In this study, cotton bolls were extracted from UAV imagery acquired in two years using three object-based algorithms, and the accuracy and visual performance of each method are summarized in Table 4 and Figure 5. Both mahalanobis distance and maximum likelihood achieved satisfactory boll extraction, with overall accuracies exceeding 95% and Kappa coefficients above 0.90, indicating strong agreement. As shown in Figure 5, parallelepiped performed poorly in delineating boll boundaries, resulting in markedly lower overall accuracies than mahalanobis distance and maximum likelihood (77.79% and 88.90%, respectively), together with substantially lower Kappa values (0.2808 and 0.7007). Differences in the underlying principles of these object-based algorithms led to distinct extraction outcomes. Overall, both mahalanobis distance and maximum likelihood provided reliable boll extraction, with overall accuracies above 96% and Kappa coefficients greater than 0.91.

SBW estimation models were developed using the selected features in combination with ridge regression, random forest regression, and neural network regression, and the results are shown in Figure 9. Models built on features derived from maximum likelihood-based boll extraction achieved higher accuracy than those based on mahalanobis distance and parallelepiped (R² = 0.68-0.80; RMSE = 0.39-0.31 g). Across the three regression approaches, on the calibration set, the nonlinear models (random forest regression, and neural network regression) yielded higher R² values (0.90-0.96 and 0.82-0.90, respectively) than the linear ridge regression model (0.73-0.86). On the validation set, model performance decreased for all methods, yet the nonlinear models maintained their advantage. Considering the overall performance on both the calibration and validation sets, the SBW estimation model combining maximum likelihood-based boll extraction with neural network regression achieved the best performance, with a validation R² of 0.80 and an RMSE of 0.31 g.

SBW was spatially mapped across the cotton field using the optimal modeling strategy that combines maximum likelihood-based boll masking with neural network regression, with the results shown in Figure 10. The results from both 2023 and 2024 indicate that this workflow can reliably estimate the spatial distribution of SBW, and predicted values were generally close to the measured values with relatively low errors (Figure 10a). In addition, the zoomed-in inset maps (Figure 10b) highlight pixel-level predictions for representative low- and high-density plots and provide a direct comparison between masked (object-based) and unmasked (canopy-based) estimation. Boll masking removed 74.72% of background pixels in the representative areas, demonstrating its ability to suppress background interference before aggregating pixel-level predictions to plot-level values for breeding evaluation.

The robustness of the maximum likelihood-neural network regression-SBW model was further confirmed by the results shown in Figure 11. Overall, the model exhibited strong robustness, with estimated values closely matching the measured values. A slight overestimation was observed in the low-value range, whereas a slight underestimation occurred in the high-value range. Overall, the maximum likelihood-neural network regression-SBW model achieved accurate SBW estimation. Notably, the relative errors across different combinations of years, varieties, and planting densities were all below 15%, further demonstrating the robustness of the maximum likelihood-neural network regression-SBW model.

Object-based approaches are well suited for delineating morphologically complex, small-scale targets in high-resolution imagery (Blaschke, 2010). Consistent with this, our results showed that all three object-based supervised classification algorithms effectively separated cotton bolls from background components. Among them, maximum likelihood achieved the highest overall accuracy (≥96.6455%) and Kappa coefficient (≥0.9147) across the two growing seasons. The Mahalanobis distance classifier demonstrated higher sensitivity to the distribution and quantity of the training samples provided. In contrast, the parallelepiped classifier, due to its reliance on simple, axis-aligned decision rules, proved particularly prone to misclassifying pixels in spectrally ambiguous zones such as boll boundaries and shaded areas.

Importantly, the quality of segmentation is not only reflected by classification accuracy metrics but also by the strength of subsequent spectral-trait relationships. Vegetation indices extracted using the maximum likelihood-based mask generally exhibited higher correlations with SBW and higher model importance than those derived from the other two algorithms. This indicates that high-quality boll extraction can effectively improve the “purity” of spectral signals by reducing interference from soil and residual leaves, thereby providing more reliable inputs for SBW retrieval. Previous studies on deep-learning-based boll detection and boll-index construction have likewise shown that extraction quality is a prerequisite for accurate yield estimation (Sun et al., 2020; Li et al., 2022; Reddy et al., 2024), and our findings provide further evidence for this conclusion from the perspective of object-based classification.

Building on boll extraction, this study identified SAVI, OSAVI and their red-edge variants, as well as RDVI and RERDVI, as core predictors for SBW estimation from multiple categories of vegetation indices. These indices share two key characteristics: (i) they are constructed using the red, red-edge, and near-infrared bands, and (ii) they generally incorporate soil-adjustment terms or re-normalized formulations. This is consistent with the spectral environment of cotton fields after boll opening. On the one hand, cotton lint exhibits high brightness in the visible bands, and some wavelengths are easily confounded with exposed soil; therefore, “greenness-type” indices such as NDVI alone are insufficient to characterize cotton bolls effectively. On the other hand, defoliation increases canopy porosity and the proportion of exposed soil, thereby amplifying the influence of soil brightness variability on spectral signals. Soil-adjusted vegetation indices mitigate this effect by introducing correction factors, whereas red-edge-related indices are more sensitive to residual green tissues and structural changes around the fiber (Huete, 1988; Rondeaux et al., 1996).

