Abstract:

To address the limitations of conventional remote sensing classification methods, such as reliance on manual feature design and poor generalization ability, this study systematically evaluated three backbone architectures (ResNet34, MobileNetV2_100 and TF_MobileNetV3_Small_100) under frozen and non-frozen transfer learning strategies, utilizing Landsat 8 Operational Land Imager (OLI) multispectral imagery and employing a saline-alkali land information extraction method based on U-Net deep learning model. The experimental findings demonstrated that ResNet34 generally showed superior convergence speed, segmentation accuracy and generalization ability in comparison to the lightweight models (MobileNetV2_100, TF_MobileNetV3_Small_100). Specifically, the non-frozen ResNet34 model achieved the optimal overall performance, with a classification precision of 0.880, recall of 0.708, and F₁-score of 0.785, all of which exceeded those of other models. The lightweight model demonstrated efficacy in scenarios characterized by limited resources, which could be employed in cases where computational resources were limited and segmentation precision required was low but still necessitates high-performance backbones for complex environments. Notably, non-frozen training consistently exceeded frozen strategies, emphasizing the importance of full-parameter optimization for enhancing accuracy and generalization ability. The research not only validates the effectiveness of deep learning in the remote sensing monitoring of saline-alkali land but also provides a model selection framework for intelligent identification and monitoring of saline and saline-alkali land, offering practical guidance for ecological governance and precision agriculture.

Key words: saline-alkali land, U-Net, ResNet34, MobileNet, semantic segmentation, transfer learning, deep learning model