S. Hu, H. Qi, Z. Wang et al.
Environmental Science and Ecotechnology 30 (2026) 100682
on output uncertainty, including all interaction and correlation effects. Y | X i represents the mean model output obtained when the input parameter X i is fixed at a given value. Y | X ∼ i is the mean model output obtained when all input parameters except X i fixed, X ∼ i represents the set of all input parameters excluding X i . 3. Results and discussion
Table 1 Accuracy metrics of semantic segmentation models applied to factory functional zones. Functional zone U P R F 1 IoU Primary fiber stacking area 98.58% 93.84% 93.93% 0.93 94.00% Recovered fiber stacking area 96.45% 93.84% 85.29% 0.89 83.53% Wastewater treatment area 95.39% 92.31% 94.96% 0.93 86.39% Other stacking areas 92.55% 88.23% 77.65% 0.82 81.85% Thermal power plant area 92.90% 90.77% 83.19% 0.86 86.32% Note: U , P , R , F 1 , and IoU are accuracy, precision, recall, F 1 -score, and intersection over union, respectively. An F 1 score closer to 1 denotes better model performance. PFPP, primary fiber pulp plant; RFPP, recovered fiber pulp plant; HPPMP, heavy- weight paper product manufacturing plant; SPPMP, specialty paper product manufacturing plant; LPPMP, lightweight paper product manufacturing plant.
3.1. Classification results
3.1.1. Results of the DeepLabv3 + model We used 720 preprocessed remote-sensing images of PPPs for this study, allocating 70% to the training set and 30% to the test set. We applied five functional zones (defined in Section 2.1.1) and visualized them using distinct colors (Fig. 5): primary fiber stacking (red), wastewater treatment (green), recovered fiber stacking (blue), other stacking (yellow), and thermal power plant areas (purple). The model achieved high classification performance, with recognition accuracies exceeding 90% across all five functional- zones categories (Table 1). The highest recognition accuracy (98.58%) was achieved in the primary fiber-stacking areas, reflecting the presence of highly distinctive visual features. In addition, IoU values for all functional zones exceeded 80%, indi- cating that the model effectively captured boundary characteris- tics and discriminated among different functional zones. Based on this model, we identified the functional zones of 720 PPPs from remote-sensing images. Among these plants, 92 were classified as PFPPs and 108 as RFPPs. 3.1.2. Results of BERT model The model exhibited strong performance across the three cat- egories (HPPMP, LPPMP, and SPPMP), with classification accuracies of 91.62%, 84.55%, and 85.62%, respectively (Table 2). Precision and recall values for all three categories exceeded 80%. These results confirmed that the model effectively captured the distinctive textual features relating to different plant types. The incorporation of textual information substantially improved the classification model's accuracy. The classification of HPPMPs, LPPMPs, and SPPMPs based solely on remote-sensing images is challenging, owing to the limited discriminative visual features of such images. Hence, the textual data provided critical complementary information to overcome these limitations. The
Table 2 Evaluation metrics for text classification models. Plant type U P
F 1
R
PFPP RFPP
66.38% 73.82% 91.62% 84.55%
80.26% 62.77% 88.27% 84.39%
66.30% 81.90% 88.81% 85.28%
0.73 0.71 0.88 0.84
HPPMP LPPMP
multimodal fusion of image and textual data effectively overcame the constraints inherent in single-modality PPP classification. Using a classification framework that integrated semantic im- age segmentation with textual information, we categorized the 720 PPPs by type. HPPMPs constituted the largest group (292 plants; > 40% of the total), followed by SPPMPs (122), RFPPs (108), and LPPMPs (106), whereas PFPPs represented the smallest cate- gory (92). 0.86 Note: U , P , R , and F 1 , are accuracy, precision, recall, and F 1 -score, respectively. An F 1 score closer to 1 denotes better model performance. PFPP, primary fiber pulp plant; RFPP, recovered fiber pulp plant; HPPMP, heavyweight paper product manufacturing plant; SPPMP, specialty paper product manufacturing plant; LPPMP, lightweight paper product manufacturing plant. SPPMP 85.62% 86.26% 85.60%
3.2. Carbon accounting model
Due to potential variations in data accuracy and reporting standards in the reported ESG data, we examined the carbon emissions data prior to model construction and conducted field surveys at a set of representative PPPs to validate the reported emissions values. The survey-based estimates were generally consistent with the reported ESG data (Table 3), with discrepancies predominantly within 10%. Therefore, despite potential un- certainties, we considered the ESG data reliable enough for use in this study. Based on the carbon emissions data collected from PPPs, combined with the classification results from Section 3.1, we used the methods described in Section 2.6 to construct area-based carbon accounting models for five different PPP types defined above. Overall, the models exhibited satisfactory fitting perfor- mance, with all R 2 values exceeding 0.75 (Table 4). Of the five models, the RFPP model achieved the highest fitting accuracy ( R 2 = 0.96, MAPE = 8.10%). In contrast, the SPPMP and HPPMP models generated relatively higher MAPE values, both exceeding 19%. This increase was mainly attributable to the number of indi- vidual plants with exceptionally high carbon emissions within these categories (Supplementary Fig. S3), which led to elevated prediction errors and, consequently, higher overall MAPE values. However, because some enterprises with similar high-emission plants lack publicly disclosed ESG reports, we did not exclude these outliers to maintain the model's representativeness. Overall,
Fig. 5. Semantic segmentation results for pulping and papermaking plants. Different functional zones are identified: primary fiber stacking (red), wastewater treatment (green), recovered fiber stacking (blue), other stacking (yellow), and thermal power plant areas (purple).
7
Made with FlippingBook interactive PDF creator