Coal extraction, while fundamental to global energy and industrial production, frequently leads to the degradation of cultivated land, especially in regions where mining activities overlap with crucial grain-producing areas. Land reclamation practices are implemented to mitigate this impact and restore agricultural productivity. The ability to conduct early and accurate predictions of crop yield on reclaimed lands is therefore vital. It enables precision field management, supports the evaluation of reclamation effectiveness, and contributes to broader food security goals. Traditional yield measurement methods are often inefficient and lack the spatial granularity needed for nuanced management. In this context, remote sensing technology offers a powerful alternative. Among various platforms, unmanned aerial vehicle (UAV) remote sensing has emerged as a particularly effective tool. UAV drones provide unparalleled flexibility, allowing for the on-demand acquisition of imagery with high spatial, spectral, and temporal resolution. This capability is essential for monitoring the heterogeneous and often challenging environments of reclaimed mine sites.
Recent advancements in sensors mounted on UAV drones, such as multispectral and RGB cameras, facilitate the extraction of diverse features related to crop health and vigor. Vegetation indices (VIs), derived from mathematical combinations of spectral band reflectances, are classic proxies for biophysical parameters like chlorophyll content and biomass. Additionally, texture indices (TIs), which quantify the spatial variation and patterns within an image, can capture information about canopy structure and health that spectral indices alone may miss. The integration of these multi-modal features has shown promise in improving the robustness of crop parameter estimation. Concurrently, machine learning algorithms have surpassed traditional linear regression models in handling complex, non-linear relationships between remote sensing features and agronomic traits like yield.
Despite these advancements, research specifically targeting yield estimation for crops grown on reclaimed coal mine dumps, particularly in ecologically sensitive regions like the black soil zone, remains limited. The reconstructed soils in these areas often exhibit significant spatial variability in physical and chemical properties, posing a unique challenge for remote sensing models. This study addresses this gap. We focus on soybean, a major crop, cultivated on a reclaimed open-pit coal mine dump in the black soil region. Utilizing data collected by UAV drones, we systematically evaluate the potential of different feature sets—VIs, TIs, and their combination (VTI)—for yield estimation. We further compare the performance of four distinct modeling approaches: Multiple Linear Regression (MLR), Random Forest Regression (RFR), Back-Propagation Neural Network (BPNN), and Support Vector Regression (SVR). The overarching goal is to identify the most reliable methodology for rapid and accurate yield prediction in such reclaimed environments, thereby providing technical support for precision agriculture and reclamation assessment.
Materials and Methodology
Study Area and Field Data
The study was conducted on a reclaimed external dump of an open-pit coal mine located within the black soil region of Northeast China. This area is characterized by a continental monsoon climate with cold, dry winters and warm, rainy summers. Following the completion of dumping to the design elevation, comprehensive reclamation measures, including soil reconstruction and amendment, were undertaken to restore agricultural utility. Soybean (a semi-dwarf, dense-planting variety) was sown in late May using a standard ridging technique. The field was managed under consistent agronomic practices throughout the growing season.
At physiological maturity, yield data were collected to serve as ground truth for model calibration and validation. To facilitate this, the experimental field was divided into 20 plots. Each plot was harvested separately. The soybeans were threshed, and the grain weight was measured. The yield for each plot was then calculated as grain weight per unit area (kg/m²). This plot-based sampling strategy provided a spatially distributed dataset linking precise geographic locations to measured yield values.
UAV Data Acquisition and Processing
Remote sensing data were acquired during the grain-filling stage of soybean, a critical period closely linked to final yield. A DJI Matrice 210 RTK UAV drone platform was deployed. It was equipped with two primary sensors: a Yusense MS600Pro multispectral camera and a DJI Zenmuse X5S RGB camera. The flight mission was conducted under clear sky conditions between 11:00 and 14:00 local time.
The UAV drone was programmed to fly at an altitude of 100 meters above ground level, with a speed of 10 m/s. The forward and side overlaps were set at 85% and 80%, respectively, to ensure high-quality image stitching. The multispectral camera captured data in five central wavelengths: Blue (450 nm), Green (555 nm), Red (660 nm), Red Edge (710 nm), and Near-Infrared (840 nm). Prior to each flight, a calibrated reflectance panel was imaged to enable absolute reflectance calibration. Both the multispectral and RGB image sets were processed using Pix4Dmapper software. The processing workflow included image alignment, radiometric calibration using the panel data, and the generation of georeferenced orthomosaics for each spectral band and the RGB composite. The reflectance values from the multispectral orthomosaic and the digital numbers from the RGB orthomosaic were extracted for subsequent analysis.
