Automated Cropland Fallow Algorithm (ACFA) for the Northern Great Plains of USA

ABSTRACT

Cropland fallowing is choosing not to plant a crop during a season when a crop is normally planted. It is an important component of many crop rotations and can improve soil moisture and health. Knowing which fields are fallow is critical to assess crop productivity and crop water productivity, needed for food security assessments. The annual spatial extent of cropland fallows is poorly understood within the United States (U.S.). The U.S. Department of Agriculture Cropland Data Layer does provide cropland fallow areas; however, at a significantly lower confidence than their cropland classes. This study developed a methodology to map cropland fallows within the Northern Great Plains region of the U.S. using an easily implementable decision tree algorithm leveraging training and validation data from wet (2019), normal (2015), and dry (2017) precipitation years to account for climatic variability. The decision trees automated cropland fallow algorithm (ACFA) was coded on a cloud platform utilizing remotely sensed, time-series data from the years 2010–2019 to separate cropland fallows from other land cover/land use classes. Overall accuracies varied between 96%-98%. Producer’s and user’s accuracies of cropland fallow class varied between 70-87%.

KEYWORDS:

• Agriculture
• machine learning
• cloud computing
• image classification algorithms
• food security
• water security

1. Introduction

Cropland Fallows are defined as an agricultural practice where farmlands equipped for cropping are left unplanted, or no crop is grown during a crop growing season (Stewart and Thapa 2016). In other words, in cropland fallow agriculture, a growing season is intentionally skipped. There are various reasons for leaving croplands as fallows. These reasons include soil enrichment, water savings, government subsidies for not growing crops in certain portions of farms to manage the economics of supply and demand, and other factors like farmer decisions to manage labor and costs (Song, Prishchepov, and Song 2022). For example, fallow periods are particularly important in dryland or semiarid areas where there is not enough soil moisture to plant a cash crop every season; cropland fallows can accumulate plant available water in the soil profile (Stewart and Thapa 2016). In many marginal dryland areas, rainfall patterns vary significantly from year to year and in certain years there isn’t sufficient precipitation to provide adequate water to grow crops (Gumma et al. 2016). Cropland fallowing is also frequently practiced in areas that receive most of the annual precipitation in the winter when temperatures are too cold for crops; when winter precipitation is less than average, farmers can choose to plant less the following season (Guy and Gareau 1998). Additionally, cropland fallows serve as a break between cash crops, which can lessen pathogens and insects that reduce cash crop yield (Peairs, Bean, and Gossen 2005).

Cropland fallowing is also practiced in irrigated areas when limited water resources are in high demand for multiple uses such as for large urban areas like in California and the Southwestern United States. In some areas, farmers can temporarily sell water rights to cities, thereby offsetting their lost income by not producing crops (McLane and Dingess 2014). In other areas, farmers with junior water rights may not be allowed to irrigate (and thus not plant) in dry years (McLane and Dingess 2014). Fallowing is also financially supported by some governments; the European Union provides a subsidy to farmers in member states who leave 5% of their land fallow to provide wildlife habitat, maintain biodiversity, increase soil moisture, and stabilize grain prices (European Commission 2013).

In the United States, cropland fallows play an important part in many crop rotations, particularly in areas that are purely rainfed (with no irrigation). Such areas, for example, include the Palouse, an area where wheat is the dominant crop in Southeast Washington and Northern Idaho (Guy and Lauver 2007) and the Northern Great Plains. In the Northern Great Plains, it is traditional to plant a wheat crop every other year; since the growing period is 4 months for spring wheat and 9 months for winter wheat, many fields are fallow longer than they are cropped (Stewart and Thapa 2016). One downside to long fallow periods is soil loss from wind and rain erosion.

An alternative to conventional fallow practices is planting cover crops which can protect the land from soil erosion and add organic matter (Reeves 1994). Cover crops can also increase nutrients, such as nitrogen fixation in legumes or sequestering nutrients added from fertilizer thus stopping them from leaving the root zone (Liu, Ma, and Bomke 2005). An important factor to consider when cover crops are planted between cash crops is the ability to kill cover crops before they go to seed; otherwise they may compete with cash crops during the next season (Idowu and Grover 2014). The cost of seeding is weighed against increases in yield and or fertilizer use to determine if planting cover crops is more cost effective than leaving fields fallow (Wojtkowski 2010).

Using satellite remote sensing to map agricultural croplands has been done since the 1970s (Steven and Clark 2013): notable early projects include the U.S. Department of Agriculture (USDA) Large Area Crop Inventory Experiment (LACIE) (MacDonald 1976). These early studies classified one or few images over an area using supervised or unsupervised classifications (Boryan et al., 2011). From those early days, cropland mapping has become frequent, using multiple satellite sensor data, and over large areas. Notable recent examples include mapping crop extent over entire continents at 30-meter resolution using machine learning and cloud computing (Gumma et al. 2020a; Massey et al. 2018; Oliphant et al. 2019; Phalke et al. 2020; Teluguntla et al. 2018; Thenkabail et al. 2021; Xiong et al. 2017), as well as specialized agricultural mapping over large areas. Specialized maps include yield maps (Skakum et al. 2019), crop rotation (Alemu, Henebry, and Melesse 2019), crop phenology (Gumma et al. 2020b), field preparation (Azzari et al. 2019), and cropland fallows mapping (Wallace et al. 2017).

