Timely and accurate mapping of agricultural mulch film is essential for modern crop management, environmental monitoring, and pollution control in China. With the rapid advancement of China UAV (Unmanned Aerial Vehicle) technology, high-resolution remote sensing data have become increasingly accessible for precision agriculture. This study proposes an object-oriented method for extracting mulch film from China UAV RGB imagery, integrating image texture, shape, and spectral features with an optimal segmentation scale determined by a segmentation evaluation index and a random forest classifier. The method is tested in a representative agricultural area in Shandong Province, where winter wheat and garlic are the main crops. The results demonstrate that the proposed approach effectively identifies both large continuous mulch film and fragmented patches, achieving a user accuracy of 92.59% and an independent Kappa coefficient of 89.65%. The methodology provides robust support for large-scale agricultural surveys and environmental monitoring using China UAV platforms.
Keywords: China UAV; mulch film; object-oriented; segmentation scale; random forest
1. Introduction
Mulch film, widely used in China’s agriculture, offers significant benefits such as soil temperature regulation, moisture retention, and weed suppression, thereby enhancing crop yields. However, the extensive use of plastic mulch also leads to severe residual pollution, affecting soil structure and crop growth. Accurate extraction of mulch-covered areas is therefore a critical task for both agricultural management and environmental protection. Remote sensing, especially with high-resolution imagery, provides an efficient means for this purpose.
Previous studies have utilized satellite imagery (e.g., Landsat TM, OLI, GF-2, WorldView-2) to map plastic mulch and greenhouses, employing vegetation indices, spectral analysis, and machine learning methods. However, satellite data often suffer from limited spatial resolution and revisit frequency, making them less suitable for fine-scale monitoring. In contrast, China UAV platforms offer ultra-high spatial resolution (centimeter-level), flexible scheduling, and low operational cost, making them ideal for localized agricultural surveys. The adoption of China UAV technology in agricultural remote sensing has increased rapidly in recent years.
Existing methods for mulch film extraction from China UAV imagery can be broadly categorized into pixel-based and object-oriented approaches. Pixel-based classification suffers from the “salt-and-pepper” effect and poor performance when only RGB bands are available. Object-oriented methods, which group pixels into meaningful objects before classification, have shown superior performance but often rely on subjective visual interpretation for selecting segmentation parameters. Deep learning approaches require large training datasets and substantial computational resources. Therefore, there is a need for a robust, efficient, and automated method that can leverage China UAV RGB imagery to accurately extract mulch film.
In this work, we propose a systematic framework that combines a quantitative segmentation evaluation index (SI) for optimal scale selection, feature selection via Feature Space Optimization (FSO), and a random forest (RF) classifier. The method is validated in a typical agricultural region in Shandong Province, China. The contributions of this study are: (1) an objective selection of segmentation scale based on the weighted local variance and Moran’s I index; (2) a comprehensive feature set combining spectral, textural, geometric, and index features; (3) a demonstration of the efficacy of the object-oriented RF approach for mulch film extraction from China UAV imagery.
2. Study Area and Data
The study area is located in a typical agricultural county in Shandong Province, eastern China. The region belongs to the temperate continental monsoon climate, with an average annual temperature of 13.3°C and annual precipitation of 550 mm. The main crops include winter wheat, maize, soybean, fruit trees, cotton, and garlic. During the winter season (December), garlic fields are covered with plastic mulch film to protect against low temperatures. The landscape is a flat alluvial plain, making it suitable for UAV flight operations.
China UAV imagery was acquired on December 24, 2020, using a fixed-wing China UAV platform (Southern Tianxun MF2500) equipped with a Sony ILCE-A7RII camera. The flight altitude was 387 m, and the ground sample distance (GSD) was 5 cm. The flight was conducted between 11:00 and 14:00 local time, with forward overlap of 80% and side overlap of 60%. A total of 6,521 by 6,493 pixels covering approximately 1.06 km² were selected for this study. The imagery was preprocessed using SkyPhoto software for aerial triangulation, orthorectification, point cloud densification, and mosaic stitching. Ground control points (GCPs) acquired by RTK GPS were used to ensure geometric accuracy. The final orthomosaic contained three bands (R, G, B) with 8-bit radiometric resolution.
Major land-cover types in the study area included: mulched fields (garlic), winter wheat fields, bare soil (fallow land), roads, buildings, grass, and field ridges. The presence of shadow from trees and buildings, as well as reflective surfaces from plastic roofs, introduced spectral confusion with mulch film, posing challenges for accurate extraction.
3. Methodology
The proposed object-oriented framework consists of three main steps: (1) multi-scale image segmentation with optimal scale selection using a segmentation evaluation index; (2) feature extraction and selection via the FSO tool; (3) classification using a random forest classifier. Each step is described below.
