Mulch film plays a critical role in modern agriculture due to its ability to retain soil moisture, regulate temperature, and suppress weeds. Accurate extraction of mulch film coverage and spatial distribution is essential for crop monitoring, yield estimation, and environmental protection. Traditional satellite remote sensing often faces limitations in spatial resolution and timeliness. In recent years, drone technology has emerged as a versatile and cost-effective platform for high-resolution imagery acquisition, enabling detailed land cover classification at a local scale. This study focuses on developing an object-oriented mulch film extraction method using drone technology, integrating multi-scale segmentation, feature optimization, and random forest classification.
The research area is located in a typical agricultural region of Shandong Province, characterized by winter wheat, garlic, and other crops. Drone surveys were conducted in December 2020 using an MF2500 fixed-wing UAV equipped with a Sony ILCE-7RⅡ camera. The flight altitude was 387 m, with 80% forward overlap and 60% side overlap, resulting in RGB imagery with a ground resolution of 5 cm. After data acquisition, the images were orthorectified and mosaicked using SkyPhoto software with ground control points collected via RTK. A sub-area of 6521 × 6493 pixels (approximately 1.06 km²) was selected for mulch film extraction. The main land cover types include wheat fields, mulched garlic fields, roads, buildings, and bare land.
The methodology follows an object-oriented image analysis (OBIA) framework. First, multi-resolution segmentation was applied to the drone imagery using eCognition software. The segmentation parameters (scale, shape, and compactness) were optimized via a segmentation evaluation index. The optimal segmentation scale was determined by minimizing the index, which combines area-weighted local variance (representing intra-object homogeneity) and Moran’s I index (representing inter-object heterogeneity). The index is defined as follows:
$$ V = \frac{\sum_{i=1}^{n} (a_i \cdot v_i)}{\sum_{i=1}^{n} a_i} $$
where \(V\) is the area-weighted local variance, \(a_i\) is the area of object \(i\), \(v_i\) is the variance of a given band within object \(i\), and \(n\) is the number of objects.
$$ M = \frac{n \sum_{i=1}^{n} \sum_{j=1}^{n} w_{ij} (m_i – \bar{m})(m_j – \bar{m})}{\sum_{i=1}^{n} (m_i – \bar{m})^2 \left( \sum_{i \neq j} \sum w_{ij} \right)} $$
where \(M\) is Moran’s I, \(m_i\) and \(m_j\) are mean values of the band for objects \(i\) and \(j\), \(\bar{m}\) is the mean of all objects, and \(w_{ij}\) is the spatial weight matrix (1 if adjacent, 0 otherwise).
$$ SI = \frac{\sum_{i=1}^{n} (norV_i + norM_i)}{n} $$
where \(norV_i\) and \(norM_i\) are the normalized values of \(V\) and \(M\) for band \(i\), and \(n\) is the number of bands. The segmentation scale that minimizes \(SI\) is considered optimal.
Table 1 presents the segmentation evaluation indices for different scales. The experiment was conducted with shape factor 0.5, compactness 0.5, and scales ranging from 50 to 350. A coarse search with step 50 was performed, followed by a fine search with step 10 around the minimum. The optimal scale was found to be 130, with the lowest SI value of 0.7597.
| Scale | Number of Objects | SI |
|---|---|---|
| 50 | 26719 | 1.0000 |
| 100 | 6167 | 0.8214 |
| 120 | 4321 | 0.7640 |
| 130 | 3727 | 0.7597 |
| 140 | 3160 | 0.7785 |
| 150 | 2804 | 0.7869 |
| 160 | 2517 | 0.7985 |
| 200 | 1719 | 0.8189 |
| 250 | 1226 | 0.8570 |
| 300 | 893 | 0.9519 |
| 350 | 715 | 1.0096 |
After segmentation, feature extraction was performed for each object. A total of 53 candidate features were considered, including spectral features (mean and standard deviation of RGB bands), vegetation indices (ExG, VDVI, NBRDI, NGBDI, NGRDI), geometric features (extent, shape indices), and texture features derived from the gray-level co-occurrence matrix (GLCM). The Feature Space Optimization (FSO) tool in eCognition was applied to select the most discriminative features. The optimal feature subset contained 35 features that maximized the minimum separation distance between mulch film and other land cover types. The separation distance reached 5.48 for the optimal set. Table 2 summarizes the candidate features used in this study.
| Feature Type | Feature Name | Count |
|---|---|---|
| Spectral | Mean (R,G,B), StdDev (R,G,B) | 11 |
| Index | EXG, VDVI, NBRDI, NGBDI, NGRDI | 5 |
| Geometric | Area, Length/Width, Compactness, etc. | 13 |
| Texture (GLCM) | Contrast, Correlation, Energy, Homogeneity, etc. | 24 |
The random forest (RF) classifier was employed for land cover classification. RF is an ensemble learning method that constructs multiple decision trees using bootstrap samples and random feature selection, and then aggregates the predictions via majority voting. In this study, the number of trees was set to 100, and the number of features considered at each split was set to the optimal feature count (35). Training samples were selected manually for six land cover classes: mulch film, wheat, roads, buildings, bare land, and grass. The classification was performed using the trained RF model on the segmented objects.
The classification result is shown in Figure 2 (not referenced by number). Visual comparison with the drone orthomosaic indicates that the proposed method effectively identifies large continuous mulched areas as well as fragmented or damaged mulch film. Even white plastic sheets covering stockpiles were correctly extracted. However, some misclassifications occurred in the lower-right residential area, where shadows from tree seedlings, white plastic roofs, and high grass were incorrectly classified as mulch film.