Prior cropland fallow mapping using remote sensing include the Fallowland Algorithm based on Neighborhood and Temporal Anomalies (FANTA) (Wallace et al. 2017) and Automated Cropland Classification Algorithm (ACCA) (Wu et al. 2014). FANTA compares the current greenness of a cultivated pixel to its historical greenness and to the greenness of all cultivated pixels within a defined spatial neighborhood which can be applied across time and space and does not require training data (Wallace et al. 2017). FANTA produced a classification of cropland fallows that was 85% accurate overall when compared to field data collected for the study, with producer’s and user’s accuracy of fallowed-land 75% and 80%, respectively. ACCA has been used to produce cropland fallow maps utilizing MODIS data (Vermote 2015) with a producer’s accuracy of 93% and a user’s accuracy of 85% and was compared with USDA Farm Service Agency (FSA) crop acreage-reported data for both cultivated and fallow croplands with an accuracy of ≥ 95% for cultivated croplands and ≥ 76% for fallow croplands (Wu et al. 2014). More details of the ACCA and FANTA projects are available at (http://www.usgs.gov/WGSC/GCWP). Teluguntla et al. (2017) also mapped croplands and cropland fallows for Australia for the years 2000 through 2015 using MODIS time-series data. Cropland fallows were also mapped for Spain (Recuero et al. 2019) and the Sahel drylands (Tong et al. 2017). Recuero et al. (2019) used MODIS 250 m 8-day NDVI time-series to map fallow agricultural land in Spain from 2001 to 2012 using state collected data for training and validation and a random forest classifier (Vermote 2015). They determined how many times a given field was fallow over the previous 12 years. Groupings/classes were as follows: never fallow cropland, fallow 1–3 years, fallow 4 years, fallow 5–7 years, not cropland (Recuero et al. 2019). Users and producers’ accuracies of the fallowland classes varied between 62% to 92% for cropland fallows for the 1–3 and 5–7 years categories. However, their validation sample size was very low with only 68 samples total (Recuero et al. 2019). Tong et al. (2017) mapped cropland fallows in Nigeria primarily using MODIS Terra 8-day imagery (Vermote 2015). Interestingly, they found that fallow fields had a higher NDVI than actively cropped fields which allowed them to be separated using an unsupervised fuzzy classifier, and that mixed pixels (more than one landcover type in a single pixel) were particularly problematic. Field sizes were small in the Sahel which necessitated calculating sub pixel areas (Tong et al. 2017). In contrast, field sizes in the United States are large, particularly in the Northern Great Plains, which enables MODIS 250 m imagery to be used for cropland and cropland fallow mapping (Didan and Huete 2015).

Currently, croplands and cropland fallows are mapped annually for the conterminous United States (CONUS) by the USDA in their annual Cropland Data Layer (CDL) using remote sensing data (USDA NASS CDL 2022). The CDL is produced by the statistical arm of the USDA, the National Agricultural Statistics Service (NASS). The modern CDL mapping program started in 2006 using Erdas Imagine with decision tree models (Boryan et al. 2011; Hexagon AB. 2021). CDL products are created at the state level during the growing season for internal crop monitoring and are released to the public in February–March of the following year. The current version (2010-2019) maps 110 crop classes and 23 land cover classes. For major commodities (corn, soybean, wheat, cotton, and rice) in large production areas (commercial farms) crop type accuracy is often greater than 90% (USDA-NASS-RDD Spatial Analysis Research Section, 2016). However, (Lark 2017) has identified that CDL has relatively low accuracy in certain specific crops as well as non-cropland covers such as grasslands and cropland fallows. Lark (2017) determined that the fallow class of the 2012 CDL across the CONUS had an overall producer’s accuracy of 34.7% and user’s accuracy of 66.9%. While fallowland statistics are available for some places, it is widely accepted that they have very high uncertainty associated with them, since these fallowland extents/areas have not been systematically studied (Wu et al. 2014). The cropland fallows have varied composition (e.g. barren, sparse grass/shrub cover, dense weeds), produce complex spectral signatures, have widely varying temporal dynamics, and lack field observations and reference data (Wu et al. 2014).

Given the importance, complexity, and uncertainty in mapping cropland fallows, the goal of this study was to develop and implement the automated cropland fallow algorithm (ACFA) with the ability to hind-cast, now-cast, and future-cast cropland fallows over large areas using MODIS 250 m time-series imagery and cloud computing (Didan and Huete 2015; Heck et al. 2019). The process involves gathering refined and accurate training and validation data, stacking MODIS time series imagery, extracting pixel values for training data, coding the knowledge into rule-based ACFA decision trees, assessing accuracies, errors and uncertainties, testing the ACFA model on independent years, and implementing the ACFA model to past years (hind-cast), present year (now-cast), and future years (future- cast) with high degrees of accuracy.

2. Methods

2.1. Study area

The study is the Northern Great Plains section of the United States which includes all of North Dakota, the majority of South Dakota and Montana, and parts of Wyoming, Colorado, Nebraska, and Minnesota (Figure 1). Figure 1 shows the CONUS divided into nine agricultural zones established by the USDA Farm Resources Regions (USDA ERS 2000) which was also used to segment agriculture for CONUS in the Global Food Security-support Analysis Data at 30-meter (GFSAD30) study (Massey et al. 2018; Thenkabail et al. 2021). This region encompasses parts of seven Level III ecoregions (U.S. Environmental Protection Agency 2013) but agricultural land primarily is on the Northern Glaciated Plains, Northwestern Glaciated Plains, Lake Agassiz Plain, and the High Plains ecoregions. Precipitation generally decreases from East to West. Before European settlement, the area was covered with mixed tall grassland with frequent wetlands; afterward, most land suitable for growing crops or raising livestock was converted to agriculture with the notable exception of land controlled by governments for conservation (Taylor et al. 2015).

Figure 1. Study area. The nine broad agroecological zones of the Conterminous United States (CONUS) (USDA ERS 2000). The Northern Great Plans study area is outlined in violet.

Crops in the Northern Great Plains have traditionally been wheat (Spring and Durum) along with hay and pasture. Although it was common practice to fallow fields every other year, now cover crops such as triticale, legumes, and brassicas are frequently planted to increase soil health and reduces soil erosion (Stewart and Thapa 2016; Bourgault et al. 2022). However, wheat yields can decrease in dry years following cover crops due to water they consume, and it is important to kill cover crops before they go to seed (Bourgault et al. 2022). In recent years planting corn and soybean has become more common (Auch et al. 2018). Many studies attribute this increase in corn and soybean to federal subsidies of the ethanol market (Lark, Salmon, and Gibbs 2015) and most have concluded based on CDL data that cropland has expanded to meet this demand (Wright and Wimberly 2013). However, Auch et al. (2018) found that most cropland expansion of corn and wheat was not on native grassland but was instead on farmland used to grow hay and wheat in the 1980s using USDA’s National Agricultural Statistics Service (NASS) Agriculture Census data from 1980 to 2013 (USDA NASS CDL 2022). Table 1 shows cropland areas for the seven states in the Northern Great Plains region derived from the 2019 CDL.

Table 1. Shows the relative cropland area in each state in the Great Northern Plains Agricultural Zone (Figure 1) relative to the total state area derived from the 2019 USDA Cropland Data Layer (CDL) (USDA NASS CDL 2022).