In this study, the Pearson correlations analysis and SHAP-based importance rankings were highly consistent, suggesting that vegetation indices incorporating red-edge information and designed to reduce soil background effects are not only strongly associated with SBW in univariate analyses but also contribute robustly in multivariate models. Compared with studies that estimate cotton yield or SBW from canopy-averaged spectra (Ma et al., 2022), our approach derives spectral predictors from boll pixels identified by object-based extraction. This design provides a closer correspondence between remote-sensing signals and yield components and supports a more physically interpretable relationship between spectral features and SBW.

Using the selected features, we further evaluated the potential of different machine-learning algorithms for SBW estimation. Overall, nonlinear models outperformed the linear model, indicating a certain degree of nonlinearity between boll-scale spectral signals and SBW, which is consistent with previous findings on estimating cotton and other crop yields using UAV imagery and machine learning (Ashapure et al., 2020). Among the tested approaches, the neural network model built on maximum likelihood-based boll extraction achieved a favorable balance between the calibration and validation sets, with a validation R² of 0.80 and an RMSE of 0.31 g. Moreover, the relative errors remained below 15% across different years, varieties, and planting densities, demonstrating good robustness under the experimental conditions.

Notably, error analysis revealed a systematic pattern: slight overestimation at low SBW values and some underestimation at high SBW values. This pattern is more than a simple “regression-to-the-mean” statistic; it reflects fundamental constraints in spectral-trait modeling. We attribute this non-linear bias to two intertwined factors. First, at the high-SBW end, the observed underestimation is likely driven by spectral saturation effects, where the sensitivity of broadband vegetation indices diminishes as biomass or yield increases, compressing the dynamic range of the spectral response (Ashapure et al., 2020). Concurrently, occlusion and shading within dense canopies may exacerbate signal saturation. Second, at the low-SBW end, the overestimation may stem from increased interference from background soil and senescent material in mixed pixels, coupled with a potential data imbalance where extreme low-value samples are underrepresented in the training set. Together, these physical and statistical factors cause the model’s predictions to be biased toward the mean, particularly at the trait distribution tails. Future research may address this issue by employing less-saturating hyperspectral or red-edge indices, extending the saturation point through fusion of multimodal data, and implementing strategic oversampling for extreme phenotypes.

From an application perspective, SBW spatial maps can be used together with boll-number or yield maps. In breeding trials, such joint interpretation can help distinguish genotypes characterized by “many bolls with low SBW” from those with “few bolls but high SBW.” In production management, SBW maps can provide complementary information for optimizing planting density, guiding defoliant application, and determining harvest timing, thereby extending existing yield-estimation frameworks that rely on canopy height or fractional cover (Niu et al., 2023; Huang et al., 2025).

It should be noted that this study has several limitations. First, the experiment was conducted at a single site in northern Xinjiang, where soil type and management practices were relatively consistent. Although multiple varieties and planting densities were included, the transferability of the proposed model across different agro-ecological zones and cultivation systems remains to be validated. Yield estimation models are often affected by factors such as variety type, mulching practices, and management level when applied across regions (Niu et al., 2023; Ma et al., 2022). Second, the current object-based segmentation relies on manually labeled training samples and is also dependent on image spatial resolution and radiometric calibration quality. To scale up to large production fields, it may be necessary to adopt deep-learning frameworks such as instance segmentation to reduce annotation costs and improve adaptability under complex backgrounds and small-target scenarios (Li et al., 2022; Reddy et al., 2024). In addition, the regression models in this study were built solely on spectral features, without integrating structural information derived from digital elevation models or 3D point clouds, or agronomic covariates such as meteorological and soil variables. Multi-source data fusion has been shown to improve the robustness and cross-year generalization of crop yield estimation models (Ashapure et al., 2020).

Therefore, future work should pursue two complementary directions to enhance model generalizability. First, expanding sampling campaigns across multiple ecological regions is essential to build a more representative and diversified training dataset. Second, on the methodological front, it is crucial to explore techniques such as transfer learning or domain adaptation. These approaches could potentially leverage the model knowledge gained from data-rich source environments (like the current study site) to improve performance in new target regions with limited ground truth data, thereby reducing the dependency on large-scale labeled samples from every new zone. Ultimately, integrating these efforts within a framework of “high-quality segmentation + multi-source feature fusion” will support the joint retrieval of SBW and other yield-component traits. This will provide more comprehensive phenotypic information to guide precision cotton production and breeding decisions.