LRTACF-NET: Lowest-Resolution Temporal Attention and Cross Feedback for Multi-Temporal Crop Classification

Accurate land cover and land use (LCLU) classification is critical for environmental monitoring, agricultural planning, and sustainable development. However, distinguishing spectrally similar land categories, such as crop types, remains challenging due to the limited ability of traditional methods to extract discriminative features. To address this, we propose a multi-feedback mechanism with a lightweight self-attention model, where multi-scale feature maps are progressively enhanced through deep supervision for robust feature extraction and fusion. Leveraging the high-resolution satellite time series data, LRTACF-NET in this paper demonstrates significant improvements over state-of-the-art approaches, achieving +10% mIoU and +9% mF1 in quantitative metrics. Notably, while maintaining high accuracy, our Model-7 reduces computational costs by 28% in FLOPs compared to UNet++. Although the best-performing model incurs higher computational cost in terms of FLOPs compared to the baseline, it achieves superior classification accuracy over existing LCLU approaches—including state-of-the-art foundation models—especially in mitigating misclassification among spectrally similar crops. Extensive experimental results demonstrate that LRTACF-NET achieves the highest scores in mIoU, mF1, and overall accuracy, thereby offering a scalable solution for precision LCLU mapping, particularly in crop classification.