prediction

Accurate measurement and management of soil nutrients are crucial for smart agriculture. Among various soil nutrients, Soil Organic Carbon (SOC) is one of the most essential measures that form the basis of this work. Unlike approaches that use individual Machine Learning (ML) and Deep Learning (DL) models with minimal emphasis on data cleaning and feature selection, this paper introduces a novel methodological framework called DSeP-xNet. The framework includes preprocessing techniques like noise removal with the Savitzky–Golay filter and feature selection using XGBoost, as minute changes in features using such techniques significantly affect the model accuracy. It also anchors around previously established models to configure a stacking-based ensemble model, DSeP-Net, combining the strengths of three base models: a Deep Neural Network (DNN), a Squeeze and Excitation Convolutional Neural Network (SECNN), and Partial Least Squares Regression (PLSR), with Linear Regression as the meta-model. The study uses data points from two publicly available hyperspectral imagery data: Africa Soil Information Service (AfSIS) and Land Use and Coverage Area Frame Survey (LUCAS-2009) for SOC prediction. The datasets leverage the individual pixels of satellite images over the narrow band of reflected light as features. The DSeP-xNet framework demonstrates a remarkable R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and RMSE score of 0.967 and 0.2202, respectively, on the AfSIS dataset and 0.9448 and 22.7193 on the LUCAS-2009 dataset, attesting to its high accuracy and robust nature in SOC prediction. These findings pave the way for more precise and efficient data-driven decision-making in agriculture and environmental conservation.