Mapping Zero Tillage Adoption in The Indo-Gangetic Plains of India

You might also like

Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 18

Mapping zero tillage adoption in the Indo-Gangetic Plains of India

Abstract
In recent decades, the zero tillage (ZT) technique has been highly promoted in the rice-
wheat systems across the Indo-Gangetic Plains (IGP) for its ability to improve yields and reduce
production costs. However, separating ZT and conventional tillage (CT) and estimating the ZT
adoption areas for the entire IGP is challenging. In this study, we created multi-state and multi-
year generalized classification models using the data from high spatiotemporal resolution
satellite imagery, Sentinel-2, and the random forest machine learning algorithm. Based on the
field data across the IGP in 2018, the classification accuracy of the model generated with the
imagery composites from a 3-month sowing time period is 75.51% for the entire IGP with
higher accuracy in western IGP (~80%) and lower accuracy in eastern IGP (~70%). We also found
that the sowing time period is more critical than the peak time period when mapping the ZT
adoptions. In addition, the classification accuracy will decrease with the reduction of training
field data, but the accuracy only dropped 3% when we reduce the number of training polygons
by 75%. Furthermore, the classification model that used multi-year and multi-state field data in
the training data set could be applied to the previous years’ data, hence the compatibility of the
model was highly improved. The results of this study demonstrate that using the high
spatiotemporal resolution satellite imagery, ZT adopted areas can be reliably estimated across
years and across states in the smallholder farming systems of the entire IGP of India.

Keywords:
Sentinel-2, zero tillage, random forest, google earth engine
1. Introduction
Tillage practices, the methods of plowing soil for seeding, are essential for agricultural
areas’ environmental outcomes and productivity (Hagen et al., 2016). Zero tillage (ZT), one of
the conservation tillage methods, is planting the seed in the existing crop residue. This
technique has the ability to reduce soil and water loss and control greenhouse gas emissions
through retaining the organic matters in the agricultural soil (Gebhardt et al., 1985; Horowitz et
al., 2010). In contrast to the conventional tillage (CT), ZT has the potential to increase farmers’
income and benefit-cost ratio through lowered costs of production in human labor, machine
labor, and irrigation (Prasad et al., 2018; Tripathi et al., 2013). In recent decades, the studies in
the economic and productive benefits of zero tillage adoption is spreading all over the world,
including the U.S., Brazil, and India (de Freitas and Landers, 2014; Erenstein, 2009; Erenstein et
al., 2008; Horowitz et al., 2010; Pandit et al., 2010). Giving the increase of ZT adoption, it is
important to identifying ZT from CT and estimating the ZT adopted areas. Remote sensing
approaches may offer a way to effectively mapping ZT adoption, especially across large
spatiotemporal scales.
Multiple studies have attempted to use remote sensing imagery and machine learning
algorithms to identify the zero tillage practices and map crop residue cover (CRC) in the
agricultural areas. Zheng et al. (2013) used Landsat Thematic Mapper (TM) in their study of
tillage practices in the U.S. and found that the minimum normalized difference tillage index
(NDTI) of the planting season is an effective indicator of crop residual cover. Jin et al. (2015) had
consistent results with Zheng et al. and further proposed that three other vegetation indices –
simple tillage index (STI), normalized difference index 7 (NDI7), and shortwave red normalized
difference index (SRNDI) – were also vital for their study of maize residue cover in northeast
China. Azzari et al. (2019) used the combination of Landsat 5, 7, 8, and Sentinel-1 to map tillage
practices in 10 states in the U.S. from 2005 to 2016. Their best model has accuracies around
75% when mapping the entire north-central U.S. region.
Despite the feasibility of mapping zero tillage, there are two challenges of mapping ZT
adoption in the smallholder farming system. First, the size of the agricultural plots in the
smallholder farming system required high spatial resolution remote sensing images in the
analysis. The farm size of smallholders varies from 0.5 to 4.3 ha (Alexandratos, 1995) and
specifically in India, the average size of holdings is 1.15 ha (DAC&FW, 2015). Converted to pixel-
level, this is an average of 12 pixels per agriculture plot from the Landsat imagery which is the
most common data source for the previous research. This data amount is not ideal to represent
an agricultural field, but until the launch of Sentinel-2 in 2015, there isn’t high spatial resolution
imagery that was opened for research globally. Sentinel-2 has a spatial resolution of 10m that
means for the same area, it could provide 9 times as much data as Landsat. Thus, in this study,
we used the Sentinel-2 imagery to overcome the challenge of small farm size.
The second challenge is the computational complexity of the remote sensing data
processing for the study of large spatiotemporal mapping. One Sentinel-2 image, which can
cover a million ha on the ground, has a data volume of 600 MB. Therefore, when the research
area is large and the time span is long, the remote sensing image acquisition will be large and
image download, storage, and processing will face challenges. To overcome this issue, we used
the cloud-based platform Google Earth Engine (GEE; Gorelick et al., 2017) in this study. Most of
the image processing and analysis steps could be processed on the cloud server, so there is no
need to download and store all the satellite images on the personal computer. Besides, the
computing speed on the cloud server is much faster than the personal computer.
The aim of this study is to fill in the gap in mapping ZT adoption in the smallholder
farming systems at a large spatial scale using high spatiotemporal resolution satellite imagery
from the Sentinel-2 mission. Thus, the specific research questions and objectives of this study
are: 1) develop a multi-state generalized model that can classify zero tillage and conventional
tillage area accurately across the four states – Punjab, Haryana, Uttar Pradesh, and Bihar – in
the IGP region, 2) determine the critical time period for identifying the zero tillage adopted
areas, 3) detect the accuracy change when reducing the amount of field data used for training
the multi-state generalized model, and 4) generate a multi-year generalized model that can be
applied to identify ZT areas in previous years.

