Hyperspectral image processing, a critical domain in remote sensing and Earth observation, has witnessed exponential growth in feature extraction methods. In light of this, our study delves into the ongoing debate between spectral and spatial feature extraction strategies, a question of paramount importance. To address this, we embarked on a rigorous investigation using two eminent deep learning architectures, namely 3DCNN(Three dimensional convolutional neural network) and ResNet(Residual Neural Network) and employed two benchmark hyperspectral datasets: Indian Pines and Salinas. Our research was meticulously structured, encompassing spectral feature extraction methods such as mean, mode, standard deviation, and skewness, as well as spatial feature extraction techniques employing advanced convolutional operations. The spectral standard deviation feature extraction method using 3DCNN yields an overall accuracy of 98.71% for the Indian Pines dataset. In the Salinas dataset, the spectral and spatial mean feature extraction methods achieve accuracies of 99.78% and 99.89% respectively. When employing the ResNet architecture, the spectral skewness feature extraction method attains an accuracy of 99.83% for the Indian Pines dataset. Meanwhile, in the Salinas dataset, spectral standard deviation feature extraction excels with an accuracy of 99.99%. Analyzing computational times, the 3DCNN on the Indian Pines data set shows that spatial mode feature extraction has the shortest time of 0.33 s, while in the Salinas dataset, spatial skewness feature extraction requires 1.60 s. For ResNet, Indian Pines’ spatial mode is the quickest at 2.10 s, and Salinas’ spatial mode is 10.66 s. In the comparison of spectral and spatial feature extraction methods, the likelihood of achieving lower computational times favors spectral methods. Additionally, the probability of attaining higher accuracy is notably higher with spectral feature extraction methods. Therefore, spectral approaches emerge as a superior choice.