<p>Remote sensing image classification using deep learning methods faces challenges such as high complexity, significant computational demands, and inefficiency on resource-constrained devices, while also being affected by issues like class similarity and spatial distribution. Current convolutional neural networks rely on stacking small convolutional kernels for feature learning, which results in relatively low classification accuracy, while their dependence on centralized learning architectures with high-performance GPUs/CPUs incurs substantial training costs. Therefore, this paper proposes a distributed rapid classification method for high-similarity natural scene remote sensing images using an improved VGG19 model (RS-VGG19) that combines residual connections and attention mechanisms. By introducing residual connections, the method improves training convergence speed and high-level feature learning ability, effectively preventing gradient vanishing during training. Embedding the SENet visual attention module in the tenth convolution layer allows the model to more specifically extract similar and significant features in remote sensing images. By employing a combination of cross-entropy and center loss functions, the model is able to learn features with reduced intra-class variance and increased inter-class variance, further enhancing classification accuracy. The distributed inference framework Spark is employed for decentralized model training, storing large-scale remote sensing images in the distributed file system HDFS, and accessing the pre-trained RS-VGG19 model in Docker containers on cluster nodes for distributed inference and classification using PySpark. Experimental results show that on two commonly used high-similarity remote sensing image datasets, NWPU-RESISC45 and UC Merced Land-Use, the RS-VGG19 model improves classification accuracy by 6.57% and 8.76% respectively compared to the original VGG19 model, and significantly enhances accuracy compared to other related classification models. This demonstrates the superior performance of the proposed structure and loss function fusion strategy in remote sensing image classification tasks. On the large-scale remote sensing image inference dataset NWPU-RESISC45, while maintaining classification accuracy, the distributed inference framework achieved a speedup of 11.9 when using six nodes, an improvement of 98.33% over theoretical linear speedup (6.00), reducing dependency on high-end hardware resources and significantly improving the classification speed of high-similarity natural scene remote sensing images.</p>