Change detection in bitemporal remote sensing images aids in immediate disaster relief efforts by identifying and analyzing affected areas. However, due to the diversity of backgrounds such as forests, grasslands, and deserts, existing mainstream methods lack generalization when detecting disaster-stricken regions. This leads to issues such as blurry edges and missed detections of small targets. To address these problems, we propose a Change Detection Network (called DCAFNet) that integrates diverse convolutions and attention mechanisms. Based on an encoder-decoder structure, DCAFNet employs a diverse convolution module as the backbone to extract edge features. Additionally, it utilizes a lightweight attention module to enhance the flow of contextual information, improving the detection of small targets. Experiments on a representative landslide dataset validate the capability of DCAFNet. The results show that DCAFNet achieved a Kappa coefficient and F1 score of 81.86%, 91.27% respectively on this dataset, demonstrating the effectiveness of DCAFNet.