Image classification has an important role in many applications e.g., Land Use Land Cover (LULC) and emergency responses. Different methods have been used by researchers to improve performance by using Low-Level, High Level and fusing features. Using the Convolutional Neural Networks (CNNs) provide a great performance but with a great cost of computational complexity, so custom networks start to gain a lot of attention. This paper presents an Ultra-Light Weight Fused Network (ULWFN) architecture that depend on Atrous Convolution with four different fusion types. The ULWFN performance was compared with pre-trained networks, custom networks and Low-Level features on three different datasets AIDER-V1, AIDER-V2 and UC-Merced. Results shows that ULWFN significantly surpasses baseline networks in computational efficiency on datasets. The proposed ULWFN outperforms all low-level feature-based and custom network approaches while delivering competitive classification performance on the AIDER-V1 and AIDER-V2 datasets, achieving up to 79.05% accuracy and 0.8872 F1-score respectively. Although performance gaps emerge on the fine-grained UC-Merced dataset relative to large pre-trained models an expected trade-off of aggressive parameter reduction the model maintains a superior balance between classification accuracy and computational efficiency, requiring as few as 1.70 billion FLOPs, 91K parameters, and 0.35 MB memory while achieving inference speeds of up to 242 FPS, making it a highly suitable solution for real-time emergency response on resource-constrained edge devices.