Fourier Transform Infrared Spectrometers (FTIR) play a crucial role in remote sensing, atmospheric monitoring, and gas composition analysis. However, their measurement accuracy is highly susceptible to temperature drift, which results in spectral deviations and compromises the reliability of collected data. This issue is particularly critical in applications where environmental temperatures fluctuate due to variations in geographical location, flight altitude, and seasonal or diurnal temperature changes. As a result, the spectrometer’s ability to provide precise and stable measurements is significantly challenged, limiting its effectiveness in practical scenarios. Temperature-induced spectral drift occurs when changes in ambient conditions affect the optical and electronic components of the spectrometer, leading to systematic errors that must be corrected to maintain measurement integrity. Traditional calibration methods often fail to account for the complex, nonlinear nature of temperature drift, necessitating the development of more advanced correction techniques. To address this issue, this study investigates the effects of temperature variations on spectral measurements and proposes a novel deep learning-based correction framework. The primary objective is to restore spectral data obtained at non-standard temperatures to their corresponding standard-temperature equivalents, thereby improving measurement accuracy and enhancing the robustness of FTIR spectrometers for real-world applications. In our experimental setup, spectral curves were collected from a single emission source while the environmental temperature was systematically varied between 55 ℃ and 5 ℃, with 25 ℃ designated as the standard temperature. This dataset allowed for an in-depth analysis of spectral drift characteristics across different thermal conditions. Initially, a Long Short-Term Memory (LSTM) neural network was employed for spectral correction, leveraging its ability to capture temporal dependencies in sequential data. The LSTM model successfully learned the nonlinear relationships between temperature variations and spectral deviations, leading to a significant improvement in correction accuracy compared to traditional approaches. However, while LSTM effectively addressed short-term dependencies, it exhibited limitations in capturing long-range correlations across the spectral data. To further enhance correction accuracy and improve generalization performance, we introduced a Hybrid Model (HM) that integrates LSTM with Transformer. This model exploits the advantages of both architectures: LSTM excels in sequential feature extraction, while Transformer’s self-attention mechanism enables global feature modeling, allowing the network to identify complex dependencies between different spectral wavelengths. By combining these two approaches, the Hybrid Model offers a more comprehensive correction mechanism, reducing errors and improving spectral alignment across varying temperatures. Experimental results demonstrate that, compared to standalone LSTM, the Hybrid Model significantly improves key performance metrics, including mean squared error (MSE) and coefficient of determination (R²). Specifically, the Hybrid Model achieves a 86% reduction in prediction errors relative to standard spectral values, demonstrating its superior capability in mitigating spectral drift. These findings confirm that the proposed method effectively enhances the accuracy and stability of FTIR spectrometer measurements under varying temperature conditions. This study highlights the effectiveness of combining LSTM and Transformer in a Hybrid Model to address spectral temperature drift in FTIR spectrometers. By leveraging LSTM’s sequential learning capabilities and Transformer’s ability to model global dependencies, this approach offers a more accurate and robust correction framework than conventional deep learning techniques. The Hybrid Model significantly improves spectral correction accuracy, ensuring that FTIR spectrometers maintain high measurement precision even in harsh and rapidly changing environmental conditions. The results also suggest that this method has broad applicability beyond FTIR spectrometers, as it can be extended to other spectroscopic instruments experiencing temperature-induced deviations. This has important implications for various fields, including remote sensing, environmental monitoring, industrial quality control, and atmospheric research. Looking ahead, future work will focus on optimizing the Hybrid Model’s computational efficiency to enable real-time spectral correction, improving its adaptability to a wider range of operational scenarios. Additionally, further investigations will explore integrating physics-informed models with deep learning techniques to enhance interpretability and reliability. As advancements in remote sensing and spectroscopy continue, refining temperature drift correction algorithms will be essential to achieving more precise, stable, and versatile spectroscopic measurements across diverse applications.