TL;DR
This study demonstrates the feasibility of using Sentinel-2 satellite imagery and Google Earth Engine to map individual cover crop species (cereal rye and red clover) in southeastern Michigan, achieving moderate to high accuracies. This advancement is crucial for understanding the environmental benefits of diverse cropping systems, as different cover crop species provide distinct ecosystem services.
What Problem Does This Paper Address?
Cover crops are a vital agricultural practice for improving soil health, crop yields, and reducing nutrient runoff. However, understanding their adoption rates and environmental impacts at large scales is challenging due to the infrequent and coarse resolution of traditional agricultural census data. While remote sensing has been used to map the presence of cover crops, identifying specific species has remained a significant gap. This is important because different cover crop species, belonging to different plant functional types (e.g., legumes vs. grasses), offer unique ecosystem services. Existing datasets like the USDA Cropland Data Layer have shown low accuracies for individual cover crop species, highlighting the need for more precise mapping methods.
Methods
The researchers employed a random forest classifier to map cover crop species. They utilized Sentinel-2 satellite imagery, known for its moderate spatial resolution and spectral bands, including red-edge bands which are sensitive to vegetation properties. The analysis was conducted within the Google Earth Engine platform, enabling efficient processing of large volumes of satellite data.
Key methodological aspects include:
- Feature Engineering: A suite of vegetation indices and individual spectral bands were calculated from Sentinel-2 imagery. These were chosen for their ability to capture differences in leaf area index, plant senescence, and chlorophyll content, as well as potential spectral differences due to flowering (e.g., red clover).
- Data Selection: Eleven Sentinel-2 images from April to August 2018 were selected to capture the phenological cycles of cover crops and cash crops.
- Ground Truth Data: Extensive field surveys were conducted in the River Raisin watershed to collect ground truth data on land cover types, including fallow fields, grass/pastures, winter wheat, cereal rye, and red clover. This data was used for both training and validation of the random forest model.
- Variable Importance Analysis: The study analyzed the importance of different spectral bands, vegetation indices, and image acquisition dates for classification accuracy, providing insights into what features are most discriminative.
Datasets
- Sentinel-2 Satellite Imagery: Level-1C TOA (Top-of-Atmosphere) data from April 4th to August 11th, 2018, covering the River Raisin watershed.
- Ground Truth Data: Manually collected field data from 461 fields in the River Raisin watershed, classified into five categories: fallow fields, grass/pastures, winter wheat, cereal rye, and red clover. This dataset was split into training (70%) and validation (30%) sets.
- USDA NASS Cropland Data Layer (CDL): Used as a reference for cultivated land in the study region to guide the selection of fields for ground surveys.
Key Insights and Ideas
- Sentinel-2's Potential for Species-Level Mapping: The study demonstrates that Sentinel-2, with its red-edge bands and moderate resolution, is capable of differentiating individual cover crop species, a significant improvement over mapping only the presence of cover crops.
- Importance of Red-Edge Bands: Red-edge bands were found to be particularly important for classification, highlighting their sensitivity to vegetation characteristics that can distinguish different species.
- Timing is Crucial: Specific time points in the growing season are critical for accurate classification. Early April (for cereal rye establishment) and early July (for red clover biomass and winter wheat senescence) emerged as key periods.
- Phenology and Spectral Signatures Matter: Differences in the phenological cycles (growth, flowering, senescence) and spectral signatures (especially for flowering species like red clover) are the underlying drivers that enable species differentiation.
- Misclassification Drivers: The study identified that inter-seeding cover crops into existing crop stubble or planting them into residue leads to spectral similarities with fallow fields, contributing to misclassifications.
Key Results
- Overall Accuracy: The random forest model achieved an overall accuracy of 83.01%.
- Cover Crop Accuracy:
- Red Clover: Achieved an F1 score of 0.77, with user's accuracy of 86.24% and producer's accuracy of 69.29%.
- Cereal Rye: Achieved an F1 score of 0.70, with user's accuracy of 78.08% and producer's accuracy of 63.33%.
- Cash Crop Accuracy: Wheat and fallow fields were classified with higher accuracies (F1 scores of 0.93 and 0.87, respectively).
- Most Important Predictors: The normalized difference tillage index (NDTI) from April 20th was the most important predictor. Red-edge bands (B5, B6, B7, B8A) were also highly influential.
- Critical Time Periods: Images from early July (July 9th) and late April (April 20th-30th) were identified as the most important for classification.
Limitations and Future Work
- Limited Sample Size for Cover Crops: The study acknowledges a relatively small sample size for cover cropped fields, particularly cereal rye (n=30), due to their lower adoption rates (4-8%) in the region. This may have impacted classification accuracy.
- Generalizability: The findings are specific to the River Raisin watershed. Further research is needed to assess the generalizability of these methods to broader regions within Michigan, across the Midwest, and in other cropping systems globally.
- Heterogeneity: Cover crop management and performance can be heterogeneous, which is expected to increase across larger study areas.
- Future Directions:
- Investigate if classification accuracies improve with larger sample sizes of cover cropped fields.
- Explore the generalizability of the model to different regions and cropping systems.
- Examine how these mapping techniques can be used to better characterize diversification and link specific cover crop species to expected environmental benefits.
Summary updated