Enhancing change detection in multi-temporal optical images using a novel multi-scale deep learning approach based on LSTM

Multi-temporal change detection is crucial for effective environmental monitoring and resource management, enabling informed decision-making in diverse applications such as urban planning, agriculture, and disaster response. Traditional methods, such as pixel-based differencing, Change Vector Analysis, threshold-based approaches applied to spectral indices like NDVI, and Support Vector Machine post-classification, often fail to effectively capture spectral-spatial relationships. This limitation arises from their focus on individual pixels without considering neighboring pixel information, resulting in lower accuracy for change detection and classification tasks. This study aims to enhance change detection performance in remote sensing applications using Long Short-Term Memory (LSTM)-based deep learning models. Despite the success of deep learning in image processing, challenges remain in obtaining diverse and sufficient training data and addressing spatial scale variations. To tackle these issues, we implemented a semi-automated stage that performs clustering on temporal information from optical time series images. Experts then assign semantic information to these clusters, ensuring representative and diverse training samples while reducing redundancy. The collected data are used to train our proposed deep models, which are specifically designed for change detection. We introduce two innovative deep learning models to improve change detection by extracting comprehensive spectral-spatial features. The first model, 3Branch-3DConvNet (3BDCN), uses data patches of various sizes to handle diverse spatial scales in the images. The second model, 3Branch-3DConvLSTMNet (3BDCLN), replaces the conventional CNN with a 3D-ConvLSTM block to better capture temporal and spectral dependencies in satellite data. Classified maps generated by these models are then processed to produce from-to-change maps, highlighting the changes detected in the study areas. Experiments conducted on three case study areas demonstrate that our models outperform conventional methods and are competitive with state-of-the-art approaches. On average, across all regions, the kappa coefficient improved by 1.5 %, 3.12 %, and 8.5 % compared to 3DCNN, MSCNN, and ANPC, respectively, underscoring the effectiveness of 3BDCLN in enhancing change map accuracy. This research also offers a scalable solution for effective environmental monitoring and change detection.