3.1 Multi-scale Image Segmentation
Multi-scale segmentation is a bottom-up region-merging technique that groups pixels into homogeneous objects. The segmentation parameters include scale, shape weight, and compactness weight. We employed eCognition Developer software for segmentation. The shape and compactness parameters were both set to 0.5 based on preliminary experiments. The scale parameter was varied from 50 to 350 with an initial step of 50, and then refined with a step of 10 around the preliminary optimum.
To objectively evaluate segmentation quality, we defined a Segmentation Index (SI) that combines intra-object homogeneity and inter-object heterogeneity. The intra-object homogeneity is measured by the area-weighted local variance (V), while the inter-object heterogeneity is measured by Moran’s I (M). The formulas are as follows:
$$
V = \frac{\sum_{i=1}^{n} (a_i \cdot v_i)}{\sum_{i=1}^{n} a_i}
$$
where \(n\) is the number of objects in a single band, \(a_i\) is the area of object \(i\), and \(v_i\) is the variance of pixel values within object \(i\). A smaller \(V\) indicates higher internal homogeneity.
Moran’s I is defined as:
$$
M = \frac{n \sum_{i=1}^{n} \sum_{j=1}^{n} w_{ij}(m_i – \bar{m})(m_j – \bar{m})}{\sum_{i=1}^{n} (m_i – \bar{m})^2 \left( \sum_{i \neq j} \sum_{j} w_{ij} \right)}
$$
where \(m_i\) and \(m_j\) are the mean values of objects \(i\) and \(j\) in a given band, \(\bar{m}\) is the overall mean, and \(w_{ij}\) is the spatial weight matrix (1 if adjacent, 0 otherwise). A smaller absolute value of Moran’s I indicates lower spatial autocorrelation and thus higher heterogeneity between objects.
For each band (R, G, B), we computed \(V\) and \(M\), then normalized them to [0,1] using min-max normalization. The final evaluation index \(SI\) is the average of the normalized sum across bands:
$$
SI = \frac{\sum_{b=1}^{3} \left( \text{norm}(V_b) + \text{norm}(M_b) \right)}{3}
$$
The optimal segmentation scale corresponds to the minimum \(SI\) value, indicating the best balance between intra-object homogeneity and inter-object heterogeneity.
3.2 Feature Extraction and Selection
After segmentation, each object is characterized by a set of features. We initially considered 53 features from four categories: spectral (mean and standard deviation of R, G, B), geometric (area, length/width, shape index, etc.), texture (Gray-Level Co-occurrence Matrix, GLCM, including contrast, correlation, homogeneity, etc.), and vegetation indices (ExG, VDVI, NBRDI, NGBDI, NGRDI). The full list of candidate features is shown in Table 1.
| Feature Type | Feature Names | Count |
|---|---|---|
| Spectral | Mean (R, G, B), Std Dev (R, G, B) | 6 |
| Index | ExG, VDVI, NBRDI, NGBDI, NGRDI | 5 |
| Geometric | Area, Length, Width, Length/Width, Shape Index, Compactness, Border Index, etc. | 13 |
| Texture (GLCM) | Homogeneity, Contrast, Dissimilarity, Entropy, Angular Second Moment, Correlation, Mean, Std Dev (8 x 3 bands) | 24 |
| Other | Brightness, Max diff | 5 |
To reduce redundancy and improve classification efficiency, we applied the Feature Space Optimization (FSO) tool in eCognition. This tool uses a separability measure (Jeffries-Matusita distance) between classes to evaluate feature subsets. Starting from an empty set, features are added iteratively until the minimum class separability does not improve significantly. The optimal number of features was found to be 35, which yielded a minimum separability distance of 5.48 between mulch film and the most similar class. The selected features included 10 spectral, 4 index, 8 geometric, and 13 texture features.
3.3 Random Forest Classification
Random Forest (RF) is an ensemble learning method that constructs multiple decision trees using bootstrap samples and random feature subsets. It aggregates the predictions by majority voting. The RF classifier is robust to noise, handles high-dimensional data, and provides feature importance. We implemented RF using the eCognition classification engine. The number of trees was set to 100, and the number of features per split was set to the square root of the total number of selected features (i.e., about 6). Training samples were selected manually for six land-cover classes: mulch film, winter wheat, fallow land, roads, buildings, grass, and field ridges. Approximately 50–100 object samples per class were chosen, ensuring representative coverage of spectral and spatial variability.