State Name	State Area within Zone^a	Percent state area covered^2	Cropland area covered in zone^b	Percent of state cropland area in zone^c
Colorado	3,250,000	12.03	850,000	23.82
Minnesota	3,270,000	14.97	1,730,000	21.04
Montana	17,880,000	46.95	4,940,000	75.10
Nebraska	7,930,000	39.57	1,200,000	15.29
North Dakota	18,310,000	100.00	9,830,000	100.00
South Dakota	16,930,000	84.74	5,310,000	71.54
Wyoming	7,960,000	31.41	560,000	60.21
Total	75,530,000	44.28	24,420,000	55.01

^a Rounded to nearest 10,000 ha.

^b Based on USDA Cropland Data Layer for 2019, rounded to nearest 10,000 ha.

^c Includes Cropland and Fallowland.

2.2. Satellite imagery

Terra MODIS 16-day EVI image composites (MOD13Q1) Version 6 were used for this study (Heck et al. 2019; Didan and Huete 2015). The images were obtained from Google Earth Engine (GEE) (Gorelick et al. 2017). The MOD13Q1 was chosen as it is a robust standardized vegetation time-series dataset with a 250-meter ground resolution (Didan and Huete 2015). This makes analysis much easier and more repeatable. Terra is less sensitive to clouds than Aqua (the other satellite that has a MODIS sensor) due to its earlier approximate flyover 10:30 am vs. 1:30 pm respectively.

Preliminary testing by the authors and (Wallace et al. 2017) found that EVI produced better classifications than NDVI, which is generally supported by research as EVI includes the Blue band as well as Red and NIR and is less prone to saturation (Walker, de Beurs, and Wynne 2014; Wardlow, Egbert, and Kastens 2007). This study opted to use MODIS data instead of finer spatial resolution imagery such as Landsat due to the difficulty in producing consistent cloud-free image composites (Liang and Gong 2015). MODIS imagery is available from 1999 to the present and the red and near infrared bands have a nominal pixel ground resolution of 250 meters (Román et al. 2024). Harmonized Landsat-Sentinel 2 fused imagery was investigated, but was not used because it is only available for CONUS from 2017 to present (Masek et al. 2021; LP DAAC 2021). From studying crop calendars and referencing the EVI time-series, we decided to restrict the annual time series to a time period that is typically snow free (beginning of March to end of October) (Whitcraft, Becker-Reshef, and Justice 2015). Sixteen MOD13Q1 images were acquired during this period spanning the beginning of March to the end of October (Figure 2). Since 16-day EVI composites were used, there are nominally 2 composites per month; for simplicity here they are labeled M A, or M B, where M is the given month and A and B refer to it being the first or second composite of the month respectively (Figure 2). To further focus on the peak of the growing season, a 17th band was created by summing the five composites from May B to July B to calculate an approximate integral of crop growth per year. Other multi-date composites were examined, but they were highly correlated with EVI May B to July B and were not chosen as factors in classification models. Thus 17 EVI time-series bands were combined to form an Analysis Ready Data (ARD) cube on which the classification was performed (Oliphant et al. 2024).

Figure 2. ARD-cube. Analysis ready data cube (ARD-cube) of the study area composed using MODIS EVI imagery (Oliphant et al. 2024). Note that an 8-band stack is shown, though 17 bands were used in this classification for each year.

2.3. Reference sample generation

Reference data were gathered using USDA CDL. As widely accepted, CDL has high levels of accuracy for the cropland classes but a high degree of uncertainty for the cropland fallow class. Therefore, our approach was to select cropland reference data from the USDA CDL.

For cropland fallows we adopted a critical review of each sample based on: (a) the quality layer in CDL, and (b) very high spatial resolution imagery (VHRI). CDL is independently calculated for each year which makes it resistant to propagating errors over multiple years. When comparing two CDL products that are separated by more than 1 year, (Lark et al. 2017) recommends reviewing CDL products from the years in between because a direct comparison can have an additive impact in classification errors and produce erroneous results. Historically CDL has under-predicted total cropland area; however, estimates have become more accurate over time and CDL reports a greater increase in cropland extent over time compared to other reliable sources (Johnson and Mueller 2021; Lark et al. 2021). For each year between 2010 and 2019, 70,000 random locations, hereafter referred to as samples, were generated within GEE to serve as reference samples for product training and validation. A series of filtering steps defined below were used to reduce those to approximately 6000 samples per year. We reviewed the potential samples based on: (a) location, (b) quality layer in CDL, and (c) verification using very high spatial resolution imagery (VHRI), as follows:

1. Location – Spatial autocorrelation avoidance: Remove samples generated within 1.2 km of another sample. This ensures all samples were over 1 km away from the nearest sample, thus reducing the effect of autocorrelation.

2. Location – Samples located on Cropland reference data: All samples outside the GFSAD30 reference Cropland Extent product, (see (Thenkabail et al. 2021)) were discarded.

3. Location – Samples within homogeneous Areas to avoid mixed pixels: A nearest neighbor analysis was used to ensure that training and validation samples were not placed within 1 MODIS pixel (250 m) of an alternate landcover to avoid mixed pixel signatures (Figure 3).

4. Quality – High confidence samples: All cropland classification layers of USDA CDL produced in 2013 or later have a ‘confidence layer’ which ranges from 0-100. The confidence layer is not an accuracy layer but rather a ‘goodness of fit’ estimate that indicates how confident the classifier was in assigning the particular class. By taking the confidence of a pixel classification in consideration, preliminary research by the authors of this study has shown to improve the user accuracy of a particular class (USDA NASS CDL 2022). Prior experimentation found that pixels with confidence values (CV) under 80 were much less accurately classified than samples with CV over 80 so samples with CV less than 80 were discarded.

5. Verification – Fallowed samples were overlaid on very high-resolution satellite imagery and were examined to remove samples that were clearly not fallow. It is the quality of samples that matter rather than quantity. While adequate sample size is essential and these samples are collected spatially well distributed, it is important to emphasize the high level of confidence in selecting the samples.

6. Reference samples for training and validation split: After reference samples were filtered, they were split into training and validation datasets by randomly dividing the reference samples into 2/3 training and 1/3 validation.

Figure 3. Sample distribution. Image showing random sample distribution across the Northern Great Plains study area (Oliphant et al. 2024).

