Micro-near-infrared (Micro-NIR) spectroscopy has emerged as a promising technique for accurate and cost-effective estimation of soil attributes compared to traditional wet chemistry methods and conventional NIR spectroscopy. Despite its potential, the full extent of its capabilities and applications under different soil conditions remained unexplored. This study has evaluated the potential of a low-cost micro-NIR sensor for predicting soil organic carbon (SOC) and clay contents in agricultural soils under dry and fresh conditions. Two sets of soil samples i.e., A (92 samples, Netherlands) and B (92 samples, Belgium, France, and Germany) were collected and analysed for SOC and clay contents using standard laboratory methods. A micro-NIR of 2000 ∼ 2450 nm spectral range with 18–28 nm resolution (NIRONE D2.5, Spectral Engine, Germany) was used to scan both fresh and air-dried soil samples. Partial least squares regression (PLSR) models were developed for each dataset separately with and without feature selection using competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and slime mould algorithm (SMA). Micro-NIR sensor performed differently for the two datasets, predicting SOC and clay contents accurately for the dataset A while failing for the dataset B. The best accuracy achieved for SOC in the dataset A-fresh (coefficient of determination in prediction (RP2) = 0.76 and root mean square errors in prediction (RMSEP) = 0.27 %) was improved with the dry samples (RP2 = 0.81 and RMSEP = 0.27 %). The prediction of clay content was rather poor (RP2 = 0.48 and RMSEP = 5 %). Feature selection by SMA and CARS methods improved SOC prediction for the dataset A of fresh (Rp2 = 0.79 and RMSEP = 0.25 %) and dry soils (Rp2 = 0.84 and RMSEP = 0.25 %), respectively. Using CARS improved the results of clay prediction for the dataset A of dry soil (Rp2 = 0.56; RMSEP = 4.62 %). Poor performance for the dataset B could be attributed to the non-normal data distribution (SOCskewness = 4.30 and clayskewness = 3.05) and the larger soil variability encountered including soil types. Therefore, this study indicates that a micro-NIR sensor is a potential innovation and can predict SOC accurately and clay with rather moderate accuracy for a normally distributed dataset. •Micro-NIR sensor can accurately predict soil organic carbon.•Micro-NIR sensor can predict soil clay content to a certain degree of accuracy.•Feature analysis can enhance the accuracy of micro-NIR for SOC and clay content predictions.•Sample’s attributes distribution has influence on the prediction using micro-NIR.