With the advancement of artificial intelligence, deep learning has become instrumental in land cover classification. While there has been a notable emphasis on refining model structures to improve classification accuracy, it is imperative to also emphasize the pivotal role of data-driven optimization techniques. This paper presents an in-depth investigation into optimizing multi-class land cover classification using high-resolution multispectral images from Worldview3. We explore various optimization strategies, including refined sampling strategies, data band combinations, loss functions, and model enhancements. Our optimizations led to a substantial increase in the Mean Intersection over Union (mIoU) classification accuracy, improving from a baseline of 0.520 to a final accuracy of 0.709, which represents a 35.2% enhancement. Specifically, by optimizing the classic semantic segmentation network in four key aspects, we improved the mIoU by 15.5%. Further improvements through changes in data combinations, sampling methods, and loss functions led to an overall 17.2% increase in mIoU. The proposed model optimization methods enabled the OUNet to outperform the baseline model by providing more precise edge detection and feature representation, while reducing the model parameters scale. Experimental evidence shows that in the application of multi-class land surface classification, increasing the quantity and diversity of samples, avoiding data imbalance issues, is equally valuable for improving overall classification accuracy as it is for enhancing model performance.

With the rapid development of artificial intelligence, significant progress has been made in land cover classification using deep learning methods. However, in existing research, most studies focus more on improving classification accuracy by optimizing the model structure and less on mining the value of the data itself. In this paper, experiments on remote sensing multi-class land cover classification were conducted based on Worldview3 data, and strategies to improve classification accuracy were proposed in terms of sampling methods, band combination, loss function, and model optimization. Experiment results show that the proposed improvement strategies are effective for multi-class land cover classification, with recall, F1, and IoU improved by 29%, 17%, and 19%, respectively. The significant improvement in classification accuracy for less-represented targets confirms that enhancing data richness and balance leads to greater improvement than just optimizing the model.