2. Method
2.1. Study area
Our study area is in the IGP region (Figure 1) which has a high concentration of the 10.3
million hectares rice-wheat systems in India (Timsina and Connor, 2001). In the IGP, the
growing season for wheat is from November to March and the growing season for rice is from
June to October. The ZT was applied to the wheat crop in the rice-wheat systems as a
technology of resource conservation. The ZT was first been adopted in the northwest IGP
(Punjab, Haryana, western Uttar Pradesh) in the 1990s and then been picked up in eastern IGP
(eastern Uttar Pradesh, Bihar) in the 2000s. Since then, the estimated area of ZT adoption and
the number of ZT machinery sold in the IGP have been growing steadily (Erenstein and Laxmi,
2008).

2.2. Field data


The field data of tillage before the wheat planting season was collected by the
International Maize and Wheat Improvement Center (CIMMYT) for three years – 2016, 2017,
and 2018. The data was collected from all four states in the IGP in 2018, whereas the 2017 data
is only from Uttar Pradesh and Bihar, and the 2016 data is only from Punjab and Haryana. The
spatial distribution of the field polygons is concentrated in two clusters – western IGP (Punjab
and Haryana) and eastern IGP (eastern Uttar Pradesh and Bihar) (Figure 1). The data in Uttar
Pradesh is only from the east part, and there is no field data in western Uttar Pradesh. The
ground truth data simply labeled whether the zero tillage technique is adopted in the
agricultural field or not, whereas the amount or percentage of crop residue cover in the field
has not been recorded.
The field data were collected in three different years and the formats of field data are
slightly different. In 2018, the field team visited the agriculture field and for each agricultural
field, they collected 5 GPS points, 4 points on the corners and a point in the center. For the
2017 season, the locations of the fields were reported by the farmers. The farmers could record
the GPS location of their field when they stand in the middle of the field with an APP on the
phone. The field data in 2016 was provided in the form of polygons. For the data from 2018 and
2017, we digitized the field polygons based on the GPS points. We used the code in R (R
Development Core Team, 2019) to automate the process of digitization for the 2018 data, and
manually created the polygons for the 2017 data. All the field polygons were double-checked
using the high-resolution images in Google Earth Pro before the analysis.
The adoption of zero tillage is still a gradual process in the IGP, which means that the
area of conventional tillage agricultural fields is much more than that of zero tillage adopted
fields. To avoid data from conventional tillage areas dominating the machine learning
classification models, a balanced number of agricultural polygons were randomly selected from
both types of tillage. Table 1 shows the number of field polygons that are used in the study
after the random selection. The total number of field polygons collected in 2018 is 2830 with
1415 CT and 1415 ZT, which is more than that in 2016 (372 polygons) and 2017 (192 polygons),
and there is more data from Punjab and Haryana than from Uttar Pradesh and Bihar.
Figure 1. Map of the study area and field data

