Crop Type Detection
Crop Identification
Regrow’s crop detection model detects and identifies commodity crops growing on a field, at scale. Crop detection models are regionally tuned, and support the common crops grown in each region.
Crop detection is the foundation of all Monitor’s practice detection insights. The crop and growing period are directly used to evaluate the impact of tillage activity and cover cropping practices on a field.
Predicting Crop Type with Machine Learning Models
Regrow’s CropID is a machine learning model trained to detect and identify common crops in a field. It analyzes multi-spectral Sentinel 2 satellite data throughout the growing season to determine crop type based on historical data & crop-specific spectral signatures.
The model is trained on industry-standard datasets, including government estimates of previous crops, farmer-reported data and survey data collected through government and academic institutions. By learning the reflectance values for each crop, Crop ID can predict crop types just one to two months after harvest, and in some cases, even during the growing season as the model improves.
Detecting Crop Planting and Harvest
In addition to identifying crop types, Monitor's CropID ML model also predicts the planting and harvest dates.
The model is trained on both satellite imagery and vegetation indices, such as NDVI and EVI which measure the chlorophyll content in crops, and region-specific crop calendars, which account for local farming practices. Satellite images help detect when crops emerge and begin to die (senescence) within the expected crop calendar, offering detailed timing. Using annotated NDVI time series data, the ML model can predict emergence and senescence of the crop.
It's important to note that while Monitor refers to these as planting and harvest dates for simplicity, it's really detecting the crop's emergence and senescence. So the actual planting date may differ from the detected emergence, and harvest timing may vary depending on how much green vegetation remains after harvesting.
Crop Detection Confidence
Monitor's ML Crop ID model provides a confidence score alongside each crop type prediction, indicating how likely the prediction is correct. Confidence is influenced by the quality of remote sensing imagery, which can be affected by factors like cloud cover.
Confidence values are reported on a scale of 0-100. Crop type confidence can be interpreted as a statistical representation of the likelihood of the practice. For example, a confidence score of 80 for a crop indicates that 8 out of 10 times the model will get this crop type correct. Regrow recommends using crop type predictions when model confidence is 75 or higher.
Accuracy + Model Validation
The accuracy and validation of the Crop ID model are crucial for ensuring it can classify a wide range of field conditions, from fully productive to partially managed and newly converted croplands. The model must also account for crop diversity, especially in regions with varied crop rotations, while avoiding bias in areas dominated by a few crops.
Regrow's Data Science team uses established techniques to validate predictions against ground truth data, analyzing metrics like precision, recall, F1 scores, and confusion matrices. Multiple datasets, including government and farmer-reported data, are used to ensure unbiased validation.
Tillage Detection
Regrow’s tillage detection algorithms provide a field-level assessment of the impact of tillage activity at key agronomic points of crop and field management cycles. Fields are observed and analyzed for tillage impact in the time period prior to the planting of a commodity crop, as well as the period after the harvest.
Tillage Algorithm Methodology
Fields are monitored for tillage activity at key agronomic periods: the time period before a commodity crop is planted, and after the time period following the harvest. This results in two distinct tillage practice determinations for each commodity crop cultivation cycle.
Satellite data can’t pinpoint the exact day a field is tilled, the equipment used, or soil disturbance depth due to the limited resolution (10-30 meters) and frequency (5-10 days) of satellite imagery. Cloud cover, especially in fall and winter, further restricts visibility.
Instead of tracking tillage directly, Regrow estimates residue cover using indices like the Normalized Difference Tillage Index (NDTI) and Crop Residue Cover Index (CRC). Residue cover acts as a proxy for soil disturbance. No till practices will leave high levels of residue cover on the field, while conventional tillage will leave minimal residue on the soil surface (see Tillage Intensity table below).
Weekly residue cover percentages are calculated over the two 8-week observation periods, pre-plant and post-harvest. Observations with the highest confidence (where the the greatest area of the field was observed and where residue cover estimations within the field were consistent) are identified, and a tillage intensity classification is made based on the median residue percent of those observations. This provides two tillage intensity estimations for each crop cycle.
Summarizing Tillage Intensity
There are various types of tillage methods and implements that result in wide gradients of tillage intensity and disturbance. The USDA has provided classifications that relate residue observed on fields to tillage intensity, providing guidance that can be applied across projects and regions. The classifications relate directly to reducing erosion and emission, maintaining or increasing soil health and organic matter, and increasing plant-available moisture.
Regrow uses the USDA residue thresholds as the basis for relating residue amount to a tillage practice. While these thresholds have shown to be a reliable proxy for relating residue amounts to tillage practice for crops that produce a large volume of biomass (ex: corn), we've found them to be often unreliable for crops that produce little biomass, or cases where the biomass breaks down quickly (ex: soybeans, cotton). As a result, Regrow uses two sets of residue thresholds to determine tillage practice, depending on the fragility of the crop's residue.
