Enhancing business recommendation through satellite image analysis with multi-visual feature fusion

Satellite imagery offers potential for applications in disaster response, business recommendation, agricultural decision-making, and urban planning, owing to its accessibility and ability to capture details of the Earth’s surface over time. However, traditional methods struggle to extract information from these images. This study leverages deep learning and fusion of visual features from satellite imagery to determine suitable businesses for specific locations in Malaysia. By scrutinizing the surrounding structure and characteristics, this study examines the visual context of business environments. Satellite and map images are processed by transfer learning deep learning models to extract deep features, while hand-crafted statistical features and Scale-Invariant Feature Transform (SIFT) features are extracted from road network images to represent structural patterns of business locations. It is found that combined feature sets are more reliable than individual feature sets for business recommendation from satellite imagery. This study analyzed 12,500 satellite images across five classes, generating four feature sets of 128 and 512 dimensions. The 512-dimensional concatenated feature set achieved the highest accuracy: 0.5932 with an artificial neural network (0.13 improvement) and 0.6068 with a support vector machine (0.25 improvement) over the best individual feature sets.