1. Introduction
Farmer livelihoods and food production are affected by myriad ongoing changes in climate, markets, and policies. Accurate data on cropping systems are essential to monitor and understand the effects of these changes, yet such data are often lacking. One major aspect of cropping systems are the crops that farmers choose to plant, which typically change from season to season as farmers rotate crops or shift into new crops [
1]. Information on crop choice is helpful for various applications, including modeling land use decisions, mapping yield variations, and forecasting regional food production.
Given the widespread demand for crop type information, maps of crop types have been developed from a variety of sources and at a range of spatial and temporal resolutions [
2,
3]. In a small number of countries, such as the United States [
4], Canada [
5], and France [
6], detailed crop type maps at the field scale are publicly available for each growing season, based either on farmer surveys or a combination of ground and satellite sources. In most countries, however, timely data is much harder to obtain. Several gridded datasets with global coverage have been developed, but these are often based on census data more than a decade old [
2]. Given the dynamic nature of agriculture, including evidence of rapid cropland expansion in some regions and cropland abandonment in others [
7], decades old data are insufficient for many uses. Moreover, global products typically have a resolution of 10 km or coarser [
2], which limits their utility for applications requiring field-scale data.
As a result, there is a continued need for improved approaches to mapping crop types [
2,
3]. This need is recognized, for example, by the new WorldCereal effort that aims to create global, annual maps for wheat and maize at 10 m resolution (
https://esa-worldcereal.org/, accessed on 1 March 2023). Remote sensing offers clear advantages for large-scale crop type mapping, with proven success in many local or national scale studies [
4,
8,
9,
10,
11,
12]. Yet a major challenge remains that models require large amounts of training data, and models trained in one region for a single season often do not transfer well to other regions or seasons. One sensible way to address this challenge, as in the WorldCereal project, is to invest in large amounts of field data collection around the world, so that models can be locally trained anywhere. Other efforts have focused on developing models that are better able to maintain performance in years or locations outside of their training domain [
13,
14,
15].
A third, complementary approach has been to seek training data derived without the need for field data collection. In recent work [
16], we demonstrated the promise of one such source of data—LiDAR measurements acquired by the Global Ecosystem Dynamics Investigation (GEDI) [
17]. GEDI LiDAR returns provide information on canopy heights with a nominal spatial resolution of 25 m and a vertical precision of roughly 50 cm [
17]. Although many crops have similar heights, some of the key commodity crops grown throughout the world, especially maize and sugarcane, are typically 1 m taller than other common crops such as wheat, rice, or soybean (
Figure 1).
Indeed, in many landscapes the two main crops are one tall crop and one short crop (e.g., maize and soybean, or sugarcane and rice), such that the ability to discern tall from short crops goes a long way toward mapping individual crops (
Figure 2). For example, roughly two-thirds of the area sown to maize in the world lies in regions where maize is 90% or more of the area covered by tall crops.
GEDI alone, however, only samples a very small fraction of the landscape, so therefore rather than use GEDI directly, Di Tommaso et al. [
16] use GEDI to train a random forest model that predicts crop height class based on Sentinel-2 (S2) optical data. This combined GEDI-S2 approach was found to map crop types nearly as well as a model trained on thousands of local ground training points in the US, France, and China.
Here, we develop and test an approach that combines GEDI and S2 to map tall and short crops throughout the world for a three-year period (2019–2021). We extend the initial insight from Di Tommaso et al. [
16], namely that GEDI signals are informative in cropped landscapes, in several important ways. These include a method to automatically identify the most appropriate months for tall crop delineation, an investigation into the effects of view angle and topography on GEDI signals in the context of crop discrimination, and global scale implementation of the GEDI-S2 approach. We also conduct an evaluation of GEDI-S2 models in a much broader set of countries and cropping systems, using various independent datasets on crop types during the study period. Overall, we find that GEDI is a useful resource for advancing the goal of low-cost, timely, and accurate global mapping of crop types. At the same time, we identify some important areas for improvement to guide future research efforts.
