Search Results (245)

The Probability Hypothesis Density (PHD) filter based on the Track-Before-Detect (TBD) approach is a key technique for detecting weak targets whose numbers are unknown and time-varying. To overcome the limitations of existing algorithms, such as high computational cost, poor real-time performance, and low tracking efficiency in dense clutter, this paper proposes a fast non-sequential particle PHD filter for TBD. Specifically, an adaptive particle generation method based on differential localization is introduced in the prediction stage, allowing newly generated particles to quickly concentrate around potential target locations. In the update stage, particles are divided into three groups to simplify weight calculation and improve efficiency. Furthermore, a parallel resampling strategy is adopted to further enhance real-time performance. Numerical experiments demonstrate that the proposed method maintains tracking accuracy with only a small number of particles, thereby significantly reducing computational complexity and improving real-time capability. This work offers a practical reference for the engineering deployment of TBD algorithms. Full article

Predictive quality systems are central to Industry 4.0 manufacturing, yet detecting rare defects that pass established quality gates remains an open problem. This study addresses post-quality-gate failure detection in automotive brake manufacturing, where 310 faulty units (1.20%) among 25,756 production records create a naturally occurring extreme class imbalance of 1:82. Fault labels are derived from warranty reports and linked to multi-station production line measurements, while negative samples may include latent failures, motivating a recall-focused evaluation. We propose a Quality 4.0 machine learning framework that compares five resampling methods (ADASYN, SMOTE-Tomek, KMeans-SMOTE, CTGAN, and TVAE) plus a no-resampling baseline across 24 classifiers and stacking ensembles. In total, 504 configurations are tested on a held-out test set. The proposed SVM-RBF model trained on ADASYN-augmented data achieves recall of 0.871, specificity of 0.982, balanced accuracy of 0.926, and ROC-AUC of 0.952, producing only 93 false positives (FPR = 1.8%). Stacking ensembles provide alternative operating points maximizing the detection rate (93.5%) and a separate operating point with the highest discrimination capacity (ROC-AUC = 0.975). Feature importance analysis through Permutation Importance and SHAP identifies Force Increment as the leading feature under both attribution methods. Friedman and Wilcoxon tests confirm statistically significant differences among strategies. The framework offers a practical way to add predictive capability to existing quality control systems. Full article

Greenhouse microclimate regulation is challenging due to nonlinear coupling among temperature, humidity, soil moisture, and light intensity, which limits the effectiveness of conventional threshold-based and PID control strategies under time-varying environmental disturbances. This paper presents a forecast-augmented ensemble control framework that combines Random Forest, Gradient Boosting, and Support Vector Machine classifiers with one-hour-ahead weather forecasts for closed-loop greenhouse microclimate regulation. The proposed system was deployed and validated in a working greenhouse cultivating cucumber (cv. ‘Madora F1’) over 28 consecutive days. Sensor measurements and forecast inputs were processed through a unified preprocessing pipeline, while control actions were generated through majority voting and executed on Raspberry Pi 4B edge hardware with a worst-case inference latency below 18 ms. The proposed framework achieved a temperature RMSE of 0.83 °C during field deployment. For reference, RMSE values of 3.21 °C and 1.94 °C were obtained for the threshold-based and PID baseline controllers, respectively, under the adopted disturbance-consistent evaluation protocol. Compliance rates reached 96.4% for temperature, 94.1% for relative humidity, and 97.2% for soil moisture across 40,320 resampled observation intervals (60 s analysis grid) derived from the original 10 s acquisition stream. Integration of short-term weather forecasts enabled anticipatory irrigation management, reducing irrigation pump operation by 18% without compromising soil-moisture compliance and yielding an estimated annual energy saving of 158 kWh per greenhouse zone. Unlike prediction-oriented greenhouse artificial-intelligence studies, the proposed approach implements a deployable forecast-augmented closed-loop control architecture validated under continuous real-world greenhouse operation. Full article

