With climate change and the intensification of human activity, drought event frequency has increased, affecting the Gross Primary Production (GPP) of terrestrial ecosystems. Accurate estimation of the GPP and in-depth exploration of its response mechanisms to drought are essential for understanding ecosystem stability and developing strategies for climate change adaptation. Combining remote sensing technology and machine learning is currently the mainstream method for estimating the GPP in terrestrial ecosystems, which can eliminate the uncertainty of model parameters and errors in input data. This study employed extreme gradient boosting, random forest (RF), and light use efficiency models. Additionally, we integrated solar-induced chlorophyll fluorescence (SIF), near-infrared reflectance of vegetation, and the leaf area index (LAI) to construct various GPP estimation models. The standardised precipitation evapotranspiration index (SPEI) was utilised at various timescales to analyse the relationship between the GPP and SPEI during dry years. Moreover, the potential pathways and coefficients of environmental factors that influence GPP were explored using structural equation modelling. Our key findings include the following: (1) the model combining the SIF and RF algorithms exhibits higher accuracy and applicability in estimating vegetation GPP in the arid zone of Xinjiang, with an overall accuracy (MODIS R²) of 0.775; (2) the vegetation in Xinjiang had different response characteristics to different timescales of drought, in which the optimal timescale for GPP to respond to drought was 9 months, with a mean correlation coefficient of 0.244 between grass land GPP and SPEI09, indicating high sensitivity; (3) using structural equation modelling, we found that temperature and precipitation can affect GPP both directly and indirectly through LAI. This study provides a reliable tool for estimating the GPP in Xinjiang, and its methodology and conclusions are important references for similar environments. In addition, this study bridges the research gap in drought response to GPP at different timescales, and the potential influence mechanism of natural factors on GPP provides a scientific basis for early warning of drought and ecosystem management. Further validation using a longer time series is required to confirm the robustness of the model.

As a typical arid region, the change in Xinjiang’s vegetation carbon footprint is crucial for assessing ecological restoration and resource allocation. This study analyzes the changes in the vegetation carbon footprint and its influencing factors in Xinjiang by employing a range of models, including Net Ecosystem Productivity (NEP), carbon emission fitting, carbon footprint analysis, and structural equation modeling (SEM). Furthermore, using the carbon deficit vegetation investment estimation method, we quantify the additional vegetation area and investment required for Xinjiang to achieve carbon neutrality. The results show the following: (1) Net Ecosystem Productivity (NEP) increased slowly, with six regions (Altay, Bortala, Bayingolin, Kizilsu Kirghiz, Tacheng, and Yili) contributing 66.95% of the total NEP, forming the main carbon sink. Meanwhile, carbon emissions rose significantly, coming largely from Urumqi, Changji, Kumul, and Karamay (61.31% of total emissions). (2) The carbon footprint expanded 3.44 times, from 30.41 × 10⁴ km² to 104.49 × 10⁴ km². Human activities were the main positive driver, while vegetation factors negatively influenced the carbon footprint. (3) Based on the 21-year average carbon deficit, achieving carbon neutrality in Xinjiang requires an estimated investment of USD 106.77 × 10⁸ to expand cropland, woodland, and grassland by 8029 km², 1710 km², and 35,016 km², respectively. Implementing vegetation expansion, improving carbon markets, and transforming carbon-source economies are essential to achieving the “double carbon” goal. This study clarifies regional carbon sources/sinks and supports the carbon neutrality strategy in arid ecosystems.