The following section describes the various datasets used in the study, including any initial processing steps for the data.
Section 3 then describes the methods used to map crop height class and evaluate the predictions.
Section 4 presents the main results, while
Section 5 discusses various sources of errors and potential future directions for improvement. Finally,
Section 6 briefly summarizes the main conclusions.
2. Datasets
This study utilized five main data sources: a global cropland mask used to define cropland areas, GEDI shot returns for cropland areas, Sentinel-2 optical imagery, reference data on crop types from three regions used to train the GEDI classification model, and reference data on crop types from throughout the world used to evaluate the performance of our GEDI-S2 tall/short crop predictions. Below we describe each of these, as well as supplementary datasets used to analyze and interpret our results, including a global map that defines the number of growing seasons in each location, and reference crop type maps used to analyze the relationship between peak crop biomass and model errors.
2.1. Crop Mask
To identify cropped areas we used the European Space Agency (ESA) [
19] and ESRI [
20] global Sentinel-based 10 m global land cover maps available in the Google Earth Engine (GEE) [
21] official and community data catalogs, respectively [
22]. Both the ESA WorldCover 2020 product and ESRI 2020 Global Land Use Land Cover provide a global land cover map for 2020 at 10 m resolution, the former based on Sentinel-1 and Sentinel-2 data, and the latter based on Sentinel-2 alone. We primarily used the ESA mask, which, based on visual inspection using Google’s high-resolution basemap, better captured cropland in most areas where the two maps disagreed. However, for Kenya and Uganda the ESA mask tended to greatly underestimate cropland area, and to better capture cropland for these countries we therefore merged the two masks, defining a pixel as cropland if either of the two classified it as cropland.
2.2. GEDI Data
GEDI is a sensor onboard the International Space Station (ISS) that acquires LiDAR waveforms between 51.6°N and 51.6°S to observe the Earth’s surface in 3D. It is the first spaceborne LiDAR instrument specifically optimized to measure vegetation structure [
17]. It contains three lasers emitting near-infrared (1064 nm) light. Two of the lasers are full-power lasers, with the other coverage laser split into two beams, producing a total of four beams. Each beam is then optically dithered across-track resulting in eight ground tracks (four full power and four cover tracks) spaced 600 m on the ground. Shots have an average footprint of 25 m in diameter and are separated 60 m along the track.
GEDI spatial coverage changes in time. In particular, in early 2020 the ISS lifted its orbit, causing GEDI to have “orbital resonance” which means it goes over the same tracks repeatedly while leaving big gaps in between (
Figure 3a–c). While orbital resonance does not change the number of shots acquired in a time period, it reduced the spatial coverage of GEDI in 2020 relative to 2019. When GEDI samples of agricultural areas are less geographically uniform and more clustered, we expect GEDI-based crop type classification accuracy to decrease.
Another important aspect of GEDI is that while its viewing angle is typically near-nadir, it can be rotated by up to 6°, allowing the lasers to be pointed up to 40 km on either side of the ISS ground track. This capability is used to sample the Earth’s land surface as completely as possible, but can also complicate interpretation of the GEDI returns [
23].
For this study, we used the GEDI dataset Level 2A (L2A) and Level 2B (L2B) from April 2019 to December 2021, available in GEE data catalog. Level 2 data provide information about the vertical distribution of the canopy retrieved from the waveform return at footprint level. The main GEDI product used is GEDI’s L2A Geolocated Elevation and Height Metrics Product, which is primarily composed of Relative Height (RH) metrics, which collectively describe the waveform collected by GEDI. Relative Height (RH) metrics give the height at which a certain percentile of energy is returned relative to the ground. RH are reported at 1% intervals, resulting in 101 metrics. The GEDI L2A dataset (
LARSE/GEDI/GEDI02_A_002_MONTHLY) is a rasterized version of the original GEDI product, with each GEDI shot footprint represented by a 25 m pixel [
24]. This rasterization process can introduce an additional geolocation error to the initial GEDI shot error. The raster images are organized as monthly composites of individual orbits in the corresponding month. RH values and their associated quality flags and metadata are preserved as raster bands.