Green prepared dishes are an emerging food-consumption format that links convenience, food safety, and sustainable consumption. In this study, “green” denotes a sustainability-oriented product profile involving food-safety assurance, resource-conscious packaging or sourcing, and waste-reduction potential, rather than formal organic certification. However, existing studies have mainly relied on linear behavioral models and have paid limited attention to nonlinear and asymmetric consumer decision mechanisms. This study integrates the stimulus–organism–response framework with explainable machine learning to predict consumers’ sustainable purchase intention toward green prepared dishes. Based on 805 valid questionnaires collected in Jilin Province, China, predictors were organized into three dimensions: environmental and health cognition, socioeconomic and infrastructural conditions, and sustainable behavioral propensity. The sample represents a regional online consumer profile in Jilin Province rather than a national probability sample. Six classifiers were trained using SMOTE–Tomek resampling and Optuna-based hyperparameter optimization. XGBoost achieved the best predictive performance, with an F1-score of 0.894, an AUC of 0.934, and an MCC of 0.702. Unlike conventional black-box machine learning, the SHAP-based interpretation translated ensemble predictions into transparent feature-level and case-level explanations. Accordingly, the model interpretations are framed as predictive associations rather than causal mechanisms. The study reveals an asymmetric decision pattern in which core behavioral willingness functions as a non-compensatory barrier, while channel convenience, delivery efficiency, and after-sales support facilitate purchase intention among consumers who already show high behavioral readiness. The findings suggest that green prepared-dish strategies should prioritize trust-based advocacy and word-of-mouth, reliable channel design, low-risk trial experiences, and collaborative food-safety governance rather than relying only on short-term traffic acquisition. Full article

Introduction: The most common method for controlling invasive fishes is mechanical removal, although it is time-consuming and operationally demanding. In Portugal, commercial inland fisheries are permitted, but the extent to which they could represent a complementary tool for invasive species control remains unknown. Objective: We compared fish assemblages in two sections of a branch of the Cedillo Reservoir where commercial fishing is allowed (i.e., downstream the Lentiscais bridge) and where it is prohibited under the regulations of the International Tagus Natural Park (i.e., upstream). Methodology: Fish were sampled in 2023 and 2024, using a total of 116 gillnets: 72 downstream and 44 upstream. The proportion of blank nets between sections was compared using Fisher’s exact test. Variations in community composition were assessed using NMDS and PERMANOVA, while homogeneity of multivariate dispersion was evaluated with PERMDISP. Total CPUE was compared between sections using Mann–Whitney tests. At the species level, CPUE were assessed using permutation-based Mann–Whitney U tests adjusted for multiple comparisons using the FDR procedure. The direction and magnitude of between-section differences were quantified using the rank-biserial correlation, and 95% confidence intervals were estimated by section-stratified bootstrap resampling. Results: In total, 20 gillnets yielded no fish, and there was no significant difference in their proportions between sections. Total CPUE per gillnet was significantly higher upstream than in the commercially fished section. Fish assemblage composition differed significantly between sections, and there were no dissimilarities in multivariate dispersion, indicating a genuine, although partial, separation between assemblages. The species contributing most to the dissimilarity between the two sections were Silurus glanis, Cyprinus carpio, Sander lucioperca, and Luciobarbus bocagei, which are also among the main target species for national inland commercial fisheries. Among these species, L. bocagei and S. glanis showed significantly lower CPUE in the commercially fished section, while S. lucioperca showed higher CPUE. Variations in C. carpio CPUE were barely significant. Conclusions: These results suggest that commercial fishing may at least partially influence fish catches and assemblage structure in this reservoir branch. Future studies should partition the influence of commercial fishing from other drivers of assemblage variation to further evaluate whether it may represent a complementary tool for managing fish invasions under specific management strategies. Full article