Feature Extraction: Vegetation and Texture Indices
Two categories of remote sensing features were extracted from the processed UAV drone imagery: Vegetation Indices (VIs) and Texture Indices (TIs).
Based on a review of literature concerning soybean growth monitoring, eight VIs were selected for their known sensitivity to vegetation biomass, chlorophyll content, and photosynthetic activity. These indices are listed in Table 1.
| Vegetation Index | Abbreviation | Formula |
|---|---|---|
| Normalized Difference Vegetation Index | NDVI | $$NDVI = \frac{R_{NIR} – R_{Red}}{R_{NIR} + R_{Red}}$$ |
| Green Normalized Difference Vegetation Index | GNDVI | $$GNDVI = \frac{R_{NIR} – R_{Green}}{R_{NIR} + R_{Green}}$$ |
| Normalized Difference Red Edge Index | NDRE | $$NDRE = \frac{R_{NIR} – R_{Red Edge}}{R_{NIR} + R_{Red Edge}}$$ |
| Green-Red Vegetation Index | GRVI | $$GRVI = \frac{R_{Green} – R_{Red}}{R_{Green} + R_{Red}}$$ |
| Simplified Canopy Chlorophyll Content Index | SCCCI | $$SCCCI = \frac{NDRE}{NDVI}$$ |
| Enhanced Vegetation Index | EVI | $$EVI = 2.5 \times \frac{R_{NIR} – R_{Red}}{R_{NIR} + 6 \times R_{Red} – 7.5 \times R_{Blue} + 1}$$ |
| Modified Chlorophyll Absorption in Reflectance Index | MCARI | $$MCARI = [(R_{Red Edge} – R_{Red}) – 0.2 \times (R_{Red Edge} – R_{Green})] \times (\frac{R_{Red Edge}}{R_{Red}})$$ |
| Optimized Soil Adjusted Vegetation Index | OSAVI | $$OSAVI = \frac{R_{NIR} – R_{Red}}{R_{NIR} + R_{Red} + 0.16}$$ |
Texture Indices were derived from the pan-sharpened multispectral imagery and the RGB imagery using the Gray-Level Co-occurrence Matrix (GLCM) method. The GLCM calculates how often pairs of pixels with specific values and in a specified spatial relationship occur in an image. From this matrix, eight common second-order texture measures were computed: Mean (Me), Variance (Var), Homogeneity (Hom), Contrast (Con), Dissimilarity (Dis), Entropy (Ent), Second Moment (SM), and Correlation (Cor). A moving window size of 5×5 pixels was used for these calculations, which was found to be suitable for capturing canopy texture at the spatial resolution provided by the UAV drones.
Feature Selection and Yield Estimation Modeling
The extracted VIs and TIs for each of the 20 plots were averaged. To identify the most relevant features for yield prediction and reduce dimensionality, Pearson correlation coefficients (PCC) between each index and the measured plot yield were calculated. The indices were ranked based on the absolute value of their correlation with yield. To ensure a fair comparison across models, the top six indices from different feature groups were selected to form three distinct input feature sets:
- VI Set: The six VIs with the highest correlation to yield.
- TI Set: The six TIs with the highest correlation to yield.
- VTI Set: A combined set featuring the three most correlated VIs and the three most correlated TIs.
Four regression algorithms were employed to build yield estimation models using these feature sets:
- Multiple Linear Regression (MLR): A baseline statistical model that assumes a linear relationship between the input features and yield.
- Random Forest Regression (RFR): An ensemble learning method that constructs multiple decision trees during training and outputs the mean prediction of the individual trees. It is robust to overfitting.
- Back-Propagation Neural Network (BPNN): A classic artificial neural network with one hidden layer (5 neurons), trained using the error backpropagation algorithm to model complex non-linear relationships.
- Support Vector Regression (SVR): A powerful algorithm that finds the hyperplane that best fits the data within a specified margin of error (ε). A Radial Basis Function (RBF) kernel was used, with hyperparameters (C and γ) optimized via grid search.
The dataset of 20 plots was used for model training and testing. The performance of each model (MLR, RFR, BPNN, SVR) with each feature set (VI, TI, VTI) was evaluated and compared.