Table 1. Number of field polygons used from each year and each state
Year 2018 2017 2016
CT ZT CT ZT CT ZT
State Total Total Total
polygons polygons polygons polygons polygons polygons
Punjab 368 368 736 0 0 0 119 119 238
Haryana 407 407 814 0 0 0 67 67 134
Uttar Pradesh 251 251 502 36 36 72 0 0 0
Bihar 389 389 778 60 60 120 0 0 0
Total 1415 1415 2830 96 96 192 186 186 372

2.3. Sentinel-2 imagery


2.3.1. Image pre-processing
The satellite imagery used in this study is from the Sentinel-2 mission of the European
Space Agency (ESA, 2015). The Sentinel-2 imagery has 13 bands with spatial resolutions ranging
from 10 to 60 meters. In this study, all the bands are resampled to a spatial resolution of 10
meters per pixel for consistency. Bands 1 (Aerosols), 9 (Water Vapor), and 10 (Cirrus) were
excluded from the input features for creating the models because the data from these three
bands reflect the atmospheric conditions rather than the reflectance of the surface, and they
have a relatively coarse resolution of 60 meters. Besides the original bands from Sentinel-2,
seven vegetation indices – the green chlorophyll vegetation index (GCVI; Gitelson et al., 2003),
normalized difference vegetation index (NDVI; Carlson and Ripley, 1997), NDTI (Van Deventer et
al., 1997), normalized difference index 5 (NDI5; Mcnairn and Protz, 1993), NDI7 (Mcnairn and
Protz, 1993), CRC (Sullivan et al., 2006), and STI (Van Deventer et al., 1997) – were also been
calculated during the image pre-processing (Table 2). These vegetation indices have been used
in previous research of identifying the zero tillage adopted areas in the north-central U.S.
(Azzari et al., 2019).
The Sentinel-2 imagery is available on the Google Earth Engine platform. However, the
Level-2A data, which is the bottom-of-atmosphere reflectance product, only available after
December 2018. The Sentinel-2 product available on GEE prior to that date is the Level-1C data
which is the top-of-atmosphere reflectance in cartographic geometry. To avoid the
inconsistency from the imagery surface reflectance correction process, all the Sentinel-2 images
acquired from GEE are the Level-1C product, which then was transformed to bottom-of-
atmosphere reflectance using the GEE Python API with the code adapted from Murphy (2018).
The core of his repository was based on the Py6S which is an interface to the Second Simulation
of the Satellite Signal in the Solar Spectrum (6S) atmospheric Radiative Transfer Model through
Python (Wilson, 2013). Besides surface reflectance correction, a cloud mask was applied to all
the images during the image pre-processing. The cloud mask was adapted from the cloud
scoring algorithm for Landsat (ee.Algorithms.Landsat.simpleCloudScore()) on GEE. The cloud
mask could calculate a score of relative cloudiness from 0 to 100 for Sentinel-2 pixels and then
masked the pixels with the score greater than a threshold. Besides, the cloud mask also
integrated the QA60 bands, which is a bitmask band with cloud information from Sentinel-2
(ESA, 2015). We tried different values of the cloud score and finally took 26 as the most suitable
value for the IGP region. All the vegetation indices in Table 2 were calculated after applying the
cloud mask and the surface reflectance correction.

