Enhanced ResNet for lake wetland components classification based on Sentinel-2 composites: A case study of Taihu Lake, eastern China

Abstract

Lake wetland ecosystems perform critical ecological functions such as water purification, biodiversity maintenance, and climate regulation, making accurate and fine lake wetland components classification essential for ecological health assessment and productivity accounting. Although deep convolutional neural networks (CNNs) have demonstrated strong potential in image recognition, their application to fine-grained classification of lake wetlands remains limited due to the complex spectral characteristics created by water, land and vegetation interactions. This study developed an enhanced Multi-level Dual-Attention(MLDA)-ResNet50 deep learning model using multi-temporal Sentinel-2 data to achieve integrated classification of wetland components in Taihu Lake. A median compositing strategy based on phenological windows was implemented to address severe cloud contamination. Sample scarcity issue at local scales was resolved through upsampling, enabling native-resolution pixel-level classification via probability-based sliding window accumulation. Key improvements on the architectural framework of CNN included: embedding CBAM modules within residual blocks to enhance the discriminative power for key feature extraction, Reducing spatial attention kernel to prevent edge distortion in upsampled data., and proposing a multi-level dual-attention feature fusion (MLDA-FF) mechanism to integrate shallow texture features with deep semantic features. Experimental results showed an overall accuracy at 95.6% in the classifications, representing a 4.8% improvement over the baseline ResNet50, with substantially escalated performance in spectrally similar land cover mixtures and small-scale feature areas. This research validates the applicability of CNNs for lake wetland components classification in Taihu Lake, and offers new methodological insights for future studies through its improved classification framework and data preprocessing strategy.