Model Evaluation Metrics
The accuracy and predictive capability of the yield estimation models were assessed using three standard metrics:
- Coefficient of Determination (R²):
$$R^2 = \frac{ \sum_{i=1}^{n} (x_i – \bar{x})^2 (y_i – \bar{y})^2 }{ \sum_{i=1}^{n} (x_i – \bar{x})^2 \sum_{i=1}^{n} (y_i – \bar{y})^2 }$$
where \(x_i\) is the measured yield, \(y_i\) is the predicted yield, \(\bar{x}\) and \(\bar{y}\) are their respective means, and \(n\) is the number of samples. Higher R² values indicate better model fit. - Root Mean Square Error (RMSE):
$$RMSE = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} (y_i – x_i)^2 }$$
Lower RMSE values indicate higher prediction accuracy. - Relative Percent Difference (RPD):
$$RPD = \frac{ STD_{measured} }{ RMSE }$$
where \(STD_{measured}\) is the standard deviation of the measured yield values. The RPD is used to evaluate the model’s predictive power: RPD < 1.4 indicates poor model, 1.4 ≤ RPD ≤ 2.0 indicates fair model, and RPD > 2.0 indicates an excellent model with good predictive ability.
Results and Analysis
Correlation Analysis of Features with Yield
The correlation analysis revealed strong relationships between many of the UAV-derived features and soybean grain yield on the reclaimed dump. Most VIs and TIs showed statistically significant correlations. Among the texture indices, Correlation (Cor) exhibited the highest positive correlation with yield (r ≈ 0.97), indicating that areas with more uniform and structured canopies tended to have higher yields. Variance (Var) and Entropy (Ent) also showed strong positive correlations. Among the vegetation indices, NDVI demonstrated the strongest correlation (r ≈ 0.91), followed closely by GNDVI and GRVI. Indices like OSAVI and SCCCI showed relatively lower correlations. This analysis confirmed the relevance of both spectral and textural information obtained from UAV drones for assessing yield potential. Based on the ranking, the feature sets were constructed as follows: VI Set = {NDVI, GNDVI, GRVI, MCARI, EVI, NDRE}; TI Set = {Cor, Var, Ent, SM, Me, Con}; VTI Set = {Cor, Var, Ent, GNDVI, GRVI, NDVI}.
Performance of Yield Estimation Models
The performance of the four modeling algorithms with the three different feature sets is summarized in Table 2. The results highlight clear trends regarding the superiority of certain feature combinations and algorithms for this specific application.
| Feature Set | Model | R² | RMSE (kg/m²) | RPD | Predictive Capability |
|---|---|---|---|---|---|
| Vegetation Indices (VI) | MLR | 0.51 | 0.0072 | 1.43 | Fair |
| RFR | 0.69 | 0.0055 | 1.86 | Good | |
| BPNN | 0.73 | 0.0049 | 2.10 | Excellent | |
| SVR | 0.73 | 0.0048 | 2.14 | Excellent | |
| Texture Indices (TI) | MLR | 0.58 | 0.0066 | 1.55 | Fair |
| RFR | 0.65 | 0.0060 | 1.71 | Good | |
| BPNN | 0.67 | 0.0059 | 1.74 | Good | |
| SVR | 0.80 | 0.0047 | 2.19 | Excellent | |
| Vegetation & Texture Indices (VTI) | MLR | 0.62 | 0.0063 | 1.63 | Good |
| RFR | 0.78 | 0.0049 | 2.10 | Excellent | |
| BPNN | 0.81 | 0.0045 | 2.28 | Excellent | |
| SVR | 0.84 | 0.0039 | 2.63 | Excellent |
1. Effect of Feature Set: The results consistently demonstrate that combining vegetation and texture indices (VTI set) leads to superior model performance across all non-linear algorithms (RFR, BPNN, SVR) compared to using either set alone. For instance, the SVR model’s R² increased from 0.73 (VI) and 0.80 (TI) to 0.84 (VTI), while its RMSE decreased and RPD increased correspondingly. This underscores the complementary nature of spectral and spatial/textural information captured by UAV drones. Spectral indices like NDVI are sensitive to chlorophyll and green biomass, while texture indices like Cor describe canopy architecture and homogeneity—both factors influence final yield.
2. Performance of Modeling Algorithms:
- MLR consistently yielded the lowest R² and RPD values, regardless of the feature set. Its performance was only “fair” (RPD ~1.4-1.6), indicating that the relationships between the features and yield in this complex reclaimed environment are predominantly non-linear, which simple linear regression cannot adequately capture.
- RFR and BPNN showed significant improvement over MLR. They effectively modeled the non-linearities, achieving “good” to “excellent” predictive capability. Their performance was notably enhanced when using the combined VTI feature set.
- SVR emerged as the most robust and accurate algorithm in this study. It achieved the highest R² (0.84) and RPD (2.63), and the lowest RMSE (0.0039 kg/m²) when trained on the combined VTI feature set. Its performance was also the most stable across the three different feature inputs, consistently delivering excellent results. The optimization of its kernel and penalty parameters likely contributed to its strong generalization ability and resistance to overfitting, even with a limited sample size.
