Enhancing U-Net performance for high-resolution land cover classification using a dynamic epoch-centric optimizer (DECO)

Abstract

In recent years, deep learning models—particularly U-Net—have garnered significant attention for applications such as high-resolution land cover mapping. A key challenge in improving these models' performance lies in the proper selection and tuning of optimizers: each algorithm (e.g., Adam, Nadam) offers distinct strengths and weaknesses, and reliance on a single optimizer may not yield optimal results across all training stages. Here, we introduce DECO, a novel hybrid optimizer that dynamically switches among multiple optimizers across epochs to enhance overall convergence and stability. U-Net trained with DECO on aerial imagery of buildings, forests, roads, and water in the Minski region of Warsaw, Poland, achieved 96.13 % overall accuracy, a Kappa coefficient of 91.49 %, an F1 score of 96.08 %, and a Jaccard index of 64.53 %. To assess generalizability, the model was further evaluated on a test region in the Malopolskie province, yielding 86.74 % accuracy, 73.75 % Kappa, 87.29 % F1, and 55.02 % Jaccard. Moreover, to demonstrate DECO's broader applicability, we implemented it on the DeepLab v3+ architecture, observing likewise improvements in validation accuracy and training stability. These findings substantiate that dynamic, epoch-centric optimizer switching can substantially boost the precision and robustness of deep learning models for high-resolution land cover classification.