Table 2. Bands and Indices used in the classification


Band Description Index Description
B2 Blue GCVI (NIR/Green)-1
B3 Green NDVI (NIR−Red)/(Red+NIR)
B4 Red NDTI (SWIR1−SWIR2)/(SWIR1+SWIR2)
B5 Red Edge 1 NDI5 (NIR−SWIR1)/(NIR+SWIR1)
B6 Red Edge 2 NDI7 (NIR−SWIR2)/(NIR+SWIR2)
B7 Red Edge 3 CRC (SWIR1−Green)/(SWIR1+Green)
B8 Near-infrared (NIR) STI SWIR1/SWIR1
B8A Red Edge 4
B11 Short-wavelength Infrared 1 (SWIR1)
B12 Short-wavelength Infrared 2 (SWIR2)

2.3.2. Composite method


Though the application of cloud mask could reduce the noise of clouds, it brought up
the problem of missing pixels. Even though the winter growing season in IGP is not the time of
the year for heavy precipitation, the hazy and cloudy pixels caused massive data gaps,
especially in January and February. To solve the problem of missing pixels, we created the
multi-date image composites. Another reason for using a multi-date image composite is that
the spatial extent of the study area is large. The total area of the four states in the IGP is 432
thousand square kilometers and the study area goes east-west, so the Sentinel-2 satellites
cannot cover the entire study area with a single path. Therefore, in each repeat cycle, the
revisit dates of the satellite images for the adjacent area would not be exactly the same.
Instead, they would be within a small time interval. So the image of a particular day for a
western area may not have a corresponding image for an eastern area on the same day. Thus,
creating a multi-date image composite would ensure that the remote sensing data used for
classification were captured from the same time period for any part of the study area.
The determination of the image composite time range was based on not only the
minimum time required to piece together a whole image that could cover the IGP without
missing pixels but also the agricultural phenology of winter wheat. The winter growing season
in the IGP is generally from November to April (Pathak et al., 2003), but the season varies by
state. The sowing date gradually delayed from western to eastern IGP (Figure 2). Since the
tilling happens before planting seeds, a three-month sowing season composite was generated
to capture all the surface reflectance around the sowing time period. In addition, a
corresponding three-month peak season composite was created for the growing time period.
According to the wheat phenology curve, the sowing season is from October 15 th to January
15th, and the peak season is from January 15th to April 15th.

Figure 2. Phenology curve of wheat in the 2018 winter growing season for each state

The approach for generating the seasonal composite is based on the distribution of pixel
values. For both sowing and peak seasons, five percentiles were used to represent the
distribution of each band and vegetation index of the season – the 0 th, 25th, 50th, 75th, and 100th
percentiles. Considering that 10 original bands from Sentinel-2 and 7 vegetation indices were
extracted from each image, the percentile approach will quintuple the number of features used
in the classification models to 85 features per season.

2.4. Training and test data sets


In this study, all the training and validation was done at the pixel level. To avoid the
effect of different agricultural field sizes, a subset with the same number of pixels is extracted
from each field polygon. The average size of agricultural fields in the west is greater than that in
the east, so they had more pixels within a field polygon in the remote sensing images. The
average size of field polygons in the eastern Uttar Pradesh and Bihar in this study is around 17
pixels, so a subset of 10 pixels was used to represent each field polygon. If an agricultural
polygon has more than 10 pixels, a subset of 10 pixels will be randomly selected from that
polygon. If the field size is smaller than or equal to 10 pixels, then all the pixels within the
polygon will be extracted.
The ground truth data was split into a training data set (70%) and a test data set (30%).
The training data set was used to create the classification models, and the rest 30% of the field
polygons set aside for validation. The number of zero tillage and conventional tillage polygons
in the training data set was balanced, as was that in the test data set. But in the training and
test data sets, the number of ZT pixels and CT pixels were slightly different due to the presence
of polygons with less than 10 pixels.