Understanding the structural characteristics of Pinus taiwanensis plantations in climatically transitional regions is essential for developing science-based management strategies under global change. This study investigated 23 plots in Huangbai Mountain Forest Farm, Henan Province, China, classified into low-, medium-, and high-density stands (n = 9, 9, and 5, respectively). Diameter distributions were fitted using six probability functions, and four spatial structure parameters—mixing degree (Mc), size ratio (U), uniform angle index (W), and forest layer index (S)—were quantified. In addition, five comprehensive spatial structure indices—average superiority coefficient index (SPV), spatial structure comprehensive index (Q), stand spatial structure distance index (FSI), Comprehensive Distance Evaluation (CDEV), and Comprehensive Assessment of Proximity Vector (CAPV)—were constructed using a combined analytic hierarchy process and entropy weight method. Given the unbalanced sample sizes, non-parametric Kruskal–Wallis tests were employed for comparisons, and bootstrap resampling (1000 iterations) was performed to assess the reliability of mean estimates. The results showed that both the Gamma and Weibull distributions were equally suitable for describing diameter distribution under different stand densities, as their AIC differences were below 2 for all density classes. Correlation analysis indicated that the relative importance of spatial parameters followed the order S > U > Mc > W. Medium-density stands exhibited the most optimal spatial structure, whereas low-density stands showed the poorest performance. These findings suggest that both overly dense and sparse stands negatively affect spatial organization. Appropriate management practices, such as thinning or enrichment planting, are recommended to optimize stand structure and enhance ecological resilience. Full article

Background/Objective: Long-term tenofovir disoproxil fumarate (TDF) therapy is associated with progressive bone mineral density loss in patients with chronic hepatitis B (CHB), yet existing fracture risk algorithms, such as FRAX, were not designed for this population. We aimed to develop and internally validate a clinical nomogram for identifying TDF-associated osteoporosis using penalized regression on demographic, virological, and biochemical predictors. Methods: In this single-center retrospective cohort study, 237 adult CHB patients receiving TDF for at least 12 months underwent dual-energy X-ray absorptiometry (DXA). Osteoporosis was defined as a T-score of −2.5 or lower at the lumbar spine or femoral neck. Thirteen candidate predictors were evaluated using LASSO regression with 10-fold cross-validation; selected variables were entered into an unpenalized multivariable logistic regression model; internal validation employed bootstrap resampling with 200 replications to derive optimism-corrected estimates of discrimination and calibration. The clinical utility was assessed using decision curve analysis (DCA). Results: Osteoporosis prevalence was 15.2% (n = 36). LASSO selected three predictors: prior fragility fracture (OR 11.45, 95% CI 4.82–27.15), the Charlson Comorbidity Index (OR 1.45 per unit, 95% CI 1.15–1.85), and alkaline phosphatase. The model demonstrated strong discrimination (apparent C-index 0.860; optimism-corrected 0.845) with excellent calibration (slope 0.94, intercept 0.02; Brier score 0.095). At a 0.15 probability threshold, sensitivity was 86.0%, specificity 78.0%, and negative predictive value 97.0%. DCA confirmed superior net clinical benefit over default strategies across the 0.10–0.30 threshold range; a pre-specified sensitivity analysis excluding fracture history retained meaningful discrimination (corrected C-index 0.791). Conclusions: This nomogram offers a clinically actionable, disease-specific tool for stratifying osteoporosis risk in TDF-treated CHB patients, particularly well suited for safely deferring DXA imaging in low-risk individuals. External validation in multicenter and ethnically diverse cohorts is required before widespread implementation. Full article

