Integrating image segmentation and auxiliary data for efficient estimation of FVC and AGB
Abstract:: Accurate estimation of fractional vegetation cover (FVC) and aboveground biomass (AGB) is essential for large-scale grassland monitoring. However, this process is often constrained by the labor-intensive nature of field surveys. In this study, we propose a fully non-destructive framework...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772375525004642 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract:: Accurate estimation of fractional vegetation cover (FVC) and aboveground biomass (AGB) is essential for large-scale grassland monitoring. However, this process is often constrained by the labor-intensive nature of field surveys. In this study, we propose a fully non-destructive framework that integrates deep learning, image segmentation, vegetation indices, and structural auxiliary variables to retrieve FVC and AGB in meadow grasslands. By combining a U-Net network with convex-hull masking, we automatically delineate 1 m × 1 m quadrats from 574 smartphone RGB images collected in western Jilin, China. This approach improves the mean intersection-over-union (mIoU) from 61.8 % (using raw U-Net bounding boxes) to 90.1 % (with final quadrat masks). We evaluate five vegetation indices and identify the color index of vegetation extraction (CIVE) as the most robust. CIVE achieves correlation coefficients of 0.85 for FVC and 0.54 for AGB in fresh grass plots. To address spectral confusion between dry grass and soil—particularly in images where dry grass covers >10 %—we introduce a “segment-then-reclassify” pipeline. This pipeline incorporates TurboPixels superpixels, edge-guided watershed segmentation, and k-means clustering for improved discrimination. Two predictor sets are used to train five inversion models (ridge regression, k-nearest neighbors, support vector regression, random forest, and partial least squares regression): one using only the vegetation pixel ratio, and the other combining the vegetation pixel ratio with vegetation density and mean height. Modeling the FVC of fresh and dry grass separately improves performance, increasing the R2 by an average of 0.23 and reducing the root mean square error (RMSE) by up to 41 %. Structural variables play a key role in AGB estimation, improving R2 by up to 0.28 and decreasing RMSE by 17 %. The proposed framework, based solely on lightweight RGB imagery and low-cost computation, enables real-time and non-destructive data acquisition in the field and offers reliable support for multi-scale grassland remote sensing and ecological assessments. |
|---|---|
| ISSN: | 2772-3755 |