A secondary dataset L2B was used to retrieve the GEDI view angle (i.e., local beam elevation property). This is available in GEE as a table of points (LARSE/GEDI/GEDI02_B_002) with a spatial resolution (average footprint) of 25 m. At the time of writing, the raster version of the L2B dataset (LARSE/GEDI/GEDI02_B_002_MONTHLY) is only partially ingested in GEE, and we therefore used the table.
2.3. Sentinel-2
We used S2 surface reflectance data (Level-2A) present in GEE and filtered out clouds using the S2 Cloud Probability dataset provided by SentinelHub in GEE. The Sentinel-2A/B satellites acquire images with a spatial resolution of 10 m (Blue, Green, Red, and NIR bands) and 20 m (Red Edge 1, Red Edge 2, Red Edge 3, Red Edge 4, SWIR1, and SWIR2 bands), and together they provide images at a 5-day interval.
To capture crop phenology, we extracted S2 imagery for 2019–2021 from 1 January to 31 December for the northern hemisphere, and from 1 July of one year to 30 June of the next for the southern hemisphere. Features were extracted from S2 time series by fitting harmonic regressions to all cloud-free observations in cropped areas. For each spectral band or vegetation index
, the harmonic regression takes the form
where
are cosine coefficients,
are sine coefficients, and
c is the intercept term. The independent variable
t represents the time an image is taken within a year expressed as a fraction between 0 and 1. The number of harmonic terms
n and the periodicity of the harmonic basis controlled by
are hyperparameters of the regression.
To determine n and , we sampled multiple locations around the world and compared the harmonic fit of the time series by varying the hyperparameters. We found that third order harmonics () with were a good fit for both regions with one or multiple growing seasons.
We computed harmonic coefficients for four bands and one vegetation index: NIR, SWIR1, SWIR2, RDED4 and GCVI. GCVI is the green chlorophyll vegetation index [
25] computed as
This yields seven features per band, for a total of 35 coefficients. Previous crop type classification studies [
26,
27] have reported the efficacy of using these four bands and VI, demonstrating performance comparable to classification models using all optical bands and a variety of other VIs.
2.4. GEDI Model Training Dataset
To train the GEDI model to distinguish tall from short crops we used high-accuracy crop type labels from 2019 from the three regions used in prior work [
16] and mapped in red in
Figure 4: Jilin province in China, Grand Est region in France, and Iowa state in USA. These regions are major agricultural production areas containing a mix of tall and short crops and have accurate, field-scale crop type maps that are publicly available. Although at similar latitudes, these regions are located in three separate continents and management practices do differ. Maize in France in particular exhibits a wide range of GCVI, and China exhibits very small fields. Differences in agricultural practices across regions for the same classes could translate to differences in the GEDI waveforms, helping the GEDI model to be more flexible and adaptable in other regions as well.
For Jilin, China we used the 2019 crop type map produced by You et al. [
9]. It maps three major crops in the area (maize, soybean, and rice) at 10 m with an accuracy of 87%, and F1-scores of 85% for maize. For Grand Est, France we used the Registre Parcellaire Graphique (RPG) 2019 dataset downloaded from
https://www.data.gouv.fr/ (accessed on 1 July 2021). It is a public georeferenced vector product derived via survey. For Iowa, USA we used the U.S. Department of Agriculture’s 2019 Cropland Data Layer (CDL) at 30 m resolution available in GEE [
4]. It has an overall accuracy of 90%, and precision and recall for maize exceed 95%.
2.5. Evaluation Datasets
To evaluate our product, we sought high-quality crop type datasets for a diverse set of cropping systems and regions for the 2019–2021 period. We used a combination of field data and crop type maps that were produced by combining field and satellite data. In regions with multiple growing seasons, we filtered for crop type labels matching the growing season that the GEDI model predicted for. A map of location and type of reference data is shown in
Figure 4 and a summary of data characteristics including sample size are given in
Table 1.
