Improving Remote Sensing Change Detection via Spatial Autocorrelation Regularization and Momentum Orthogonalization Rahmat Faisal
ESRI Indonesia
Abstract
Change detection in remote sensing is vital for monitoring environmental shifts, urban expansion, and disaster impacts. Accurately identifying spatial changes over time is critical for effective policy development, sustainable resource allocation, and responsive early warning systems. However, many existing deep learning models treat change detection as an independent pixel-wise classification problem, failing to account for the spatial correlations embedded in geospatial imagery. This often leads to outputs that are noisy, fragmented, and spatially inconsistent.
To overcome these challenges, we introduce a novel deep learning framework that explicitly incorporates spatial autocorrelation into the training process. Our approach enhances the loss function with a regularization term derived from Moran^s I statistic and spatial neighborhood smoothness, guiding the model to preserve local spatial structures and produce more coherent prediction maps.
In addition, we employ the MuOn optimizer-short for Momentum Orthogonalized by Newton-Schulz-a cutting-edge optimization method that improves training dynamics by orthogonalizing the momentum vector with respect to the gradient direction. This process reduces redundant updates, enhances gradient diversity, and accelerates convergence, which is especially advantageous in high-dimensional remote sensing models.
By combining spatial autocorrelation-aware regularization with the MuOn optimizer, our framework delivers improved spatial coherence and classification accuracy. This makes it a robust, efficient, and interpretable solution for high-resolution remote sensing change detection.