4. Results and Discussion
4.1 Optimal Segmentation Scale
The SI values for different segmentation scales are listed in Table 2. As the scale increases from 50 to 130, the SI decreases, indicating improved segmentation quality. After 130, the SI begins to increase, reflecting over-merging. The minimum SI value (0.7597) is achieved at a scale of 130, which is therefore selected as the optimal scale. Visual inspection confirms that the segmented objects at scale 130 delineate agricultural fields well, with clear boundaries and minimal under- or over-segmentation in cropland areas. However, in the southwestern part of the study area (residential zones with shadows and complex roofs), some over-segmentation persists due to the high heterogeneity of built-up land.
| Scale | Number of Objects | SI |
|---|---|---|
| 50 | 26,719 | 1.0000 |
| 100 | 6,167 | 0.8214 |
| 120 | 4,321 | 0.7640 |
| 130 | 3,727 | 0.7597 |
| 140 | 3,160 | 0.7785 |
| 150 | 2,804 | 0.7869 |
| 200 | 1,719 | 0.8189 |
| 250 | 1,226 | 0.8570 |
| 300 | 893 | 0.9519 |
| 350 | 715 | 1.0096 |
4.2 Classification Results and Accuracy Assessment
Using the optimal segmentation (scale 130) and the selected 35 features, the RF classifier was applied to produce a land-cover map. The mulch film class was successfully identified across the entire study area. Visual comparison with the orthomosaic shows that large contiguous mulch films, as well as small fragmented patches, are well captured. Figure 1 presents the segmentation overlay and classification results.
We generated 95 random validation points using ArcGIS and visually interpreted their true land-cover types by referencing the high-resolution orthoimage and field knowledge. A confusion matrix was constructed, as shown in Table 3. The overall accuracy (OA) for all classes is 78.95%, while the user accuracy for mulch film reaches 92.59% and the producer accuracy is 86.21%. The Kappa coefficient for the mulch film class is 89.65%.
Errors for mulch film mainly occurred at object boundaries where shadows from adjacent trees or buildings caused spectral mixing. Two validation points of mulch film were misclassified: one as field ridge and one as grass. These points were located at the edges of mulched plots. Additionally, some bright roofs of plastic sheds and concrete roads were occasionally confused with mulch film, especially in the high-reflectance regions. However, the overall performance demonstrates that the object-oriented RF method using China UAV imagery is highly effective for mulch film extraction.
| Reference \ Classified | Mulch film | Field ridge | Wheat | Road | Fallow | Grass | Building |
|---|---|---|---|---|---|---|---|
| Mulch film | 25 | 1 | 0 | 0 | 0 | 1 | 0 |
| Field ridge | 2 | 5 | 1 | 1 | 0 | 0 | 0 |
| Wheat | 0 | 1 | 22 | 0 | 0 | 0 | 0 |
| Road | 0 | 0 | 0 | 5 | 0 | 0 | 0 |
| Fallow | 0 | 3 | 4 | 2 | 11 | 0 | 1 |
| Grass | 0 | 0 | 0 | 1 | 1 | 4 | 0 |
Figure 1: Example of China UAV orthomosaic and corresponding classification result. The red areas represent mulch film. The high spatial resolution (5 cm) enables detailed mapping of plastic-covered garlic fields and damaged film fragments.
4.3 Discussion
The proposed method demonstrates several advantages for China UAV-based mulch film extraction. First, the quantitative segmentation evaluation index (SI) eliminates subjective bias in selecting the scale parameter. The optimal scale of 130 is suitable for the object sizes of agricultural fields in our study area. Second, the FSO-based feature selection reduces the dimensionality from 53 to 35, improving computational efficiency without sacrificing accuracy. Third, the random forest classifier provides robust performance even with limited training samples.
However, some limitations remain. The method is sensitive to the acquisition time: in our case, the December imagery captured garlic mulch film with high contrast against bare soil and senescent vegetation. In other seasons, when green vegetation is abundant, spectral separability may decrease. Additionally, the presence of shadows, white plastic roofs, and concrete surfaces can cause false positives. Future work could incorporate temporal information (multi-temporal China UAV imagery) or additional spectral bands (e.g., near-infrared if available) to mitigate these issues. Moreover, the segmentation scale is optimized for the entire scene, but local variations in object size (e.g., narrow field ridges versus large fields) might benefit from adaptive segmentation scales.
5. Conclusion
This study developed an object-oriented mulch film extraction method using China UAV RGB imagery. By integrating a segmentation evaluation index for optimal scale selection, feature space optimization for feature reduction, and a random forest classifier, the proposed framework achieved a user accuracy of 92.59% and a Kappa coefficient of 89.65% for mulch film. The method successfully mapped both continuous and fragmented mulch film in a winter agricultural landscape. The results confirm that China UAV platforms, combined with advanced object-based image analysis, provide a cost-effective and accurate solution for agricultural plastic waste monitoring and crop management.
The approach can be readily applied to other regions with similar crop types and landscape characteristics. It also holds potential for integration into operational monitoring systems that require rapid, high-resolution mapping of mulch film distribution. Future research should focus on improving robustness in complex environments, exploring deep learning alternatives with limited training data, and extending the method to multi-temporal China UAV datasets for change detection of plastic mulch.