2.5.1. Ground-Based Reference Data
Europe
Schneider et al. [
28] contains harmonized agricultural parcels information data from regions in Austria (2019), Denmark (2020), and Slovenia (2019). The parcel data are based on publicly available self-declared crop reporting datasets, gathered for the purposes of subsidy payments. We focus on Austria and Slovenia, since the Denmark dataset is outside the GEDI latitude coverage.
Canada
Agriculture and Agri-Food Canada [
29] is a collection of thousands of points identifying crops types and occasionally other land cover types across Canada from 2011 to 2021. These point sources are used by Agriculture and Agri-Food Canada (AAFC) as training or reference sites for the creation of the Annual Crop Type map.
Malawi
Field boundaries for three crops—groundnut, maize, and soybean—were collected in five districts of Malawi (Lilongwe, Ntchisi, Kasungu, Salima, and Mzimba) for 2021 as part of research on the groundnut value chain conducted by the AgroHitech Innovation and Advisory Consortium for the Peanut Innovation Lab.
Mali
Field boundaries in Mali were collected by the NASA Harvest team during the 2019 growing season [
30]. The crop type growing in each field was observed by a surveyor. In total, the dataset contains 148 fields. The data were released as part of the CropHarvest dataset and is also available on Radiant MLHub.
Kenya
The Global Agriculture Monitoring initiative of the Group on Earth Observation, called Copernicus4GEOGLAM, collected ground reference data during field surveys in three countries—Uganda, Tanzania, and Kenya [
31]—for both the 2021 long rain and 2021–2022 short rain seasons. The georeferenced ground data were used by Copernicus4GEOGLAM to train random forest models to map crop type using S2 imagery as input. Given the fairly low accuracies of the resulting maps (e.g., maize F1 scores were often below 0.6), we utilized only the field data for our evaluation. We focused on the Kenya point dataset for the 2021–2022 short rain season, which had the greatest number of points overlapping with the season of our GEDI-S2 predictions.
India
India crop type labels are crowdsourced from farmers via Plantix, a free Android application created by Progressive Environmental and Agricultural Technologies (PEAT). The Plantix app is used by farmers who submit photos of their crops seeking help to diagnose and treat crop diseases. As part of the disease diagnosis, PEAT uses a convolutional neural network to assign crop labels based on the submitted photos. We used these data in the Indian states of Maharashtra and Telangana, where the accuracy of Plantix crop type labels exceeds 0.90 for most major crops. These data have been cleaned to remove location inaccuracy (keeping only submissions with GPS accuracy better than 10 m), as suggested by previous work by [
32]. To match the timing of the GEDI-S2 predictions, we filtered the Plantix data for the 2021 kharif season based on photo submission timing.
2.5.2. Satellite-Based Reference Data
United States
The Cropland Data Layer (CDL) produced by the United States Department of Agriculture (USDA) provides yearly crop type maps across the conterminous US at 30 m spatial resolution [
4]. Maps are based on Landsat and other satellite imagery using training data from the Farm Service Agency (FSA). For validation we chose two states, North Dakota and Alabama, that were far from the conditions and locations of the Iowa locations used in the training data. Accuracy of CDL on FSA labels are available in the CDL metadata, with precision and recall for maize for 2019–2021 higher than 81% and 85% in North Dakota and Alabama, respectively.
Germany
National scale crop type maps for Germany were recently produced for 2017–2019 [
33] and 2020 [
34]. These maps are generated using a random forest classifier based on Sentinel-1, Sentinel-2 and Landsat time series, with parcel data used for training. More details about the underlying data and methods can be found in Blickensdörfer et al. [
12]. Overall accuracy for 2019 is 78%, with precision and recall for the maize class of 90% and 83%, respectively.