2.4. ACFA model development using data from a wet, dry, and normal year

Because most of the agriculture in the Northern Great Plains is rainfed, the amount and timing of precipitation have a large impact on yield and planting success or failure (Franzluebbers et al. 2011). Although, weather is difficult to predict, forecasts have become increasingly accurate and useful to inform farmers planting practices (Basso and Liu 2019). For the model to be able to be applied to different years covering a range of temperatures and amounts of precipitation (i.e. flood, normal, and drought years), we selected three years to train the model corresponding to a climatically wet, normal, and dry year. The wet, normal, and dry years were determined based on the Standardized Precipitation Index (SPI) maps across the growing season for every year between 2010–2019 (McKee et al. 1993). This helped us determine which years were wet, normal, and dry based on precipitation for the Northern Great Plains (McKee et al. 1993). As a confirmation for our determinations based on the SPI maps, we also reviewed the weekly National Drought Monitor report to determine which years had the most and least drought across the Northern Great Plains (National Drought Mitigation Center 2015). Additionally, we also reviewed the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Information (NCEI) climate and weather data. NCEI provides archival weather and climate data for the USA and around the globe and has many well regarded monthly and annual summary products including the Annual National Climate Reports which assesses regional climatic conditions based on NOAA’s extensive network of ground and satellite based meteorological measurements (National Centers for Environmental Information 2020).

From the SPI and National Drought Monitor we determined that 2019 was a wet year, 2015 was a normal year, and 2017 was a dry year for the Northern Great Plains and used these years for automated cropland fallow algorithm (ACFA) model development. This assessment was confirmed by the 2019 Annual National Climate Report which stated 2019 was a particularly wet year; it was the wettest year on record for North Dakota, South Dakota, and Minnesota; Nebraska’s precipitation was also well above average (National Centers for Environmental Information 2020). The 2015 and 2012 Annual National Climate Reports also confirmed that 2012 was a near record dry year for the Northern Great Plains and 2015 had a mild climate with typical precipitation and temperature (National Centers for Environmental Information 2020). After selecting the years to train the model, we randomly divided the reference data from 2019, 2015 and 2017 into 3 parts with 2 parts used for model training and 1 part used for model validation. The number of cropland and cropland fallow samples used in the ACFA algorithm is shown in Table 2. Table 3 shows how training samples were filtered.

Table 2. Distribution of samples used in the Automated Cropland Fallowland Algorithm (ACFA) model (Oliphant et al. 2024).

Training	Wet Year	Normal Year	Dry Year	Total
	2019	2015	2017	17, 15, 19
Cropland	3773	3590	3992	11,355
Fallow	203	334	81	618
Total	3976	3924	4073	11,973
Validation	Wet Year	Normal Year	Dry Year	Total
	2019	2015	2017	17, 15, 19
Cropland	1839	1792	1963	5592
Fallow	95	157	53	305
Total	1934	1947	2016	5897

Table 3. Reference training and validation data used in Automated Cropland Fallowland Algorithm (ACFA) model, showing how samples were refined and filtered (Oliphant et al. 2024). GFSAD30 = Global Food Security-support Analysis Data at 30 m.

Class	Full Dataset	GFSAD30	Confidence ≥ 80	Used in Training	Used in Validation
Mixed Pixels	61,432	38,346	18,666
Cropland	30,863	30,005	22,600	11,355	5,592
Fallow	1,647	1,624	1,194	618	305
Non-Cropland	114,548	3,672	2,376
Total	208,490	73,647	44,836	11,973	5,897

2.5. Knowledge-base generation for decision trees

Creating the knowledge-bases of EVI scaled reflectance values that characterizes each class is necessary to ensure cropland fallows can be separated from croplands. Once we establish a clear separability of classes, we can code that into the ACFA models. To achieve this goal, we visualized the MODIS EVI value distribution of the training samples using a series of box and whiskers plots created with ggplot2 v 3.3.5 within R version 4.4.1. (Wickham 2016; R Core Team 2021) and are shown in Figure 4. The four sub-figures in Figure 4 show wet (2019), normal (2015), dry (2017), and combined three years, respectively. Plots show box and whiskers plot which presents the MODIS EVI value for each set of training samples representing the cropland and cropland fallow classes. The black line within the box is the mean value while the lower and upper boundaries of the box are the first and third quartile respectively (Wickham 2016). The vertical lines or whiskers show the samples within 1.5 times the inter-quartile range and the black dots are samples that are beyond that range and are outlying samples (Wickham 2016).

Figure 4. Knowledge-base generation. Knowledge-base generation using (a) wet year 2019, (b) normal year 2015, (c) dry year 2017, (d) combined wet, normal, and dry years (Oliphant et al. 2024). Plot shows bar and whiskers plot that depict the distribution of MODIS pixel EVI values for each set of training samples. The black line within the box is the mean value while the lower and upper boundaries of the box are the first and third quartile respectively. The vertical lines or whiskers show the samples within 1.5 times the inter-quartile range and the black dots are samples that are beyond that range and are outlying samples (Wickham 2016).

2.6. Decision tree (DT) rule-based automated cropland fallowland algorithm (ACFA) model development and testing

Decision Trees (DT) use machine learning to create a series of binary splits in data to categorize data based on a training dataset. Unlike many other algorithms, input data do not need to be normalized, and the resulting models are easy to interpret and implement. The classifications for this study were determined by two simple DT. A frequently used implementation of DTs is random forest classifiers, which create a large number (5-500) of DT and use each to classify every pixel; a majority rule is then applied to produce a final classification of each pixel (Oliphant et al. 2019). Although random forest classifiers tend to produce higher accuracy models than single DT, single DT are superior to Random Forest for research because the trees, are viewable and are easier to understand, making analysis straight forward and more repeatable.

The R package rPart (Therneau et al. 2019) was used to create decision trees for the years 2019, 2015, and 2017 which correspond to wet, normal, and dry climatic years respectively for the Northern Great Plains region (Oliphant et al. 2024). The reference samples from the three years were partitioned, with two-thirds used for model training and one-third used for model validation. The binary splits in the two decision trees were chosen by rPart program based on the training data provided; they were not defined or adjusted by the authors (Therneau et al. 2019; Oliphant et al. 2024). Various optimization parameters were modified to improve the classification results. Ultimately, we used a penalty matrix that prioritized accurate cropland fallow classification 2x over cropland classification for the final models to account for the large number or imbalance of cropland samples compared to fallow samples as shown in Table 3.

