Abstract: A high-precision satellite imagery has enabled to near real-time monitor the land use change via the classification maps using remote sensing. However, the traditional classification cannot fully meet the current requirements of land change monitoring. Particularly, the images were often blurred with the feature edge and noisy information, leading to a low accuracy classification. In this study, a hybrid Convolutional Neural Network (CNN) model was proposed to accurately classify the land use using a “residual error-squeeze excitation” unit. An expanded convolutional layer was also used in the 3D-2D-CNN, instead of the part of the network structure than before. The “spectral-spatial” features were first extracted from the image using an inflated convolutional layer, and then the perception field of the convolutional kernel was expanded to accommodate the extracted features with different sizes in the less complexity of the model, finally a “residual error-squeeze excitation” unit was introduced in this framework. The input of the unit was directly added to activate the output in the residual link, where a better performance of the model was achieved to avoid the degradation of deep neural networks during natural constant mappings. A squeeze excitation module was adaptively recalibrated the feature responses between channels during the explicit modeling. As such, the interdependence between channels was obtained to realize the feature reuse for the selective emphasis on the informative features. The noise was suppressed from the interdependencies between channels using explicit modeling. After that, a depth-separable convolutional layer was used for the further feature extraction. The final classification of remote sensing images was achieved through a fully connected layer using the Softmax activation function. Taking the 2020 Jiaozuo Landsat-8 image data as an example, the overall accuracy of the model was improved 11.15, 11.18, 0.06, and 2.46 percentage points, respectively, compared with the K-Nearest Neighbor, Support Vector Machine (SVM), 2D-CNN, and Hybrid SN. Furthermore, the Kappa coefficient and the average accuracy were ranked the first, 0.97 and 91.81%, respectively. It infers that the improved model can be widely expected to classify the common feature types. In addition, the least loss of information was obtained about the linear features (such as roads and rivers) and the edge of features in the test area, while ensuring high accuracy. Anyway, the improved model can present a high universality to effectively achieve the higher accuracy and quality classification of Landsat than before.