Automated visual inspection of photovoltaic (PV) cells is an important component of solar-module quality assurance. However, adapting modern pre-trained vision backbones to PV defect classification remains challenging because full fine-tuning requires substantial memory, naturally imbalanced datasets can reduce sensitivity to rare defect classes, and edge-oriented inspection workflows impose computational constraints. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA), have been widely studied in natural language processing, but their use for PV defect classification remains underexplored. This study presents a controlled benchmark of LoRA and QLoRA against full fine-tuning for PV cell defect classification. Four adaptation strategies—full fine-tuning, LoRA with rank 8, LoRA with rank 16, and 4-bit QLoRA with rank 16—are evaluated using a ConvNeXt-Tiny backbone on a 17,377-image polycrystalline PV cell electroluminescence dataset referred to as POLY, covering five classes: intact, cracked, broken, surface-diffuse, and surface-point. The natural 6.7× class imbalance is preserved without synthetic resampling, and a group-aware StratifiedGroupKFold protocol based on available cell or panel-image identifiers is used to reduce identifiable leakage across folds. All PEFT variants slightly outperform full fine-tuning in macro-F1 while training 26–52× fewer parameters. QLoRA_r16 achieves the highest macro-F1 score of 79.92 ± 0.75%, compared with 78.26 ± 0.94% for full fine-tuning, while training the same number of parameters as LoRA_r16 (1.060 M; 3.67% of the adapted model). QLoRA_r16 also improves F1 on the intact (+4.75 points) and surface-diffuse (+2.62 points) classes relative to full fine-tuning. This class-wise pattern suggests that quantized low-rank adaptation may influence minority and visually ambiguous categories; however, the present experiments do not isolate the independent effect of NF4 quantization from adapter rank, batch size, or optimization dynamics. Under the training configuration used, QLoRA_r16 records the lowest observed peak training GPU memory, approximately 30% below full fine-tuning (1727 MB versus 2478 MB). Because QLoRA_r16 was trained with batch size 16 whereas the other methods used batch size 32, this reduction should be interpreted as an end-to-end configuration effect rather than as the isolated effect of 4-bit quantization. Overall, the results indicate that PEFT is a promising and resource-efficient alternative to full fine-tuning for PV defect classification, although batch-matched memory experiments, direct embedded-device profiling, and cross-dataset validation remain necessary before making deployment-level claims. Full article

Although the traditional Gmapping algorithm incorporates optimized proposal distribution and resampling strategies within the RBPF-SLAM framework, it remains susceptible to particle degradation during intensive particle iterations. This degradation compromises map integrity and localization accuracy. To address this limitation, this study proposes an enhanced Gmapping system integrated with an improved particle swarm optimization (PSO) algorithm. The proposed PSO incorporates an adaptive inertia weight and a Gaussian distribution model to guide swarm dynamics, thereby effectively accelerating convergence. Furthermore, during the resampling phase, the system adopts an SDPR strategy to reduce computational complexity, shorten runtime, and alleviate particle degradation. The improved PSO algorithm was first validated through MATLAB R2022b simulations, and the integrated system was subsequently implemented and tested on a ROS-based Unmanned ground vehicle (UGV) platform within the Gazebo simulation environment (Gazebo Garden). From the results, compared with classical Gmapping using 50 particles, the proposed method using 50 particles reduces the ATE RMSE from 0.154 m to 0.104 m, corresponding to a 32.5% reduction. The RPE translation RMSE decreases by 31.0%, and the map-scale MAE decreases by 44.6%. The average time per frame is also slightly lower than Gmapping-50 because SDPR reduces the frequency and cost of full resampling. Experimental results demonstrate that the proposed system yields significant improvements in both accuracy and robustness for localization and environmental mapping. Full article