The following table contains the full list of fragile vs. non-fragile crops. Please note that this list is inclusive of crops across all regions. Some crops may not be available in every region.
| Fragile crops | Non-fragile crops |
| asparagus, barley, beet, berry, blueberry, broccoli, cabbage, camelina, canary_seed, canola, canteloupe, carrot, cauliflower, celery, chickpea, clover, corn_silage, cotton, cranberry, cucumber, dry_bean, eggplant, grape, hop, legume, lentil, lettuce, melon, mustard, oat, onion, pea, peanut, pepper, potato, pumpkin, radish, rye, rye_spring, safflower, soybean, speltz, squash, strawberry, sugar_beet, sunflower, sweet_potato, tobacco, tomato, triticale, turnip, vetch, watermelon, wheat_durum, wheat_spring, wheat_winter | alfalfa, buckwheat, corn, flax, grass_annual, grass_perennial, hay, millet, miscanthus, other_hay, pasture, pop_or_orn_corn, rice, ryegrass, shrub, sod_grass, sorghum, sugarcane, sweet_corn, switchgrass |
Tillage Detection Confidence
Monitor provides a confidence score for each tillage intensity determination. The confidence score is dependent on the amount of the field observed and the variability of residue cover each week. Every weekly observation receives a confidence score from 1 to 3 (with 3 being highest confidence), which would mean a majority of the field was observed with low variability in residue cover estimates across the field (we estimate residue cover for every 10m x 10m pixel). The weeks with the highest confidence in the 8-week observation period (for example, all weeks with confidence score = 3) are used to make the determination, and that confidence score is provided.
Confidence values are reported on a scale of 0-3. Regrow recommends using tillage type predictions when model confidence is 3.
Accuracy and Validation
To assess the accuracy of both residue cover estimations and tillage classification we compare the estimates provided by Monitor to ground truth data provided at the field scale as well as to regional summary statistics of tillage practices where available. Accuracy in identifying conventional till is generally high, but the intricacies and gradients of residue cover that separate reduced till from conventional till can be more difficult to identify, especially as different crop types can have very different post-harvest residue even when the tillage practice is the same. Our most recent validation, using over 22,000 observations, demonstrated a 71% accuracy in identifying conventional tillage vs. conservation tillage (inclusive of no till and reduced till) in the United States (continental). We are working to provide greater specificity based on crop type and region to achieve higher accuracies in this realm.
Cover Crop Detection
Monitor’s Green Cover algorithm predicts the presence or absence of a cover crop during the agronomic periods of time when cover crop activity can happen (between harvest and planting of two commodity crops).
When are fields evaluated for a Cover Crop?
Fields are evaluated for cover crops after the harvest of one commodity to the planting of the next commodity crop. We refer to this time period as an 'observation window'. Because the harvest and planting dates are variable per field, the cover crop observation window is dynamic to each cultivation cycle.
Fields are monitored during the observation for emergence, persistence and vigor of green cover. To avoid inclusion of temporary regrowth, weeds, volunteers and other non-intentional green signals in cover crop determinations, cover crop observations are only made if there are at least 8 weeks between two the commodity crops. We also account for temporary regrowth, weeds and volunteers through regional parameters that reflect the level of growth in surrounding natural herbaceous vegetation.
Methodology
Cover crops are detected using NDVI, an index that is responsive to chlorophyll in the vegetation and serves as a proxy for photosynthetic potential. In other words, we can use this index to monitor how 'green' a field is at a given time, as well as over a time period. Satellite images, captured every 5-10 days, measure the strength of green cover emergence, the persistence of the cover crop over time, and the greenness relationship of the field to the surrounding area throughout the observation period.
Monthly NDVI is calculated and used to assess the persistence and vigor of green vegetation over time. The mean NDVI of the field over the cover crop observation window is compared to regional NDVI thresholds to measure the quality and extent of cover crops on a field. Fields with higher NDVI values indicate the presence of healthy, sustained vegetation, which is often a sign of intentional cover cropping practices.
Cover Crop Classifications
Cover crop classifications are based on the level of greenness sustained throughout the non-commodity crop period.
Regional NDVI thresholds are calculated annually using surrounding natural herbaceous vegetation to quantify constraints on growth due to local and inter-annual variation in weather.
Classifications are region-specific. We calculate inter-annual regional thresholds based on the NDVI of surrounding natural herbaceous vegetation to determine whether the extent and strength of green cover on a field corresponds with cover crop activity. Below are examples of thresholds for a given year across a range of regions. In some cases winter extremes can severely limit regional growth, lowering the threshold to its minimum.
Cover Crop Confidence
Confidence in the cover crop determination is based on the extent of the field observed throughout the non-commodity crop period. For each month we calculate the percent of the field observed. The average of the monthly percent observed values is used to determine the confidence.
Confidence values are reported on a scale of 0-3.
Emergence Quality and Emergence Quality Confidence
Emergence quality is a measure of how strongly a cover crop emergences after planting. This metric along with the mean NDVI provide additional insights about how a cover crop developed, and by extension the resulting cover crop classifications. Emergence quality is calculated using the first three months of the cover crop observation period, where Monitor returns a No vegetation, Weak emergence, Good emergence, or Strong emergence based on the highest NDVI value observed across the first three months. Emergence quality confidence metrics indicate the percent of the field observed during the month with the highest NDVI.
These data points provide additional context on cover crop outcomes. Winter kill and non-sustained emergence are often observed on intended cover crops (those where a field is intentionally seeded, but the cover crop failed to thrive) - in these cases the field can receive a Potential Cover Crop or Cover Crop Not Tracked classification as the average NDVI over the full observation period may be quite low. These updated metrics can indicate that initial vegetation growth was present or strong, but succumbed to adverse conditions (freezing temperatures, snow/ice, etc.) - highlighting cases when a farmer reports planting a Cover Crop, but Monitor does not return that classification.
Accuracy + Validation
Cover crop practice designations made by Monitor are compared to ground truth data provided by farmers and agronomists, and to regional summary statistics provided by government agencies (e.g. USDA) or farmer cooperatives. When compared Monitor Cover Crop results to over 25,000 field observations, Monitor displayed an overall accuracy of 85% in the United States (continental). When compared to the USDA Ag Census, Monitor cover crop results were within +/- 20% of the state level reported cover crop planting rates for 90% of states in the 2017 census and 81% of states for the 2022 census.