目的 利用深度学习开展变化检测是遥感智能解译热点研究方向之一。针对变化检测模型的轻量化部署问题，本文设计一种基于特征交互和特征融合的轻量级变化检测网络。方法 提出了一种基于特征交互和特征融合的轻量级变化检测网络(feature interact and feature mixing light network, FIFMLNet)。解码器中采用EfficientNet作为特征提取网络，其能利用模型的放缩(model scaling)能力来扩大模型的感受野。然后通过设计通道、像素交互模块(spatial and channel interact block)和浅层跳跃连接(low-level skip-connection)来实现浅层双时相的细节特征交互和上采样阶段的传递，以此增加模型对局部特征的判别精度。此外，利用特征融合分组卷积模块(feature fusion and groups convolution block，FFGCB)对双时相数据进行降维融合，来降低模型计算量。最后，设计了融合上采样模块(fusion upsampling block，FUB)对局部特征与全局特征进行融合还原，同时利用局部特征的细节、纹理来补偿全局特征细节的缺失。结果 本文方法在两个遥感影像数据集(LEVIR-CD和SYSU-CD)上与13种SOTA(state-of-the-art)方法进行比较。客观上，本文方法对比现有变化检测方法在各项定量评价指标上均具有明显优势。在LEVIR-CD和SYSU-CD数据集中，本文方法F1分别取得91.51%和82.19%，相较于对比方法的最优值分别提升了0.43%和1.58%，并且模型的每秒浮点运算量和参数量分别为1.66G和0.56M，低于所有对比方法。主观上，本文方法相对于对比方法的检测区域准确、漏检率低，具有丰富的细节。结论 本文提出的轻量级变化检测网络FIFMLNet以较少的参数量和每秒浮点运算量获得了优越的性能，改善了小目标漏检、边界误检的情况，能够获得高质量的变化检测结果。

Objective Change detection in remote sensing images is a process that leverages remote sensing technology to compare and analyze images from the same area taken at different times, and this process aims to identify changes on the Earth’s surface. The main challenge of this technique is to extract effective change features from a vast amount of image data and map them onto pixel-level change labels for high-precision detection. Methods for detecting changes in remote sensing images can be divided into traditional methods and those based on deep learning. Traditional methods primarily rely on image processing and pattern recognition techniques. However, these methods also require manual selection of suitable features and thresholds, which can introduce subjectivity and limitations. On the other hand, deep learning methods can automatically learn abstract and high-level change features from remote sensing images, enabling end-to-end change detection This significantly . improves the accuracy and efficiency of change detection. These methods can be further categorized into those based on Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Attention mechanisms. Models based on these mechanisms have made significant strides in recent years, thanks to the extensive research conducted by scholars worldwide. However, due to the complexity of image information and imaging differences, current change detection methods still face challenges in dealing with the ‘pseudo-change’ problem. Method This paper proposes a light change detection method for remote sensing images based on feature interact and feature mixing. This method aims to overcome the limitations of existing methods, such as low accuracy, high computational cost, and poor generalization ability. The main idea of this method is to use EfficientNet B7 as a lightweight backbone network to extract both deep and low-level features from bi-temporal remote sensing images. To enhance the spatio-temporal feature fusion, we introduce channel swap and pixel swap modules to enable the interaction and combination of bi-temporal features. To preserve more edge and texture details and reduce artifact generation, we employ low-level skip-connections to transfer the original image information to the up-sampling phase. To effectively fuse the deep and low-level features obtained in the down-sampling stage, we design a feature fusion group convolution module that reduces the computational overhead and the number of parameters. Finally, we use the feature fusion group convolution module and the up-sampling module to fuse and recover the deep and low-level features, and generate the pixel-level change detection map. Result In this paper, we conducted experiments on two datasets for remote sensing image change detection LEVIR-CD and SYSU-CD. We split each dataset into 7:1:2 for training, validation and testing, and segmented each image into 256×256 pixels. This facilitated the processing and increased the generalization ability of the model. We used the binary cross entropy (BCELoss) as the loss function, and evaluated the performance of our proposed method using three metrics: F1 score (F1), Intersection Over Union (IoU), and Overall Accuracy (OA). Our method achieved F1 scores of 91.51% and 82.19%, IoU 84.35% and 69.76%, and OA of 99.14% and 91.99% on the LEVIR-CD and SYSU-CD datasets, respectively. Compared with previous classical methods, such as DSIFN, DTCDSCN, STANet, SNUNet, BiT, Changeformer, DDPM-CD, USSFC-Net, ELGC-Net, LRNet, etc., our model obtained the best change detection results, especially preserving more details on the change boundary. To illustrate the effect of low-level skip-connections and channel and spatial exchange modules, we also performed ablation experiments on the same dataset. The results showed that the channel and spatial exchange module significantly optimized the utilization and representation of spatio-temporal features in the network, while the low-level skip-connection compensated for the loss of detailed features in the downsampling process and further enhanced the feature learning capability of the network. Conclusion our network used channel and spatial exchange modules to increase the utilization and understanding of spatio-temporal features, and low-level skip-connections to focus the model on local detailed features. Finally, we used a binary cross-entropy loss function at the output layer to achieve optimal change detection performance. Experiments show that the method proposed in this paper can improve the ability of recognizing changing regions while ensuring a light network, and can improve the detection performance of change detection in various environments and terrains, as well as improve the phenomenon of "pseudo-change".