Developing mature and robust DT ACFA models (Figures 5(a) and 5(b)), required us to run several iterations of the model by using EVI signatures generated. Inspection of the many DT ACFA interim model runs revealed two models that provided the best results, which are shown in Figure 5(a) and 5(b). If a pixel was classified as cropland fallow by applying either tree (Figures 5(a) and 5(b)) it was classified as cropland fallow in our final classification. Both trees were generated using the same training data. The only difference is that the complexity parameter (CP) or tree complexity was set at 0.5 for Figure 5a and 0.7 for Figure 5b.

Figure 5. Final decision trees. Final decision trees for classifying cropland fallow vs cropland created using time-series MODIS EVI data with a) complexity parameter (CP) of 0.005 and (b) CP of 0.007 (Therneau et al. 2019). White boxes show the composite band at each split level. May2July is the sum of the EVI composites from the middle of May to end of July. Green final nodes are cropland classes and yellow nodes are fallow classes.

2.7. Accuracy assessments and error matrices

The final decision tree ACFA model was tested for its performance using accuracy error matrices. First, the accuracy tests were performed separately on the wet year 2019, normal year 2015, and dry year 2017, for which the ACFA model was developed. Second, the accuracy test was performed for the composite product incorporating all three years: wet, normal, and dry. Error matrices were created for producer’s accuracies (errors of omissions), user’s accuracies (errors of commissions) and overall accuracies. Producer’s accuracies measure errors of omissions (e.g. the fallow class identified in the field is mapped as some other class in the map and thus leads to omission errors), user’s accuracy measures errors of commission (e.g. a planted crop identified in the field is mapped as cropland fallow class and thus leads to commission error). Overall accuracy measures the combined accuracy of all the map classes (Congalton and Green 2019).

2.8. Decision tree (DT) ACFA model validation

After the final Decision Tree (DT) ACFA model was developed and tested (Sections 2.6 and 2.7), the model was applied to the independent years 2010-2014, 2016, and 2018 to see how well it performed. Error matrices were generated for all 6 independent years.

2.9. Area calculations and comparisons

Once the cropland fallow, as well as cropland classes were mapped, their areas were calculated for each year between 2010 and 2019 in Google Earth Engine for the entire Northern Great Plains study area (Gorelick et al. 2017). The cropland fallow areas predicted were compared to those in the respective years’ CDL products.

3. Results

3.1. Cropland fallow products for 2010–2019 including those of 7 independent years

The final DT ACFA model was used to map cropland and cropland fallow for every year between 2010 and 2019 in the Northern Great Plains region. This included the years of model development: (a) wet year 2019, (b) normal year 2015, (c) dry year 2017. The products were also developed for 7 independent years: 2010-2014, 2016, 2018. The same two decision trees were applied to classify every year, regardless of their climatic condition. We have illustrated outputs for three independent years (Figure 6): (a) wet year 2018 (Figure 6(a)), (b) normal year 2014 (Figure 6(b)), and (d) dry year 2012 (Figure 6(c)). Each year had croplands and cropland fallows.

Figure 6. Cropland fallows maps. Cropland fallows mapped using the best decision tree algorithm for the independent (a) wet year 2018, (b) normal year 2014, and (c) dry year 2012 (Oliphant et al. 2024).

3.2. Accuracy assessments

3.2.1. Accuracies for the modeled years

The DT ACFA generated cropland and cropland fallow products were tested for accuracies using reference validation data. The accuracy error matrices for the modeled years (wet year 2019, normal year 2015, dry year 2017) are presented in Table 4. The overall accuracies were all above 97%. For the cropland fallow class, the producer’s and user’s accuracies varied between 71% to 83% (Table 4).

Table 4. Error matrices of the training and testing data for the final DT ACFA model (Oliphant et al. 2024). Reference samples generated for all years are combined, i.e. 2019, 2015, and 2017 (wet, normal and dry years).

Training	Cropland	Fallow	Row total	Commission error
Cropland	11,203	104	11,307	0.9%
Fallow	152	514	666	22.8%
Column total	11,355	618	11,973
Omission error	1.3%	16.8%
Producer accuracy	98.7%	83.2%
User accuracy	99.1%	77.2%
Overall accuracy				97.9%
Validation	Cropland	Fallow	Row total	Commission error
	5498	69	5567	12%
Fallow	94	236	330	28.5%
Column total	5592	305	5897
Omission error	1.7%	22.6%
Producer accuracy	98.3%	77.4%
User accuracy	98.8%	71.5%
Overall accuracy				97.3%

3.2.2. Accuracy assessments for independent years

The model accuracy was tested on 7 independent years: 2010-2014, 2016, and 2018. For the cropland class, the producer’s and user’s accuracies varied between 97-99% for the 7 years (Table 5). For the cropland fallow class, the producer’s and user’s accuracies varied between 62-84% and overall accuracies were between 95-98% (Table 5). These cropland fallow accuracies are among the highest in the literature, especially when evaluated over such large areas and multiple years. These results clearly indicate the model achieved a high degree of accuracy in mapping cropland fallow class over large areas. Further improvements in mapping cropland fallows are possible if more field plot data are collected and used to train the model. Cropland fallows are complex and vary in signature; they can be intentionally fallowed and managed to remain barren or they can be abandoned and become weedy.

Table 5. Error matrices showing independent samples for years 2010 through 2018 (Oliphant et al. 2024). Note: since 2019, 2015, and 2017 were used for model training, their accuracies were combined and shown in Table 4b. Acc. = Accuracy, Comm. = Commission.

