S. Hu, H. Qi, Z. Wang et al.
Environmental Science and Ecotechnology 30 (2026) 100682
(HSV) color space and thereafter performed contour extraction using OpenCV — an open-source computer vision library for image processing — to delineate PPP areas. To visually demonstrate the effectiveness of the contour extraction method, we selected several PPPs with different geometric shapes (Fig. 2). Finally, we annotated functional zones within each plant using LabelMe software, an open-source image annotation tool, following the zone definitions in Section 2.1.1. To minimize potential interference from Chinese punctuation and rare characters on downstream classification, we preprocessed the textual data using the Jieba word-segmentation tool. Specif- ically, we normalized the original text by removing Chinese punc- tuation and infrequent characters to enhance the consistency of textual representations and the robustness of the classifier. We then classified and labeled PPPs by integrating information derived from remote-sensing images with plant-level product data.
Specialty paper product manufacturing plants (SPPMPs). Plants that lack both primary fiber and recovered fiber stacking areas and produce mainly specialty papers, such as wallpaper or filter paper (Supplementary Fig. S2d). Lightweight paper product manufacturing plants (LPPMPs). Plants that lack both primary fiber and recovered fiber stacking areas, with production mainly focused on lightweight paper products, primarily tissue paper (Supplementary Fig. S2e).
2.2. Data sources
In this research, we obtained the geographic coordinates of 720 PPPs located in China, which accounted for over 90% of the country’s paper production capacity in 2022, from the Baidu Maps Development Platform (https://lbsyun.baidu.com/). We collected remote-sensing images of the plants from Google Earth and sourced information on the main products they manufactured from the China Paper Industry Yearbook [38] and the official websites of the relevant companies. We extracted carbon emis- sions data from the plants' environmental, social, and governance (ESG) reports and solar resource data (specifically for annual total solar radiation) from the Resource and Environmental Science Data Platform (https://www.resdc.cn/Default.aspx) using a spatial resolution of 1 km. The detailed data for this study are available on Zenodo, an open-access repository for research data and code (https://zenodo.org/records/16629379).
2.4. Multimodal data fusion
2.4.1. DeepLabv3 + model The DeepLabv3 + model adopts an encoder – decoder architec- ture (Fig. 3) [39]. The encoder employs ResNet-101 as the back- bone, comprising five blocks (Blocks 0 – 4) with 1, 3, 4, 23, and 3 residual blocks, respectively. For each input image, the backbone yields two feature streams: (i) shallow features from Block 1 (edge, texture, and color), which are fed directly into the decoder; and (ii) deep features from Block 4 that encode object categories and overall structure. The atrous spatial pyramid pooling (ASPP) module — four dilated convolution blocks and one pooling layer — further processed the deep features and generated five distinct feature outputs. These maps were then concatenated and passed through a 1 × 1 convolution layer before being fed into the decoder.
2.3. Data preprocessing
To eliminate interference from surrounding buildings, we first delineated plant boundaries based on manual visual interpreta- tion. We then converted the remote-sensing images from red – green – blue (RGB) color space to the hue – saturation – value
Fig. 2. Examples of contour extraction for pulping and papermaking plants with different geometric shapes. Original remote-sensing images are converted to the hue – saturation – value (HSV) color space, followed by plant boundary extraction to produce images of pulping and papermaking plants.
4
Made with FlippingBook interactive PDF creator