Brazil
Annual soybean maps were recently produced for South America at 30 m resolution between 2000 and 2020 by combining Landsat and MODIS satellite observations and sample field data [
11]. These maps are available in GEE as (
projects/glad/soy_annual_SA). For evaluation, we focused on western Bahia in 2020, since this region grows maize and soy in the main season and has only one primary growing season per year. We evaluated only recall on soy, since other short crops, e.g., cotton, are also grown in the same season but are not distinguished from other non-soy crops in their study. Accuracy for 2020 is not reported, but for the years 2017–2019 they report overall accuracies of 96%, 94%, and 96%, respectively, with high and balanced producer’s and user’s accuracies.
China
For China, we used the same 2019 crop type map described by You et al. [
9] in
Section 2.4 used in training the GEDI model. For validation, we used a random sample over the four northeast regions (Liaoning, Nei Mongol, Jilin, and Heilongjiang), which span a much larger area than used in the training sample from Jilin.
India
Lee et al. [
35] produced a map of sugarcane area in the Upper Bhima Basin, a major sugarcane producing region in Maharashtra, India. Their 10 m resolution map is based on crowdsourced Plantix data and a neural network applied to S2 data. Reported overall accuracy for sugarcane vs. not sugarcane was 77% (85% precision and 67% recall).
2.6. Number of Growing Seasons per Year
The Anomaly hotspots of Agricultural Production (ASAP) system is an online decision support tool for early warning about production anomalies developed by the Joint Research Center (JRC) of the European Commission. ASAP has produced several maps including satellite-based phenology information, which are computed from the long-term average of MODIS NDVI data at
resolution [
36]. We downloaded the phenology layer that defines the number of growing seasons (1 or 2 seasons) [
37], and aggregated this information at 5 based on the majority of the crop pixels’ seasonality.
2.7. Digital Elevation Model (DEM)
We used a DEM to investigate the effect of topography on the usability of GEDI shots for tall crop classification. The Shuttle Radar Topography Mission SRTM V3 (SRTM Plus) [
38] digital elevation data product is provided by NASA JPL at a resolution of 1 arc-second (approximately 30 m) and is available in GEE. We calculated the slope in degrees from the terrain DEM in GEE.
2.8. Reference Maps for Error Analysis
As described below, we hypothesize that errors in our GEDI-S2 predictions were often related to low biomass of the tall crop. To further investigate this, we utilized two additional crop type maps that provided wall-to-wall coverage in countries where our preferred data for evaluation covered only a subset of fields. Widespread coverage was needed to ensure a wide range of biomass values for pixels in the reference map that overlap with the GEDI shot locations.
2.8.1. Canada
The Earth Observation Team of the Science and Technology Branch at Agriculture and Agri-Food Canada (AAFC) have created Annual Crop Inventory maps that are accessible in GEE. These maps are generated using a combination of crop type labels from crop insurance data and ground-truth information collected across the country to train a decision tree model based on optical and radar satellite images. Maps have a spatial resolution of 30 m and an accuracy of at least 85%.
2.8.2. Kenya
As described above, in addition to field data, the Copernicus4GEOGLAM produces end-of-season crop type maps for each country and season where field data was performed [
31]. In our error analysis, we used the long rains map for Kenya, which possesses the highest F1 score for maize among the various countries and seasons. For this map, overall crop type accuracy is 80%, and F1 for maize is 0.64.
3. Methods
Here, we describe the steps taken to create and evaluate wall-to-wall maps of crop type height, using a combination of GEDI and S2 as input.
Figure 5 provide a graphical overview of the methods presented in this paper.
The sections below describe in detail each of the six steps in this process:
- 1.
Train a single model, which we refer to as the GEDI model, that uses GEDI features to classify locations as having short crops, tall crops, or trees;
- 2.
Apply the GEDI model to GEDI shots acquired from cropland areas globally for three years of 2019–2021;
- 3.
Tile the globe into grid cells;
- 4.
Determine the optimal month to predict tall crops for each grid-cell;
- 5.
Train a local GEDI-S2 model for each grid-cell based on GEDI predictions in the 3-month time window around the optimal month;
- 6.