2010	Cropland	Fallow	Row total	Comm. error		2011	Cropland	Fallow	Row total	Comm. error
Cropland	3813	114	3927	2.9%		Cropland	4053	98	4151	2.40%
Fallow	47	323	370	12.7%		Fallow	83	324	407	20.40%
Column total	3860	437	4297			Column total	4136	422	4558
Omission error	1.2%	26.1%				Omission error	2.0%	23.2%
Producer Acc.	98.8%	73.9%				Producer Acc.	98.0%	76.8%
User Acc.	97.1%	87.3%				User Acc.	97.6%	79.6%
Overall Acc.				96.3%		Overall Acc.				96.0%
2012	Cropland	Fallow	Row total	Comm. error		2013	Cropland	Fallow	Row total	Comm. error
Cropland	5196	116	5312	2.2%		Cropland	4702	91	4793	1.9%
Fallow	131	309	440	29.8%		Fallow	111	401	512	21.7%
Column total	5327	425	5752			Column total	4813	492	5305
Omission error	2.5%	27.3%				Omission error	2.3%	18.5%
Producer Acc.	97.5%	72.7%				Producer Acc.	97.7%	81.5%
User Acc.	97.8%	70.2%				User Acc.	98.1%	78.3%
Overall Acc.				95.7%		Overall Acc.				96.2%
2014	Cropland	Fallow	Row total	Comm. error		2016	Cropland	Fallow	Row total	Comm. error
Cropland	5645	96	5741	1.7%		Cropland	6268	75	6343	1.2%
Fallow	88	396	484	18.2%		Fallow	81	328	409	19.8%
Column total	5733	492	6225			Column total	6349	403	6752
Omission error	1.5%	19.5%				Omission error	1.3%	18.6%
Producer Acc.	98.5%	80.5%				Producer Acc.	98.7%	81.4%
User Acc.	98.3%	81.8%				User Acc.	98.8%	80.2%
Overall Acc.				97.0%		Overall Acc.				97.7%
2018	Cropland	Fallow	Row total	Comm. error
Cropland	6639	41	6680	0.6%
Fallow	39	96	135	28.9%
Column total	6678	137	6815
Omission error	0.6%	29.9%
Producer Acc.	99.4%	70.1%
User Acc.	99.4%	71.1%
Overall Acc.				98.8%

3.3. Area calculations

Mapping cropland fallows and croplands over several years, as done in this study, will allow the calculation of cropland areas as well as cropland fallow areas. These estimates are important for assessing crop productivity, water used or saved, and for a host of other purposes such as determining compliance with government mandates on retaining a certain percentage of land as cropland fallows or for assessing crop water productivity (‘crop per drop’) (Foley et al. 2020). The US Census Topologically Integrated Geographic Encoding and Referencing system (TIGER) County shapefiles were used to calculate cropland and cropland fallow areas for the counties within the Northern Great Plains region (U.S. Census Bureau 2017). Since the Economic Research Service (ERS) zone boundaries fall along county boundaries, there were no issues associated with partial counties in the Northern Great Plains region. Table 6 shows the area of croplands and cropland fallows for the portion of each state within the Northern Great Plains region for the years 2010 through 2019.

Table 6. Areas in ha calculated for cropland and cropland fallows from 2010 to 2019 for each state in the Northern Great Plains region. Crop is cropland area in ha, Fallow is cropland fallow area in ha, and % is the percent fallow / crop + fallow (Oliphant et al. 2024).

State	Crop (ha)	2010 Fallow (ha)	%	Crop (ha)	2011 Fallow (ha)	%	Crop (ha)	2012 Fallow (ha)	%
CO	689,500	220,500	24.2	648,100	261,900	28.8	467,500	442,500	48.6
MN	1,959,800	8,300	0.42	1,926,300	29,500	1.5	1,965,900	3,300	0.26
MT	3,501,300	1,068,300	23.4	3,458,100	1,162,200	25.2	3,573,800	1,057,800	22.8
NE	1,150,200	305,700	21	1,165,400	290,300	19.9	917,800	538,000	37
ND	11,562,600	189,300	1.6	10,631,500	1,093,600	9.3	11,642,600	141,600	1.2
SD	4,974,500	278,600	5.3	5,098,600	156,600	3	4,974,800	281,500	5.4
WY	325,600	99,200	23.4	305,500	119,500	28.1	194,600	231,400	54.3
Total	24,163,400	2,169,900	8.2	23,233,400	3,113,500	11.8	23,736,900	2,696,000	10.2
State	Crop (ha)	2013 Fallow (ha)	%	Crop (ha)	2014 Fallow (ha)	%	Crop (ha)	2015 Fallow (ha)	%
CO	556,100	353,100	38.8	721,600	188,300	20.7	703,100	206,900	22.7
MN	1,926,300	32,400	1.7	1,901,900	66,200	3.4	1,954,100	15,700	0.8
MT	3,323,400	1,306,200	28.2	3,397,800	1,231,400	26.6	3,397,400	1,242,100	26.8
NE	1,067,100	384,300	26.5	1,248,600	207,000	14.2	1,191,700	264,100	18.1
ND	11,005,700	730,200	6.2	11,520,000	255,700	2.2	11,612,600	178,300	1.5
SD	5,042,600	146,300	2.8	5,169,300	86,100	1.6	5,087,100	169,100	3.2
WY	236,900	177,100	42.8	354,700	70,400	16.6	293,100	132,100	31.1
Total	23,158,000	3,129,500	11.9	24,314,000	2,105,100	8	24,239,000	2,208,200	8.3
State	Crop (ha)	2016 Fallow (ha)	%	Crop (ha)	2017 Fallow (ha)	%	Crop (ha)	2018 Fallow (ha)	%
CO	703,100	206,900	22.7	652,500	257,500	28.3	681,000	229,000	25.2
MN	1,954,100	15,700	0.8	1,966,400	2,700	0.14	1,968,700	2,200	0.11
MT	3,397,400	1,242,100	26.8	3,341,400	1,297,800	28	3,964,100	677,200	14.6
NE	1,191,700	264,100	18.1	1,148,900	306,900	21.1	1,280,000	175,800	12.1
ND	11,612,600	178,300	1.5	11,560,900	228,400	1.9	11,747,300	46,300	0.39
SD	5,087,100	169,100	3.2	5,046,200	210,700	4	5,176,100	81,200	1.5
WY	293,100	132,100	31.1	300,600	125,200	29.4	342,900	83,100	19.5
Total	24,239,000	2,208,200	8.3	24,016,900	2,429,100	9.2	25,160,100	1,294,900	4.9
State	Crop (ha)	2019 Fallow (ha)	%
CO	746,000	164,000	18.0
MN	1,947,200	7,900	0.4
MT	3,963,500	675,000	14.6
NE	1,311,200	144,600	9.9
ND	11,665,400	116,200	0.99
SD	5,079,300	177,400	3.4
WY	347,900	78,100	18.3
Total	25,060,600	1,363,000	5.2

3.4. Area comparisons

After cropland and fallow areas were calculated at the state and county level for each year, the ratio of cropland fallow to total cropland was calculated for each year to determine how much cropland fallow fluctuated by year. Table 6 shows the percent cropland fallow of the total cropland area. The percent cropland fallow area of the total cropland area for each year and state is shown in Figures 7(a) and 7(c). Note that the states are divided between two sub-figures to highlight the difference in percent cropland fallow.

