In modern agriculture, plastic mulch film is widely used for its benefits in soil temperature retention, moisture conservation, and weed suppression. Accurate spatial extraction of mulch film is critical for crop monitoring, yield estimation, and environmental pollution control. The rapid development of China drone technology has provided a flexible and high-resolution data source for precision agriculture. In this study, we propose an object-oriented method that combines multi-scale segmentation optimized by a segmentation evaluation index, feature space optimization, and random forest classification to extract mulch film from China drone RGB imagery. We first preprocessed the raw China drone data (acquired with a Sony ILCE-A7RII camera) to generate an orthomosaic with 5 cm ground resolution. The study area is a typical agricultural region in northern China dominated by winter wheat, garlic (mulched in winter), and scattered settlements. We systematically evaluated segmentation scales, selected 35 optimal features, and achieved a user accuracy of 92.59% for mulch film and an independent Kappa coefficient of 89.65%. The results demonstrate that the proposed method effectively identifies both large, continuous mulch fields and fragmented, damaged patches. However, in complex land‑cover zones, some misclassifications occur due to similar spectral responses. This study showcases the great potential of combining China drone imagery with advanced object‑oriented analysis for agricultural monitoring.
1. Introduction
Mulch film plays an indispensable role in boosting crop yields by modifying the micro‑climate around plants. However, the extensive use of plastic film also leads to soil pollution and landscape degradation. Accurate mapping of mulch film extent is essential for managing agricultural residues and planning environmental remediation. Satellite remote sensing has long been used for this purpose, but its spatial resolution is often insufficient to capture fragmented film patches, especially in smallholder farming systems. China drone (unmanned aerial vehicle, UAV) technology offers a cost‑effective and timely alternative with ultra‑high spatial resolution (sub‑decimeter level). In recent years, many researchers have explored China drone data for land‑cover classification, yet extracting mulch film remains challenging due to its spectral similarity with bare soil, roads, and artificial surfaces, especially in RGB imagery lacking near‑infrared bands.
Existing methods can be grouped into pixel‑based and object‑oriented approaches. Pixel‑based classification often suffers from the “salt‑and‑pepper” effect and low accuracy when using only RGB bands. Object‑oriented image analysis (OBIA) overcomes this by segmenting the image into meaningful objects and then classifying them based on spectral, textural, shape, and contextual features. Among classifiers, random forest (RF) has demonstrated robust performance in remote sensing tasks. However, a critical step in OBIA is selecting the optimal segmentation scale, which determines the quality of objects. Many studies rely on visual interpretation to choose scale parameters, which is subjective and time‑consuming. In this work, we introduce a quantitative segmentation evaluation index that combines area‑weighted local variance and Moran’s I to objectively determine the best segmentation scale. Furthermore, we employ a feature selection tool (FSO) to reduce redundancy and improve classification efficiency. The primary objective is to develop a reliable and automated workflow for mulch film extraction using China drone imagery, and to evaluate its performance in a typical agricultural setting.
2. Study Area and Data
2.1 Study Area
The study area is located in a county within the North China Plain, characterized by a temperate continental monsoon climate with an average annual temperature of 13.3 °C and precipitation of about 550 mm. The dominant winter crops are winter wheat and garlic, the latter being mulched with plastic film from December to early spring. The terrain is flat and the fields are regularly shaped, interspersed with rural settlements, roads, and fallow lands. This region represents a typical scenario for China drone‑based mulch monitoring.
2.2 Data Acquisition and Preprocessing
The China drone (model: Southern SkySurvey MF2500) equipped with a Sony ILCE-A7RII camera was used to collect imagery on December 24, 2020, during 11:00–14:00 local time under clear sky. The flight altitude was 387 m, with 80% forward overlap and 60% side overlap, yielding a ground sample distance of 5 cm. We deployed 20 ground control points measured by RTK GPS to ensure geometric accuracy. The raw images were processed using SkyPhoto software for aerial triangulation, orthorectification, point cloud densification, and mosaic generation. A rectangular region of 6521×6493 pixels (approximately 1.06 km²) was selected as the study subset. The major land‑cover types include winter wheat, garlic mulch film, bare soil, roads, buildings, and fallow land.
3. Methodology
Our workflow consists of three main steps: (1) multi‑scale segmentation with automatic scale optimization using a segmentation quality index (SI); (2) feature extraction and selection using the FSO tool; and (3) random forest classification and accuracy assessment. All processing was implemented in eCognition Developer 9.0, Python, and ArcGIS 10.6.
3.1 Multi‑Scale Segmentation and Scale Optimization
We applied the multi‑resolution segmentation algorithm in eCognition to partition the China drone imagery into homogeneous objects. The shape and compactness parameters were fixed at 0.5. Scales were tested from 50 to 350 with initial steps of 50, then refined with steps of 10 around the optimal value. For each segmentation result, we computed the area‑weighted local variance (V) and Moran’s I (M) to evaluate internal homogeneity and inter‑object heterogeneity, respectively.
Area‑weighted local variance for a single band is defined as:
$$
V = \frac{\sum_{i=1}^{n} (a_i \cdot v_i)}{\sum_{i=1}^{n} a_i}
$$
where \( a_i \) is the area of object \( i \), \( v_i \) is its variance in the band, and \( n \) is the number of objects. A smaller V indicates higher internal homogeneity.