Evaluate results against local reference data.
3.1. GEDI Model Training
Following Di Tommaso et al. [
16], we began by defining a random forest model to classify GEDI shots in three crop height classes: short, tall or tree. The decision to use the random forest model was primarily motivated by its high accuracy, advantageous computational efficiency and seamless implementation at a large scale in GEE. To train the model we used labels from three areas with high-quality crop type maps for 2019: Jilin in China, Grand Est in France, and Iowa in the United States (see
Section 2.4). Crop type labels were sampled at GEDI shot locations and assigned a
tall label for maize class and a
short label to remaining short crops. We also defined a third
tree class, for shots with RH100 greater than 10 m. The choice of the 10 m threshold was made empirically, relying on visual assessment of GEDI shots over trees in Google’s high-resolution basemap.
We used all GEDI shots in August 2019 over the three regions since our previous study [
16] showed August to be a good time to distinguish maize from other short crops in these regions. This resulted in a total of approximately 253 k samples, with 47 k samples in Jilin, 23 k in Grand Est and 183 k in Iowa.
Although we are interested in the crop height, we found that our model worked best when multiple RHs were included to fully capture the GEDI returned waveform. To reduce the number of features, since consecutive RH metrics are highly correlated with each other, we sampled a metric every 5% and omitted RH in the middle of the RH profile based on feature importance analysis. In total, 11 RH metrics were used: RH0, RH5, RH10, RH15, RH20, RH25, RH30, RH85, RH90, RH95, and RH100.
The features and labels were used in a random forest model, implemented in GEE. Data were split into 80% training and 20% test points. To minimize spatial correlation across the training and test sets, we binned the shots by their lat/lon into bins and GEDI shots in each bin were placed entirely in either the training set or test set. The overall test accuracy across the three regions was 0.885, with F1 scores for short, tall and tree classes of 0.863, 0.898, and 1.000, respectively. The very high F1 score for the tree class is explained by the definition of tree class, as based on GEDI RH100 metric directly and not field labels.
3.2. GEDI Model Predictions
The random forest model described above was then applied to all GEDI shots in cropland pixels, according to the crop mask described in
Section 2.1. The predicted class was saved along with the prediction probabilities (the fraction of trees in the random forest model that predicted the class) as a measure of confidence.
Figure 3 illustrates these predictions for a selection of GEDI orbits, with shots colored orange for tall and gray for short based on the GEDI model predictions.
The predicted shots were filtered to retain only high quality shots to use as labels in subsequent steps. First, we removed shots with a quality flag value of zero in the original GEDI returns, which indicates poor quality, as well as shots with a non-zero degrade flag, which indicates poor geolocation. We then removed low confidence predictions (lower than 0.8) to have more confidence in the GEDI-generated labels.
Another step that proved essential was to filter out shots with low view angle and on high slope terrain since both factors can affect the accuracy of the GEDI model predictions. We refer to view angle as the angle between the off-nadir beam and the ground. Prior work has revealed that small changes in view angle can increase errors for models based on GEDI returns [
23,
39]. In particular, existing analysis recommends removing observations where the view angle was below 1.5 rad, or roughly 86° [
39].
To explore the appropriate threshold for our application, we considered shots for the US Corn Belt where we have confidence in the reference data from CDL, and where the view angle property is available in GEE at the GEDI shot level. The GEDI model prediction errors (treating CDL as truth) were evaluated for different levels of view angle, as shown in
Figure 6a. At low view angles, errors are as high as 60%. Above the recommended threshold of 1.5 rad, however, errors are below 10% and fairly insensitive to additional increases in view angle. We therefore adopted a threshold of 1.51 rad for further analysis.