Figure 7. Percentage of cropland fallow compared to precipitation per year. Bar charts (a) and (c) that show percentage of cropland fallow by state for each year in relation to total cropland i.e. %fallow = [fallow/ (cropland + fallow)]. Note the range of y axis in Figure 7a is from 0-50% while the y axis on Figure 7b is 0-10.%. Bar charts (b) and (d) showing the precipitation over the growing season from the NOAA National Centers for Environmental Information for each state in the Northern Great Plains (National Centers for Environmental Information 2020).

Additionally, to determine the interaction between precipitation and cropland fallow area, the total precipitation over the growing season for each county in the Northern Great Plains was obtained from NOAA National Centers for Environmental Information for the years between 2010–2019 (LP DAAC 2023) and is shown in 7b and 7d.

Since 250 m mapping can be used to calculate cropland fallow areas at the county level, we calculated the fallow areas in the CDL and produced maps at the county level. This is shown in a scatterplot in Figure 8 for the year 2014. The cropland and cropland fallow maps for each year from 2010 through 2019 are available in the USGS ScienceBase (Oliphant et al. 2024).

Figure 8. CDL comparison with ACFA. Scatterplot in log 10 scale comparing the cropland fallow area by county for the year 2014 between the Cropland Data Layer (USDA NASS CDL 2022) and the automated cropland fallows algorithm (ACFA) (Oliphant et al. 2024). One county per state which falls outside the linear relationship between CDL and ACFA are labeled.

4. Discussion

Although the CDL product is considered the gold standard for crop type mapping for the United States, it does have three shortcomings: 1. The classification algorithm is not available to the user: 2. The cropland fallows accuracy is relatively low: 3. There is a delay in releasing the products. Issues 1 and 2 were covered in depth in the introduction, we will now address issue 3. CDL is released well after the cropping season is over, around February of the next year mainly because USDA does not want the CDL to impact agricultural commodity prices (Boryan et al. 2011). Following that, due to the potential impact on commodity prices, the algorithms used to create CDL each year will not be publicly released which is problematic for open research and limits its usability in applying the algorithms in other areas outside CONUS (Boryan et al. 2011). This study has created an algorithm that has been applied over 10 years and has produced relatively consistent accuracies across that time-period which suggest it can be applied to years before 2010 and after 2019 (Oliphant et al. 2024).

Mapping cropland fallows is known to be difficult, partially out of the difficulty of definition and partially out of difficulty in distinguishing it from other land cover. In areas with multiple cropping seasons, defining and mapping cropland fallows is even more difficult. In areas with a long growing season, it is common to plant a crop during the wet season and occasionally plant a crop in the dry season if adequate water is available and it is economically feasible to do so. In such areas, years where only one crop was planted could be considered fallow years because there was a fallow season. Cropland fallow fields are also difficult to identify spectrally. Weeds have been found to be a contributing factor in difficulty in mapping cropland fallows (Ortiz-Monasterio and Lobell 2007). Research suggests that no-till farmed fields may be more difficult to distinguish from cropped and fallow fields due to a lack of disturbance from tilling soil and the greater presence of weeds (Thorp and Tian 2004).

Another confusing definition is the separation between temporary cropland fallow fields and long-term fallow or abandoned land. The length of time a field is left fallow before it is considered abandoned varies. A related land cover class that can be confused with cropland fallow is USDA Farm Service Agency (FSA) Conservation Reserve Program (CRP), where the USDA pays farmers rent not to plant and instead return environmentally sensitive land to a more natural state for 10–15 years (Farm Service Agency 2022). Since land enrolled in CRP must be planted for four of six crop years and physically and legally capable of being planted (Farm Service Agency 2022), there is nothing distinguishing CRP land from cropland fallow during the first few years of enrollment. Since the USDA does not release a raster or vector map of the locations of the land enrolled in CRP, additional filtering to flag CRP land is not possible.

Additional uncertainty in classification can be from the classification method chosen. This work chose to use two simple decision trees since they would be easy to record and reproduce. Vieira et al. 2021 mapped land degradation across half of the Cerrado biome of Brazil using an NDVI time-series. They found that a decision tree algorithm was able to produce detailed and accurate estimates of land degradation and was resistant to drought signature which can cause false positive indication of degradation (Vieira et al. 2021). However, decision trees are not without faults. One downside in using a supervised classifier to map a rare class is that rare classes are frequently mismapped. Most classifiers, decision trees included, prioritize overall classification accuracy. In instances where the class of interest occurs over a small area, the classifier will often poorly classify the rare class. To improve classification results, standard model optimization parameters were adjusted. Penalty matrices (PM), (which adjusts what class is given priority when the model decides splitting variables and values) and priors (which adjusts the ratio of different classes) were assessed (Kuhn and Johnson 2013). Through inspection of many model runs, it was decided that a PM with cropland fallows assigned 2x class priority performed better than other PM ratios and priors. More advanced classifiers could be utilized which should produce superior classifications, but at the expense of transparency, reportability, and reproducibility.

Single or a few decision trees are also limited in their ability to separate very intermixed data. Results from (Bofana et al. 2020) indicate that random forest and Support Vector Machines (SVM) have a 10.5% and 1.3% greater accuracy than a single Classification And Regression Tree (CART) decision tree. However, large numbers of decision trees are difficult to report and reproduce. Probst and Boulesteix (2018) found after examining 44 datasets that random forest accuracy did not appreciably improve after 250 trees which is significantly more than two trees used here. SVM are another popular supervised classification algorithm used in remote sensing and usually achieve similar classification accuracies as RF after parameter tuning. It is possible to save SVM models using libraries such as scikit-learn, however the saved models are not human readable and require the user to use the exact library version that was used to train the model to re-run the model (Pedregosa et al. 2011). Raczko and Zagajewski (2017) found that artificial neural networks classified tree species with accuracies over 10% higher than SVM or RF but required much more computation time to do so. Pandey et al. 2021 has an excellent review of different classification algorithms available and describes the appropriateness of choosing each and the emergence of spatial-spectral classifiers which can improve classification accuracy by incorporating spectral and spatial information. However, implementing more advanced classifiers in the future could likely improve the classification of cropland fallows.