Accurate prediction of loan defaults is essential for financial institutions seeking to minimize credit losses and maintain portfolio stability. In the vehicle financing segment of emerging markets, real-world datasets frequently exhibit extreme class imbalance ratios that far exceed those encountered in standard benchmark corpora, posing severe challenges for conventional machine learning pipelines. This study introduces a gradient boosting framework integrating Weight of Evidence (WoE) transformation, Bayesian hyperparameter optimization, and three complementary classifiers—Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost)—to predict vehicle loan default risk. The methodology is evaluated on a large-scale, fully anonymized Turkish vehicle loan dataset ($N=207,572$) with an extreme imbalance ratio of 1:1133 (183 defaults versus 207,389 non-defaults). A strict three-way data partition (60% training, 20% validation, 20% test) is adopted to ensure leakage-free model selection and unbiased performance estimation. A multi-stage experimental pipeline is developed encompassing: (i) statistical feature selection via Mann–Whitney U and chi-square tests with adaptive thresholding, (ii) a comparative analysis of seven resampling strategies including Synthetic Minority Oversampling Technique (SMOTE) variants, Adaptive Synthetic Sampling (ADASYN), and focal loss weighting, (iii) a greedy forward selection ensemble procedure for heterogeneous model fusion, and (iv) a systematic training-set size sensitivity analysis across eight majority undersampling ratios. Under the leakage-free evaluation protocol, the highest-AUC individual model (LightGBM with SMOTE-ENN) achieves an Area Under the Curve (AUC) Receiver Operating Characteristic (ROC) of 0.710 (95% bootstrap CI: 0.614–0.798), while CatBoost with cost-sensitive weighting exhibits superior operational metrics (KS $=0.389$, PR-AUC $=0.011$). The greedy ensemble procedure exhibits high selection instability with only 37 validation-set positives, providing a methodological finding on the minimum sample requirements for reliable ensemble construction under extreme scarcity. Ablation results confirm that WoE encoding contributes 3.1 percentage points to the overall AUC gain. Tree SHAP-based interpretability analysis identifies the financing-to-age ratio, WoE-encoded occupation group, and log financing amount as the primary predictive drivers, with cross-model stability confirmed via Spearman rank correlation. A decision support analysis provides precision–recall curves, a Brier score of 0.0082, reliability diagrams, and threshold-dependent performance at operationally plausible review rates. Fairness evaluation across gender and marital status subgroups demonstrates that threshold-dependent metrics such as Disparate Impact Ratio and Equalized Odds Gap are inherently compromised under extreme minority scarcity, whereas rank-based subgroup AUC analysis with bootstrap 95% confidence intervals preserves meaningful discriminative assessment. These findings provide an empirically validated framework for credit default prediction in highly imbalanced and data-scarce financial environments. Full article

Radio Frequency Identification (RFID) is an emerging technology in Industry 4.0 for low-cost logistics, yet direction and orientation estimation typically requires multiple antennas, and robustness under industrial multipath fading, operator variability, and signal fragmentation has not been evaluated. To address this gap, this study proposes a single-antenna RFID system that evaluated thirteen architectures spanning unsupervised methods (clustering algorithms) and supervised methods (classical machine learning, deep learning, and hybrid architectures) on Received Signal Strength Indicator (RSSI) and phase time-series reconstructed through a pipeline of Savitzky–Golay smoothing, phase unwrapping, and cubic spline resampling to $N$ = 50–300 samples, preserving signal morphology across variable-length RFID passes. The system further incorporates a physics-informed augmentation strategy that encodes multipath fading, distance variation, and fragmentation into synthetic training samples for cross-domain generalization without hardware modification. In controlled laboratory experiments, both direction and orientation tasks achieved $>$99.5% accuracy, while direction tracking was additionally validated on an industrial shop floor under varying distances, Non-Line-of-Sight (NLoS) occlusions, and signal fragmentation. Zero-shot transfer caused accuracy to degrade to near-chance levels for several configurations, confirming a pronounced domain gap. Domain adaptation with XGBoost recovered direction accuracy to $>$97% under severe fragmentation under NLoS conditions, with an inference latency of ≈150 $\mu$s. Under domain-adapted shop floor conditions, direction accuracy exceeded the 75–92% reported in prior single-antenna laboratory studies, suggesting that physics-informed domain adaptation is a promising approach for single-antenna RFID tracking in Industrial Internet of Things (IIoT) logistics environments. Full article