The GEDI view angle varies over time as shown in
Figure 6b. View angles were particularly low in June and July of 2020, causing the removal of most shots during the peak of the growing season in many regions. Other periods of frequent observations with low view angles include late 2019 and late 2021. Unfortunately, at the time of writing, information on the GEDI view angle was not yet available at the shot level for all shots globally in the GEE catalog. To create a view angle filter, we sampled orbits over a longitudinal transect and aggregated these data, averaging the view angle for each beam on each day. We then removed all shots from beams and days with an average view angle above 1.51 rad. Although this was a pragmatic way to filter out data with low view angle globally, future versions would likely benefit from accounting for view angle at the shot level, to account for variation by latitude and over time within the day.
We also removed shots on high slope terrain, defined as areas with slope higher than 5°. GEDI metrics are dependent on topographic slope [
40], and given the relatively small height signal being used by our model to classify tall vs. short crops, the effect of topographic slope are potentially important. Based on analysis of CDL in the United States, similar to the view angle analysis presented in
Figure 6a, a slope below 5° was deemed sufficient to avoid artifacts from the terrain. As cropland is typically situated on flat or nearly-flat land, this filter removed only a small fraction of GEDI shots.
3.3. Model Grids
The filtered GEDI model predictions provide labels with which to train a model that takes S2 data as input. However, we did not expect a single model to be applicable globally, since the timing of growing season and mix of crops differs across the world. Building on prior approaches [
24,
41] we instead sought to develop locally-calibrated models. We defined a grid within the GEDI coverage (between 51.6°N and 51.6°S) with
cells. Although more localized models would potentially improve performance in some regions and years, the choice of grid cell size was dictated by the orbital resonance of GEDI in 2020 and 2021. That is, a finer grid would often have cells that have very few GEDI observations because of the large gaps in GEDI coverage in those years. Furthermore, moving from the pole towards lower latitudes, the spacing between GEDI tracks becomes more substantial. This necessitated at lower latitude the adoption of larger grid cells in terms of real area to ensure sufficient data coverage. By employing 5-degree grid cells, we struck a balance between capturing the heterogeneity within each cell and ensuring an adequate sample size of GEDI data for model training and analysis.
To reduce computation, we only processed grid cells for which more than 5% of S2 pixels were classified as cropland, yielding 238 cells. The grid cells that we kept cover an area that comprises 93% of the total crop area within the latitude bands of GEDI coverage.
3.4. Optimal Timing
For each grid cell, we defined the optimal month to classify tall vs. short crops as the month in which the highest percentage of GEDI shots were predicted as tall. Specifically, we combined the 3 years of GEDI model predictions by month, computed the percentages of tall and short shots by grid cell and month, and then selected the month for each cell when the percentage of tall shots was highest. We interpret this month as the period during which tall crops have reached their peak height. Consequently, this is the time when GEDI is most likely to detect a contrast with other crops within that particular cell.
3.5. GEDI-S2 Models
For each grid cell and for each year, we separately trained a local 2-class (tall vs. short) S2 model using the GEDI predictions for the relevant time as labels and harmonic coefficients as features. Data were randomly split into 80% training and 20% test, to evaluate model accuracy. We refer to these as GEDI-S2 models, with a unique model for each grid cell and year. To account for variations in the timing of the growing season within the 5° grid cells, we considered a three-month window centered on the optimal month. We created GEDI-S2 predictions for individual months and then combined the predictions on a pixel basis, with pixels classified as tall if the predicted class was tall in any of the three months.
The result of this process was a wall-to-wall 10 m resolution map of tall and short crops for all cropland pixels in the 5° grid cells. To reduce computation, we only applied the GEDI-S2 models to grid cells where the percentage of tall shots is higher than 4%, i.e., 201 grid cells per year. Since GEDI data was not always available in all the regions in the time window of interest, the number of grid cells processed is 1562, less than the expected 1809 = 201 (grid cells) × 3 (months) × 3 (years). This resulted into 189 unique locations in 2019, 457 in 2020 and 201 in 2021, for a total of 590 grid cells for the 3 years.