The greatest challenge in mapping cropland fallows is the lack of robust training and validation data. The primary source of reference data for this study was CDL. Although CDL does have high accuracies for large cash crops, the accuracy for cropland fallows is lower. The methods used to screen CDL data such as restricting data to 80% confidence and only picking samples in the center of fields help, but some uncertainty exists. For years when very high-resolution imagery (VHRI) was available, fallow sampled fields were visually compared to VHRI usually from imagery collected by aircraft. However, even with such imagery it is difficult to distinguish based on a single date if a field was fallow or cropped, so cropland fallow fields were retained unless there were clear indications that a field was planted.

The release of Planet Labs data to all federal researchers through a blanket NASA contract will enable researchers to access near-daily 3-meter resolution 4 band (RGB-NIR) imagery (NASA 2024). This will enable one to view fields throughout a season to discern signs of fields being plowed, growing, flowering, and fruiting (for select species), being harvested, and under post-harvest land treatment. This should greatly aid in the ability to verify if a field was truly fallow and if not, what crop was grown.

In Figure 8, a high degree of linearity is observed between cropland and fallowland areas per county for 2014 in the CDL and ACFA maps; however significant outliers are present. One way to observe why this is per county is to compare the areas where ACFA and CDL disagree to National Aerial Imagery Program (NAIP) data (Aerial Photography Field Office 2021). NAIP is publicly available true color VHRI data collected by contracted aircraft for the USDA with a 1 meter pixel resolution, usually every other year. The states of Nebraska, North Dakota, and South Dakota were collected in 2014 and Colorado, Minnesota, Montana, and Wyoming were collected in 2013 and 2015. Since there are 176 counties in the Northern Great Plains, it is not feasible to discuss this comparison for each county; however, we will do so for the county in each state that had a large percentage difference between CDL and ACFA for cropland fallow area. These counties are labeled by name and leader lines in Figure 8. In Dawes Nebraska, a significant number of fields were classified as cropland in ACFA and cropland fallow in CDL; compared to the 2014 NAIP, these areas appear to be sand colored or splotchy green which indicate they were recently plowed/chemical herbicide applied or natural revegetation/volunteer vegetation is taking root respectively. In Richland North Dakota, a significant number of fields were classified as cropland fallow in ACFA and cropland in CDL; these areas appear to be dark brown or tan in 2014 NAIP, which indicates fields were recently plowed or present vegetation is dead. In Clark South Dakota, a significant number of fields were classified as cropland fallow in ACFA and non-cropland in CDL; compared to the 2014 NAIP, these areas are primarily cover small lakes/marshes due to visible water. In Jefferson Colorado, there is a small area that is mapped as cropland fallow in ACFA and cropland in CDL; according to the 2013 and 2015 NAIP imagery, these are areas near crop fields and suburban areas and have been cleared for suburban development. In Polk Minnesota, there is a significant number of fields that are mapped as cropland fallow in ACFA that are cropland in CDL; compared to the 2013 and 2015 NAIP, these appear to be green with trails or brown which indicates they are pasture or bare soil fields respectively. In Chouteau Montana, there is good agreement between ACFA and CDL for cropland fallow mapping. There is about equal area that is mapped as cropland fallow in ACFA and cropland in CDL as the opposite since these areas are on active fields by observing 2013 and 2015 NAIP and no NAIP imagery is available for 2014, it is difficult to draw conclusions why this is the case. In Johnson Wyoming, there is a small amount of area that is mapped as cropland fallow or cropland in ACFA that are non-cropland in CDL; comparing these areas to 2015 NAIP imagery, they appear as bright green areas in river valleys which appear to be natural vegetation.

As suggested in the introduction section, the variability of precipitation between years impacts the area of land that is fallow. We hypothesize that in areas with significant amount of rainfed cropland, such as the Northern Great Plains, years with less precipitation will have a greater area of cropland fallow than years that have more precipitation. In Figure 7, we generally found that to be the case with 2012, the driest year in for 6 of the 7 states over the area examined, having a higher percentage of cropland fallow area than average for 5 of the 7 states. For 2019, the wettest year for 6 of the 7 states, had less percentage of cropland fallow than average for 6 of the 7 states. A further examination of this relationship is beyond the scope of this study.

5. Conclusion

Cropland fallows of the Northern Great Plains of the United States were mapped using a newly developed decision tree-based (DT-based) automated cropland fallows algorithm (ACFA) model on Google Earth Engine (GEE) cloud utilizing MODIS 250 m time-series data. The model was developed using 3 years (2019, 2015, 2017) that represent a wet, normal, and dry year, respectively. The ACFA model was then applied to 7 independent years and evaluated for accuracy. For the cropland class, the producer’s accuracies varied between 97.5-99.4% and the user’s accuracies varied between 97.1% to 99.4% (Table 5). For the cropland fallow class, the producer’s accuracies varied between 70.1-81.5% (errors of omissions 18.5-29.9%) and User's accuracies varied between 70.2-87.3% (errors of commissions 12.7-29.8%). Overall accuracies were between 95.7-98.8% (Table 5). Further improvement in cropland fallow mapping accuracies are possible by: 1. improving reference training and validation data accuracies through field data collection, and 2. utilizing higher resolution time-series remote sensing data such as Landsat and Sentinel. The ACFA model successfully mapped cropland fallows over large areas with very high levels of accuracies (overall, producer’s and user’s) that are hitherto not achieved in the literature. The ACFA model also clearly demonstrated the ability to hind-cast and future-cast cropland fallow and cropland mapping. The products accurately estimated cropland and cropland fallow landcover over large areas on a yearly basis which are valuable in multiple applications such as assessing crop water use and food security assessment by determining areas cropped versus areas left fallow. Further ACFA models run on the GEE have the potential to be applied on other regions by modifying thresholds based on region specific training data.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data used and generated in this publication is available from the USGS ScienceBase (Oliphant et al. 2024).

Additional information

Funding

This research was supported by the U.S. Geological Survey (USGS) National Land Imaging program of the Core Science Systems Mission Area of the USGS. The major research support for this project came from the water SMART (Sustain and Manage America's Resources for Tomorrow) project funded by the Department of Interior (DOI) through USGS. The research was conducted in the science facilities of the USGS Western Geographic Science Center (WGSC). Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.