Background/Objectives: Prolonged air leak (PAL) remains a common complication after lung resection and may delay postoperative recovery and subsequent treatment. This study aimed to identify clinical factors associated with PAL after robot-assisted thoracic surgery (RATS) and to explore potential nonlinear relationships using restricted cubic spline (RCS) modeling. Methods: A retrospective cohort of 1185 patients who underwent RATS for primary lung cancer was analyzed. Multivariable Firth logistic regression was used to identify independent predictors of PAL (≥5 days). A nomogram was constructed based on the final model and internally validated using 1000 bootstrap resamples; its clinical utility was assessed using decision curve analysis. RCS analysis was performed to evaluate potential nonlinear associations. Results: A total of 98 patients (8.3%) developed PAL. Male sex was independently associated with increased PAL risk (OR 3.29, p < 0.001), whereas higher FEV1 was associated with reduced risk (OR 0.50 per 1-L increase, p < 0.001). BMI showed a modest protective effect (OR 0.91, p = 0.01). Age was not significant in the linear model (p = 0.86), but RCS analysis demonstrated a significant nonlinear association, with increased risk at older ages. The nomogram demonstrated moderate discrimination (apparent C-statistic 0.670, optimism-corrected 0.644) and good calibration, with decision curve analysis confirming net clinical benefit over treat-all and treat-none strategies. Conclusions: Male sex and impaired pulmonary function are key predictors of PAL after RATS. Nonlinear modeling revealed complex age-related risk patterns not captured by conventional approaches. The proposed nomogram may assist in preoperative risk stratification and perioperative decision-making. Full article

This study proposes an explainable framework for ESG portfolio rebalancing by integrating carbon credit signals, technical indicators, and Transformer-inspired forecasting into a unified decision process. The investable universe consists of six ESG-themed ETFs, namely ESGU, SUSA, ICLN, TAN, KRBN, and KGRN. Carbon-related sustainability information is represented by four S&P carbon indices, including GCC, CCA, EUA, and UCITS. Within the proposed framework, Transformer, Informer, and Temporal Fusion Transformer are used to predict next-day returns, and the forecast outputs are translated into portfolio decisions through threshold filtering, Softmax-based allocation, and inertia smoothing under fixed transaction costs. The empirical results show that the proposed framework remains competitive against Equal Weight, Risk Parity, and Momentum benchmarks, although its advantage is conditional rather than uniformly dominant across all metrics. Informer delivers the strongest Sharpe ratio among the model-based strategies, while Transformer exhibits a more stable risk profile. The ablation results indicate that the smoothing mechanism has the clearest effect on turnover and allocation stability, whereas the incremental value of carbon-related inputs is most visible in Informer. The uncertainty assessment further shows that many benchmark differences are not consistently significant under repeated resampling, but the performance weakening caused by removing carbon inputs in Informer remains identifiable. The subperiod analysis shows that benchmark rules are more competitive in 2024H1, whereas model-based strategies gain relative strength in 2024H2. The explainability analysis indicates that carbon-feature contributions are concentrated more strongly in Intermediate and Carbon-Sensitive asset groups and remain weaker in Broad ESG assets; feature-level and SHAP beeswarm evidence further shows that the three architectures rely on GCC, CCA, EUA, and UCITS in different ways. These findings suggest that carbon-related sustainability signals can provide economically meaningful allocation information in selected settings when they are combined with suitable model architecture and disciplined rebalancing control, thereby supporting a competitive and explainable ESG portfolio rebalancing framework. Full article