3. Optimal Model: Therefore, the Support Vector Regression (SVR) model utilizing the combined Vegetation and Texture Index (VTI) feature set is identified as the optimal approach for estimating soybean yield on this reclaimed coal mine dump. This model successfully integrates the strengths of multispectral and textural data from UAV drones with a powerful machine learning algorithm.
Spatial Yield Mapping
Using the optimal SVR-VTI model, a spatial yield map was generated for the entire study area. The map revealed a clear spatial pattern in predicted yield across the reclaimed dump. Yield generally showed a decreasing gradient from the northwest to the southeast sectors and an increasing trend from south to north. This spatial variability likely reflects underlying differences in soil properties related to the reclamation process, such as topsoil thickness, compaction, or nutrient distribution—factors that are notoriously heterogeneous in reconstructed mine soils. The map produced by the UAV drone-based model aligns with the expectation of variable crop performance in such an environment and provides a valuable tool for identifying sub-areas that may require targeted management interventions in future seasons.
Discussion
The findings of this study confirm the high potential of UAV drone remote sensing for crop monitoring and yield prediction in challenging agricultural landscapes like reclaimed mine lands. The strong correlations observed between specific VIs/TIs and yield are consistent with agronomic principles and prior research in standard farmlands. However, the performance of texture indices, particularly Correlation (Cor), being the single most correlated feature, highlights an important aspect of reclaimed environments. Soil reconstruction often results in variable seedbed conditions, leading to differences in plant density, growth uniformity, and canopy closure. Texture indices from high-resolution UAV imagery are exceptionally good at quantifying this spatial heterogeneity in canopy structure, which becomes a direct visual proxy for the varying success of crop establishment and growth across the plot.
The superior performance of the combined VTI feature set over VI or TI alone is a key result. It demonstrates that for accurate yield modeling in complex, non-uniform fields, a multi-modal sensing approach is beneficial. The spectral signals provide information on the plant’s physiological status (e.g., health, chlorophyll), while the textural signals provide information on its structural status (e.g., coverage, uniformity). Machine learning algorithms, especially SVR in this case, are adept at learning the complex, synergistic relationship between these different data dimensions and the final integrative agronomic outcome—yield.
While the SVR-VTI model showed excellent predictive capability (RPD > 2.5), certain limitations of the study must be acknowledged. The sample size (n=20 plots), though adequate for a preliminary study, is relatively small. Reclaimed dumps are areas of high spatial heterogeneity, and a larger, more geographically diverse sample set would improve model robustness and generalizability. Furthermore, soil-specific parameters (e.g., organic matter, bulk density, pH) which are critical in reclaimed settings were not integrated into the model. Future work should focus on expanding the dataset across multiple seasons and sites, and fusing UAV drone data with in-situ soil measurements or multi-temporal imagery from key growth stages. Integrating such multi-source data could further enhance prediction accuracy and provide deeper insights into the soil-plant relationships governing yield on reclaimed lands.
From a practical standpoint, the methodology presented offers a rapid, non-destructive, and spatially explicit tool for land managers and reclamation authorities. Early-season or mid-season yield forecasts can guide decisions on supplemental irrigation, fertilization, or other management practices to optimize final output. Perhaps more importantly, it provides a quantitative, objective metric for evaluating the success of reclamation efforts in restoring agricultural productivity, which is a central goal of mine land restoration.
Conclusion
This research demonstrates a effective framework for estimating soybean yield on a reclaimed coal mine dump in the black soil region using UAV drone remote sensing and machine learning. The main conclusions are as follows:
- Features derived from UAV drone imagery, particularly texture indices like Correlation and vegetation indices like NDVI, exhibit strong correlations with soybean grain yield in the reclaimed environment, validating their use as yield predictors.
- The combination of vegetation indices and texture indices (VTI) as model inputs consistently leads to higher estimation accuracy compared to using either feature type alone, underscoring the value of multi-modal data fusion.
- Among the evaluated modeling algorithms, Support Vector Regression (SVR) demonstrated the best overall performance, stability, and predictive power, achieving an R² of 0.84 and an RPD of 2.63 when using the combined VTI feature set.
- The optimal SVR-VTI model enables the generation of high-resolution spatial yield maps, revealing patterns of variability that can inform precision field management and serve as a quantitative measure for assessing reclamation effectiveness.
In summary, UAV drone-based remote sensing, coupled with advanced machine learning techniques like SVR, provides a powerful, scalable, and accurate solution for monitoring crop productivity on reclaimed mine lands. This approach offers significant advantages for supporting sustainable agriculture and ensuring the long-term success of ecological restoration in post-mining landscapes.