The Zhurong rover of the Tianwen-1 mission has detected sulfates in its landing area. The analysis of these sulfates provides scientific evidence for exploring past hydration conditions and atmospheric evolution on Mars. As a non-contact technique with long-range detection capability, Laser-Induced Breakdown Spectroscopy (LIBS) is widely used for elemental identification on Mars. However, quantitative analysis of anionic elements using LIBS remains challenging due to the weak characteristic spectral lines of evaporite salt elements, such as sulfur, in LIBS spectra, which provide limited quantitative information. This study proposes a quantitative analysis method for sulfur in sulfate-containing Martian analogs by leveraging spectral line correlations, full-spectrum information, and prior knowledge, aiming to address the challenges of sulfur identification and quantification in Martian exploration. To enhance the accuracy of sulfur quantification, two analytical models for high and low sulfur concentrations were developed. Samples were classified using infrared spectroscopy based on sulfur content levels. Subsequently, multimodal deep learning models were developed for quantitative analysis by integrating LIBS and infrared spectra, based on varying concentrations. Compared to traditional unimodal models, the multimodal method simultaneously utilizes elemental chemical information from LIBS spectra and molecular structural and vibrational characteristics from infrared spectroscopy. Considering that sulfur exhibits distinct absorption bands in infrared spectra but demonstrates weak characteristic lines in LIBS spectra due to its low ionization energy, the combination of both spectral techniques enables the model to capture complementary sample features, thereby effectively improving prediction accuracy and robustness. To validate the advantages of the multimodal approach, comparative analyses were conducted against unimodal methods. Furthermore, to optimize model performance, different feature selection algorithms were evaluated. Ultimately, an XGBoost-based feature selection method incorporating prior knowledge was employed to identify optimal LIBS spectral features, and the selected feature subsets were utilized in multimodal modeling to enhance stability. Experimental results demonstrate that, compared to the BPNN, SVR, and Inception unimodal methods, the proposed multimodal approach achieves at least a 92.36% reduction in RMSE and a 46.3% improvement in R².

Pine wilt disease (PWD) is a highly destructive forest disease. However, when PWD datasets are constructed, and image detection is conducted, Unmanned Aerial Vehicle (UAV)-based remote sensing detection often encounters the challenge of illumination-degraded UAV imagery. To address this problem, we propose a model called PineSight, which integrates three key enhancements into the Real-Time Detection Transformer (RT-DETR) framework, i.e., Contrastive Language–Image Pre-training (CLIP-LIT), Content-Aware Feature Re-Assembly (CARAFE) and Efficient Vision Transformer (EfficientViT) modules. The proposed model achieves optimal performance across multiple evaluation metrics; for example, the mAP@0.5, mAP@0.5:0.95, Precision and F1-score reach high values of 87.19%, 49.57%, 91.08% and 87.06%, respectively. Simultaneously, the parameter count is reduced from 20.08 M to 11.06 M, and GFLOPs decrease from 58.3 to 28.6. Evaluation on two independent datasets containing images captured under both normal and illumination-degraded conditions shows that PineSight exhibits robustness against illumination degradation. Even under such challenging conditions, the model maintains high performance; it shows high Precision and F1-score with values of 88.10% and 82.23%. The PineSight model not only improves the efficiency of PWD detection but also provides new perspectives and approaches for intelligent forest pest detection in illumination-degraded UAV imagery.