Background/Objectives: Vaccine-preventable diseases remain a persistent public health challenge in regions characterized by structural vulnerabilities, including suboptimal vaccination coverage, socioeconomic deprivation, and limited access to healthcare. In structurally vulnerable regions, such as the South-West Romanian region, characterized by persistent vaccination gaps and recurrent outbreaks, these conditions generate a sustained public health burden that requires ongoing preventive risk management strategies. In such contexts, digital risk stratification tools may support preventive decision-making by enabling early identification of patients at increased risk of severe outcomes. This study applied machine learning techniques to routinely collected measles surveillance data from South-West Romania to identify severe disease cases and determine key predictors of severity, offering a pragmatic alternative to outbreak forecasting in a resource-constrained setting. Methods: An open epidemiological dataset of laboratory-confirmed measles cases reported by the Regional Center for Public Health Surveillance Craiova was analyzed. The dataset defined severe cases as those with pneumonia, thrombocytopenia, a hospital stay exceeding three days, or other documented complications requiring medical intervention. Random Forest (RF) and Logistic Regression (LR) classifiers were trained and compared using a 10-fold cross-validation framework across 200 resampling iterations. Model performance was assessed using accuracy, AUC-ROC, sensitivity, specificity, positive predictive value, and F1-score. Feature importance was quantified using permutation-based measures, and the highest-ranked predictors were further evaluated through chi-square tests of independence. Results: RF significantly outperformed LR in accuracy (0.84 vs. 0.82), AUC (0.87 vs. 0.80), specificity (0.87 vs. 0.84), positive predictive value (0.89 vs. 0.86), and F1-score (0.84 vs. 0.83), with p ≤ 0.001 for most metrics. Sensitivity was equivalent between models (approximately 0.81; p = 0.328). Feature importance analysis identified seven key predictors: county of residence, vaccination status, outbreak status, presence of other symptoms, occupation, cough, and conjunctivitis. All seven were significantly associated with disease severity, and six showed significant geographic variation across counties. Vâlcea County had the highest concentration of severe cases. The model was trained on a regional surveillance cohort in which symptomatic and hospitalized cases are over-represented and should be interpreted as a triage-support tool within this surveillance context rather than as a population-level severity estimator. Conclusions: Machine learning, particularly RF, can effectively identify severe measles cases using routinely collected surveillance data in settings where robust outbreak prediction is not feasible. The county of residence functioned as a composite proxy for structural determinants, including healthcare access, vaccination coverage, and socioeconomic deprivation. These findings support the use of ML-based severity classification as a pragmatic tool for clinical risk stratification and targeted public health intervention in resource-constrained environments. Full article

Forest above-ground biomass (AGB) estimation is crucial for evaluating carbon dynamics. Although optical and synthetic aperture radar (SAR) data provide complementary spectral and structural information, limitations in existing fusion approaches restrict AGB estimation accuracy. This study proposes a multi-source data fusion framework comparing two image fusion strategies—the conventional Hue-Intensity-Saturation Wavelet (HIS-Wavelet) method and a deep learning-based HIS-Non-Subsampled Shearlet Transform combined with Pulse Coupled Neural Network (HIS-NSST + PCNN) approach—for forest AGB estimation using Gaofen-1 (GF-1), Gaofen-2 (GF-2), and Gaofen-3 (GF-3) satellite imagery in a subtropical forest area of Yunnan Province, China. Three regression models—Multiple Linear Stepwise Regression (MLSR), K-Nearest Neighbor (KNN), and KNN with Fast Iterative Feature Selection (KNN-FIFS)—were systematically compared to evaluate estimation performance and justify model selection. Results indicate that the HIS-NSST + PCNN method outperforms HIS-Wavelet in fusion quality metrics, with the GF-2 Red-Near-infrared-Blue (RNB) band and GF-3 combination using HH co-polarization achieving the highest image quality. The optimal AGB retrieval was achieved with the GF-1RNB and GF-3 combination under HIS-NSST + PCNN (coefficient of determination (R²) = 0.80, root mean square error (RMSE) = 14.79 t/ha), improving R² by 0.07 and RMSE by 2.35 t/ha over HIS-Wavelet. However, for GF-2 + GF-3, HIS-Wavelet achieved marginally better inversion accuracy (R² = 0.71) than HIS-NSST + PCNN (R² = 0.69), indicating that superior fusion quality does not directly translate to higher inversion accuracy. Bootstrap resampling analysis (1000 iterations) confirmed the statistical robustness, with the optimal KNN-FIFS yielding R² = 0.800 (95% confidence interval (CI): 0.678–0.924) and RMSE = 14.79 t/ha (95% CI: 12.51–17.22 t/ha), showing non-overlapping confidence intervals with both benchmark models. These findings demonstrate that spectral complementarity between optical and SAR data plays a more critical role than spatial resolution alone in fusion-based AGB estimation, and that adaptive feature selection is essential for maximizing inversion accuracy. Full article