To improve model convergence and performance on minority classes, I employed an exponential decay learning rate schedule, where the learning rate decreased by a factor of 0.1 every 30 epochs. This approach allowed the model to make large updates during early training while fine-tuning with smaller updates as training progressed, which was particularly beneficial for accurately classifying the underrepresented soil classes

Additionally, I incorporated a warmup period during the first 5 epochs, where the learning rate gradually increased from 1e-5 to the initial rate of 1e-3. This technique helped stabilize the early training phase and improved overall convergence, especially when dealing with the complex feature extraction required for distinguishing between spectrally similar land cover classes at 1-meter resolution.

For data augmentation, pixels with weak sun shadows such as water, plants, roads, and soil were enhanced using 60° rotation and horizontal flipping. However, for pixels with strong sun shadows such as buildings and trees, rotation and flipping decreased accuracy. This may be due to consistent shadow orientation in satellite maps taken at the same static time, and the 50ResNet101 model recognizing shadow features. Therefore, when creating satellite map mosaics, each year’s split map not only used the same month (September) but also maps with consistent sun shadow orientation when possible. Since road and soil classes had fewer training samples in the classification task, sample resampling was conducted by undersampling the majority classes while increasing the number of road and soil samples. Additionally, in remote sensing imagery land classification tasks, Focal Loss is a key technique for addressing class imbalance issues.