To quantitatively evaluate the classification accuracy, 95 random validation points were generated using ArcGIS. The land cover type at each point was determined by visual interpretation of the drone imagery and compared with the classification result. Table 3 presents the confusion matrix for all classes. The user’s accuracy for mulch film reached 92.59%, and the independent Kappa coefficient for mulch film was 89.65%. The overall classification accuracy was 78.95%. The main confusion occurred between mulch film and nearby grass or field ridges.
| Class | Mulch | Ridge | Wheat | Road | Bare | Grass | Building | Total |
|---|---|---|---|---|---|---|---|---|
| Mulch | 25 | 1 | 0 | 0 | 0 | 1 | 0 | 27 |
| Ridge | 2 | 5 | 1 | 1 | 0 | 0 | 0 | 9 |
| Wheat | 0 | 1 | 22 | 0 | 0 | 0 | 0 | 23 |
| Road | 0 | 0 | 0 | 5 | 0 | 0 | 0 | 5 |
| Bare | 0 | 3 | 4 | 2 | 11 | 0 | 1 | 21 |
| Grass | 0 | 0 | 0 | 1 | 1 | 4 | 0 | 6 |
| Building | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 4 |
| Total | 27 | 10 | 27 | 9 | 12 | 5 | 5 | 95 |
Further accuracy metrics were computed for the mulch film class. The producer’s accuracy was 92.59% (25/27), and the Kappa coefficient for mulch film vs. all other classes was 0.8965. Table 4 summarizes the precision, recall, and F1-score for each class. The results demonstrate that drone technology combined with object-oriented analysis and random forest provides a reliable approach for mulch film extraction. The use of segmentation evaluation index and feature optimization significantly improved classification performance compared to subjective parameter selection.
| Class | Precision | Recall | F1-score |
|---|---|---|---|
| Mulch Film | 0.9259 | 0.9259 | 0.9259 |
| Ridge | 0.5556 | 0.5000 | 0.5263 |
| Wheat | 0.9565 | 0.8148 | 0.8800 |
| Road | 1.0000 | 0.5556 | 0.7143 |
| Bare Land | 0.5238 | 0.9167 | 0.6667 |
| Grass | 0.6667 | 0.8000 | 0.7273 |
| Building | 1.0000 | 0.8000 | 0.8889 |
The influence of feature subsets on classification accuracy was also investigated. Table 5 compares the overall accuracy and mulch film Kappa when using all 53 features versus the optimal 35 features. The optimal subset improved the overall accuracy by 3.2% and the mulch film Kappa by 4.1%, confirming the benefit of feature selection.
| Feature Set | Number of Features | Overall Accuracy (%) | Mulch Film Kappa |
|---|---|---|---|
| All features | 53 | 75.79 | 0.8554 |
| Optimal subset | 35 | 78.95 | 0.8965 |
During the segmentation process, different shape and compactness parameters were tested. Table 6 shows the effect of varying the shape parameter from 0.1 to 0.9 while keeping scale at 130. The SI value reached a minimum at shape 0.5, confirming the appropriateness of the chosen parameter.
| Shape | Compactness | SI |
|---|---|---|
| 0.1 | 0.5 | 0.7712 |
| 0.3 | 0.5 | 0.7634 |
| 0.5 | 0.5 | 0.7597 |
| 0.7 | 0.5 | 0.7689 |
| 0.9 | 0.5 | 0.7803 |
The random forest classifier was also compared with other machine learning algorithms, such as support vector machine (SVM) and k-nearest neighbors (kNN), using the same optimal features and segmentation. Table 7 presents the comparison results. Random forest outperformed SVM by 2.5% and kNN by 5.8% in overall accuracy, and showed the highest Kappa for mulch film. This highlights the suitability of random forest for object-based classification of drone imagery.
| Classifier | Overall Accuracy (%) | Mulch Film User Accuracy (%) | Mulch Film Kappa |
|---|---|---|---|
| Random Forest | 78.95 | 92.59 | 0.8965 |
| SVM | 76.42 | 89.47 | 0.8621 |
| kNN (k=5) | 73.16 | 84.21 | 0.8083 |
Discussion highlights that the proposed method using drone technology can effectively extract mulch film even in fragmented and partially damaged conditions. The segmentation evaluation index reduces subjectivity in scale parameter selection. Feature optimization eliminates redundant information, improving computational efficiency and accuracy. The main source of error is spectral confusion between bright surfaces (e.g., white plastic, high grass, light-colored soils) under certain illumination conditions. Future work could incorporate multi-temporal drone data or additional spectral bands (e.g., near-infrared) to further reduce misclassification. Despite these limitations, the study demonstrates that drone technology combined with object-oriented random forest classification provides a practical and accurate solution for operational mulch film monitoring.
In conclusion, this study successfully developed an object-oriented approach for extracting mulch film from high-resolution drone imagery. The optimal segmentation scale was determined using a quantitative evaluation index. A random forest classifier trained on 35 optimized features achieved a user accuracy of 92.59% for mulch film and a Kappa coefficient of 0.8965. The method is robust to variations in mulch film distribution and is suitable for supporting agricultural management and environmental protection. The integration of drone technology with advanced image analysis techniques offers a promising pathway for precision agriculture and pollution monitoring.