Transformer-Guided Noise Detection and Correction in Remote Sensing Data for Enhanced Soil Organic Carbon Estimation

Soil organic carbon (SOC) is a critical indicator of soil health, directly influencing crop productivity, soil structure, and environmental sustainability. Existing SOC estimation techniques using satellite reflectance data are effective for large-scale applications; however, their accuracy is reduced due to various types of noisy samples caused by vegetation interference, sensor-related anomalies, atmospheric effects, and other spectral distortions. This study proposes a robust data refinement framework capable of handling any soil sample, whether clean or noisy, by identifying and correcting noisy samples to enable more accurate SOC estimation outcomes. The approach first explores complex global relationships among spectral bands to understand and represent subtle patterns in soil reflectance using the Transformer network. To remove redundancy and retain only essential information of the transformed features, we apply a dimensional reduction technique for efficient analysis. Building upon this refined representation, noisy samples are detected without relying on strict data distribution assumptions, ensuring effective identification of noisy samples in diverse conditions. Finally, instead of excluding these noisy samples, the proposed framework corrects their reflectance through a conditional Generative Adversarial Network (cGAN) to align with expected soil spectral characteristics, thereby preserving valuable information for more accurate SOC estimation. The proposed approach was evaluated on benchmark satellite datasets, demonstrating superior performance over existing noise correction techniques. Experimental validation using the Landsat 8 dataset demonstrated that the proposed framework improved SOC estimation performance by increasing R2 by 1.52%, reducing RMSE by 4.45%, and increasing RPD by 5.14% compared to the best baseline method (OC-SVM + Kriging). These results confirm the framework’s effectiveness in enhancing SOC estimation under noisy conditions. This scalable framework supports accurate SOC monitoring across diverse conditions, enabling informed soil management and advancing precision agriculture.