For this local training, we omitted all shots where the GEDI model predicted the tree class, as these were viewed as likely to be a mixture of crops and trees within the GEDI footprint, which at 25 m diameter is more than four times larger than the 10 m S2 pixel. Thus, predictions of tree were viewed as unreliable labels for a 2-class model focused on S2 pixels classified as cropland.
To minimize spatial artifacts when mosaicking adjacent cells, we created predictions for pixels in a buffer around each cell and mosaicked the overlapping predictions taking the predictions in the cell with higher GEDI-S2 accuracy.
3.6. Evaluation of GEDI-S2 Predictions
The first evaluation of GEDI-S2 predictions is against reference data from around the globe (
Table 1). All reference data were ingested in GEE for comparison with GEDI-S2 predictions. For regions with ground-based point or polygon data, we used all fields for evaluation. In the case of polygons, the centroid of the polygon was used to define the relevant pixel from the GEDI-S2 predictions for comparison. For regions where crop type maps were used, we randomly sampled the maps using 2000 to 4000 points and removed the ones without a specific crop type label to create a reference dataset.
Because some datasets contain as many as 100 crop types, for each reference dataset we selected the 10 most common crops for evaluation, which typically represents more than 90% of the crop areas in the reference regions evaluated. From specific crop type labels, we generated a binary tall/short classification, with maize and sugarcane defined as tall and all other crops defined as short (none of our evaluation data had sunflower or cassava among the 10 most common crops). For each evaluation dataset, we report the accuracy, precision, recall, F1 and Kappa scores, using the following equations [
42]:
where
True Positive, , is the number of samples labelled as positive by the model that are actually positive
False Positive, , is the number of samples labelled as positive by the model that are actually negative
True Negative, , is the number of samples labelled as negative by the model that are actually negative
False Negative, , is the number of samples labelled as negative by the model that are actually positive
is the proportion of observed agreement, i.e., the accuracy achieved by the model
is the proportion of agreements expected by chance
Our second evaluation compares GEDI-S2 predictions against an S2 model trained locally within each reference region. These S2-Local models provide benchmarks that represent how well a model trained on local field data would perform. To conduct this analysis, we exported S2 harmonic features at the same reference field locations and trained a local S2 model based on binarized tall/short field labels. We used the RandomForestClassfier implemented in Python’s scikit-learn package, setting similar hyperparameters to the GEDI-S2 random forest models implemented in GEE. Reference data were binned by their lat/lon into bins, and data in each bin were placed entirely in either the training set or test set using a 80%/20% train/test split. We ran the S2 classifier multiple times using each time a different train/test split and reported the average S2-Local model performance metrics.
6. Conclusions
In this study we sought to test the general applicability of an approach that uses GEDI returns to train local crop type mapping models that use S2 data as input. Our general conclusion is that the approach exhibits considerable promise for advancing crop mapping. Tall and short crops were mapped with high accuracy in the majority of maize production systems, including most of the Americas, Europe, and East Asia. Specifically, we showed that GEDI returns can first be classified into tall and short crops, that the frequency of tall crops over time can be used to identify the appropriate months for S2 training (when tall crops are at their peak height), and that S2 models trained on these GEDI shots can accurately predict the GEDI crop height class in nearly all regions. Only in rare cases, such as areas with high topographic variation, did S2 features fail to predict the GEDI crop height class. We then showed that the predictions from the GEDI-S2 agree remarkably well with independent reference data at the field scale.
At the same time, we uncovered cases where the current implementation of GEDI-S2 is problematic. The most common cause for low accuracy appears to be low biomass of tall crops, which occurs frequently in Africa and South Asia. In these regions, the GEDI classification model consistently underestimated the frequency of tall crops. S2 models trained on these shots then inherit this under-prediction of tall crop area. Although this is a notable limitation of the current approach—particularly because these regions are among those with the most limited ground data, and thus where an approach that relied on GEDI for training would be most valuable—we anticipate that future work can greatly improve the performance in low biomass regions. Progress seems most likely for semi-supervised methods that can leverage the fact that even low biomass areas typically have a significant number of fields with high biomass that are accurately captured by GEDI.