Moran’s I measures spatial autocorrelation between adjacent objects:
$$
M = \frac{n \sum_{i=1}^{n} \sum_{j=1}^{n} w_{ij} (m_i – \bar{m})(m_j – \bar{m})}{\left( \sum_{i=1}^{n} (m_i – \bar{m})^2 \right) \left( \sum_{i \neq j} \sum w_{ij} \right)}
$$
where \( m_i \) and \( m_j \) are the mean values of objects \( i \) and \( j \) in a given band, \( \bar{m} \) is the global mean, and \( w_{ij}=1 \) if objects are adjacent, otherwise 0. A lower M indicates higher inter‑object heterogeneity (i.e., less spatial autocorrelation).
After normalizing V and M across all bands, the segmentation quality index for a multiband image is:
$$
SI = \frac{1}{k} \sum_{b=1}^{k} \left( normV_b + normM_b \right)
$$
where \( k \) is the number of bands (3 for RGB). The optimal scale corresponds to the minimum SI value. Table 1 lists the SI values for various scales tested.
| 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 |
| 160 | 2,517 | 0.7985 |
| 200 | 1,719 | 0.8189 |
| 250 | 1,226 | 0.8570 |
| 300 | 893 | 0.9519 |
| 350 | 715 | 1.0096 |
From Table 1, the minimum SI occurs at scale 130, indicating the best trade‑off between object homogeneity and heterogeneity. Visual inspection confirmed that at scale 130, agricultural fields were well‑separated without over‑segmentation, while built‑up areas were slightly over‑segmented due to fine details. This scale was used for subsequent analysis.
3.2 Feature Extraction and Selection
For each image object, we initially computed 53 candidate features covering four categories: spectral (mean and standard deviation of R, G, B), index (ExG, VDVI, NBRDI, NGBDI, NGRDI), geometric (area, shape index, extent, compactness, etc.), and texture (GLCM homogeneity, contrast, entropy, etc. from 24 directions). Using the Feature Space Optimization (FSO) tool in eCognition, we evaluated the separability of mulch film versus other classes for all possible feature subset sizes. The optimal subset size was determined when the minimum separation distance (Jeffries‑Matusita distance) peaked. With 35 features, the minimum separation distance between mulch film and the second most similar class reached 5.48. The selected feature set comprised 10 spectral, 4 index, 8 geometric, and 13 texture features.
3.3 Random Forest Classification
Random forest (Breiman, 2001) is an ensemble of decision trees built on bootstrap samples and random feature subsets. We set the number of trees to 100 (a value found sufficient through preliminary tests) and the number of features per split to the square root of the total selected features (≈6). Training samples for six land‑cover classes (mulch film, winter wheat, bare soil, road, fallow land, building) were selected from the segmented objects via manual interpretation, totaling about 120 objects per class. The classifier was trained and then applied to the entire study area. Accuracy was assessed using 95 independent random points validated by visual interpretation of the original China drone orthophoto.
4. Results and Discussion
4.1 Classification Results
The final land‑cover map produced by our object‑oriented random forest method is shown conceptually (the actual image cannot be displayed here). The mulch film class was extracted with high completeness for both large continuous fields and small fragmented patches. Even torn or damaged film trapped in soil was identified, as well as white plastic covers over stacked materials. The boundaries of mulch film objects were crisp and consistent with field edges. However, in the southeastern corner of the study area, where tree shadows, glass roofs, and light‑colored roads create spectral confusion, some road segments and high grass were misclassified as mulch film.
4.2 Accuracy Assessment
The confusion matrix (Table 2) summarizes the classification accuracy. The overall accuracy was 78.95%, but the user’s accuracy for mulch film reached 92.59%, indicating that when a pixel was classified as mulch film, it had a 92.59% chance of being true mulch film. The producer’s accuracy for mulch film was 89.29% (not shown in table but derived from the matrix: 25 correctly classified out of 28 reference mulch film points). The Kappa coefficient for the mulch film class was 89.65%, demonstrating strong agreement beyond chance. Misclassifications mainly occurred between mulch film and bare soil or grass due to spectral overlap and the effect of shadows.
| Classified \ Reference | Mulch film | Ridge | Wheat | Road | Fallow | Grass | Building |
|---|---|---|---|---|---|---|---|
| Mulch film | 25 | 1 | 0 | 0 | 0 | 1 | 0 |
| 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 |
4.3 Comparison with Previous Studies
Compared with pixel‑based approaches, our object‑oriented method significantly reduces salt‑and‑pepper noise and improves classification consistency. The use of an objective segmentation quality index (SI) removes subjectivity in choosing the scale parameter, a common limitation in many object‑oriented studies. The user accuracy of 92.59% for mulch film is competitive with recent studies using China drone data and deep learning, but with much lower training sample requirements. Nevertheless, the overall accuracy of 78.95% is modest because non‑mulch classes (especially fallow land and grass) were relatively less accurate. This is partly due to the limited spectral information in RGB imagery and the similar appearance of dry vegetation and soil. Future work could incorporate multi‑temporal China drone data or integrate near‑infrared bands (e.g., by mounting a multispectral sensor) to improve separability.
5. Conclusion
This study developed an efficient and objective workflow for mulch film extraction from China drone RGB imagery. By combining a segmentation evaluation index (SI) for optimal scale selection, feature space optimization, and random forest classification, we achieved high accuracy for mulch film identification. The method successfully captured both large and damaged patches, demonstrating its practical value for agricultural management and environmental monitoring. The use of China drone platforms enables timely and high‑resolution data acquisition, which is crucial for precision agriculture. Although some misclassifications occur in complex areas, the proposed approach provides a robust baseline for automated mulch film mapping. Future research should explore multi‑temporal China drone imagery, deep learning models, and the integration of additional spectral bands to further improve classification performance.
Keywords: China drone; mulch film extraction; object-oriented; segmentation scale; random forest