An Efficient Lightweight Neural Network for Remote Sensing Image Change Detection

Abstract

Remote sensing (RS) image change detection (CD) is a critical technique of detecting land surface changes in earth observation. Deep learning (DL)-based approaches have gained popularity and have made remarkable progress in change detection. The recent advances in DL-based methods mainly focus on enhancing the feature representation ability for performance improvement. However, deeper networks incorporated with attention-based or multiscale context-based modules involve a large number of network parameters and require more inference time. In this paper, we first proposed an effective network called 3M-CDNet that requires about 3.12 M parameters for accuracy improvement. Furthermore, a lightweight variant called 1M-CDNet, which only requires about 1.26 M parameters, was proposed for computation efficiency with the limitation of computing power. 3M-CDNet and 1M-CDNet have the same backbone network architecture but different classifiers. Specifically, the application of deformable convolutions (DConv) in the lightweight backbone made the model gain a good geometric transformation modeling capacity for change detection. The two-level feature fusion strategy was applied to improve the feature representation. In addition, the classifier that has a plain design to facilitate the inference speed applied dropout regularization to improve generalization ability. Online data augmentation (DA) was also applied to alleviate overfitting during model training. Extensive experiments have been conducted on several public datasets for performance evaluation. Ablation studies have proved the effectiveness of the core components. Experiment results demonstrate that the proposed networks achieved performance improvements compared with the state-of-the-art methods. Specifically, 3M-CDNet achieved the best F1-score on two datasets, i.e., LEVIR-CD (0.9161) and Season-Varying (0.9749). Compared with existing methods, 1M-CDNet achieved a higher F1-score, i.e., LEVIR-CD (0.9118) and Season-Varying (0.9680). In addition, the runtime of 1M-CDNet is superior to most, which exhibits a better trade-off between accuracy and efficiency.

3M-CDNet-V2: An Efficient Medium-Weight Neural Network for Remote Sensing Image Change Detection

“…For example, the input of Layer 2 is a fused feature representation obtained by concatenating the high-level X 1 and low-level X 0 along the channel axis, and the input of Classifier 1 is a representation obtained by fusion or concatenation of X 0 , X 1 , and X 2 . In the original version of 3M-CDNet, there is only one skip connection sending the X 1 -X 2 fused representation to Classifier, and the authors claimed that more connections seem helpless in improving the model performance [11]. However, in our case, the opposite strategy, which can take full advantage of the improved feature representation X 0 via dense skip connections, was proved to work better.…”

“…In this way, the intermediate layers are effectively trained and weights of them can be finely updated [16], thus alleviating the presence of vanishing gradient and improving the model performance. Deep supervision strategy facilitates the network convergence during the training phase, whereas it brings about more computation and memory cost than single-head prediction [11], so the original 3M-CDNet did not apply it. However, we will prove that deep supervision is crucial in 3M-CDNet-V2.…”

“…For example, the CNN (i.e., Convolutional Neural Networks) architecture itself could achieve some degree of shift, scale, and distortion invariance [7]. Over the past years, enormous efforts have been made to use deep learning-based methods for CD using high-resolution RS images, readers are referred to [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], and [20]. A wide variety of research has indicated the superiority of these methods over the traditional machine learning or unsupervised methods, e.g., Support Vector Machine, Random Forest, Slow Feature Analysis [21], Robust Change Vector Analysis [22], and so on.…”

“…Beginning with UNet-based models namely FC-EF, FC-Siam-conc, and FC-Siam-diff [10] and their revolutionary performance on the classical CDD and BCDD datasets [5]. Siamese CNN architectures have evolved to become increasingly powerful through more dense skip connections [5], larger scale [11], more sophisticated forms of convolution [14], semisupervised architectures [12], [13], as well as hybrid loss function [15]. At the same time, the Transformer-based network architectures used for CD have also received increased attention [23], [24], [25], and they achieved competitive performances compared to their CNN counterparts.…”

“…In this article, we will report a novel Siamese CNN-self attention combined architecture (termed 3M-CDNet-V2) for change detection tasks using bitemporal high-resolution imagery. As its name implies, 3M-CDNet-V2 is the updated version of 3M-CDNet [11] which is a lightweight CD network only involving about 3.12 M trainable parameters. 3M-CDNet was designed by reducing the width and depth of the backbone network, thus producing high-resolution feature maps, and it facilitates the detection of detailed spatial information (including small changed objects) with acceptable computational costs.…”