Land Use Change Detection Using Deep Siamese Neural Networks and Weakly Supervised Learning

A weakly supervised change detection method is proposed for remotely sensed multi-temporal images, by utilizing a Siamese neural network architecture. The architecture of the Siamese network is a combination of two multi-filter multi-scale deep convolutional neural networks (MFMS DCNN). Initially, the Siamese network is trained by utilizing the image-level semantic labels of the image pairs in the dataset. The features of the image pairs are obtained using the trained network to generate the difference image (DI). Then, a combination of the PCA and the K-means algorithms has been used to produce the change map for the pair of images. Experiments were carried out using two remotely sensed image datasets. The weakly supervised method proposed in this paper offers better results in comparison to both weakly supervised- and unsupervised-based state-of-the-art models and techniques.