2.5. Random forest classification


The machine learning algorithm used in this study is the random forest algorithm
(Breiman, 2001) which is widely used in remote sensing imagery classification research. Other
advanced machine learning algorithms are used in the field of remote sensing, like support
vector machine (SVM) and artificial neural network (ANN), but the random forest is still the
most appropriate method for this study due to its high computational efficiency for large data
sets. In addition, Pal (2005) found that the random forest algorithm could produce the same
classification accuracy as SVM with fewer parameters.
Even though random forest models won’t be affected by the multi-collinearity problem
of input variables, a feature selection was performed before creating the models to avoid
redundant information. The ‘findCorrelation’ function from the ‘caret’ package (Kuhn, 2008)
was used to remove all the highly correlated features which have a correlation higher than 0.9,
the default threshold. The correlation was calculated for the sowing season, peak season, and
two-season remote sensing composites using all the pixels from all three years’ training
polygons. After applying the feature selection, 40-50% of the features were removed before
creating the random forest models. The reduction of input features greatly improved the
efficiency of the computation.
The feature selection and random forest classification were processed in R version 3.6.1
(R Development Core Team, 2019) and the random forest analysis was conducted using the
‘randomForest’ package (Liaw and Wiener, 2002). There are two main parameters used in the
random forest function, ‘mtry’ and ‘ntree’. The ‘mtry’ is the number of variables randomly
sampled at each split, and the ‘ntree’ is the number of decision trees generated. In this study,
we used the default value for ‘mtry’, which is the square root of the number of variables, and
500 for ‘ntree’, which is also the default value. The random forest models were generated in
four different ways: 1) using all the composites from 2018 and training four state-specific
models, 2) creating three multi-state generalized models with composites from all four states
and comparing the accuracies using sowing season data only, peak season data only, and both
sowing and peak seasons data in 2018, 3) choosing the most important time period from the
2018 multi-state generalized models and creating the random forest models with the bootstrap
sample size of 25%, 50%, 75%, and 100% of the training polygons, and 4) creating a final multi-
year generalized model using data from all states and all three years.

2.6. Bootstrapping
We used the bootstrapping to detect whether the classification accuracy will change
statistically with fewer training field data in the 2018 multi-state generalized model.
Bootstrapping is a common method in statistical estimates that allows assigning measures of
accuracy, like confidence intervals (Efron and Tibshirani, 1993). It based on the approach of
resampling from the data to create the bootstrap samples. In our study, the total number of
field polygons in 2018 is 2830 (1415 ZT and 1415 CT), and when creating the multi-state
generalized model, 70% of the field polygons (991 ZT and 991 CT) were assigned as the training
data set and 30% of the field polygons (424 ZT and 424 CT) were assigned as the test data set.
Thus, for the training bootstrap samples, we tried four different sizes of subset, 25%, 50%, 75%,
and 100% of the original number of training polygons in 2018. For the test bootstrap samples,
we kept the number of original test polygons (848 polygons) for all the models. For each
different bootstrap sample size, 400 different training and test subsets were randomly chosen
and 400 times of iteration were done with these 400 subsets (Davidson and MacKinnon, 2000).
Different from the previous multi-state generalize model, the polygon selection for training and
test subsets was completely random which means there is no stratified sampling by states or by
tillage type. The features used in all the random forest models were the same as the 2018
multi-state generalized model with the remote sensing data from the most important time
period.

3. Results
3.1. State-specific and multi-state models
3.1.1. State-specific models
We trained a state-specific model for each state in 2018 with both the sowing and peak
season data in 2018. Table 3 shows the confusion matrices of the four state-specific models on
the test data sets. The four random forest models were created independently, but the feature
selection process was the same, so their input features from the satellite imagery composites
were the same. These are the validation results when applying the classification model to the
test data set for each state (Table 3). The overall test accuracies ranged from 71.07% to 81.39%
where the accuracies increased from the eastern to the western states. The user and producer
accuracies all have the same spatial trends as the overall accuracies and increased from eastern
to western IGP.

Table 3. Confusion matrices of the state-specific models on validation data


State Punjab Haryana Uttar Pradesh Bihar
PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE
STATE SPECIFIC
CT 910 166 84.57% CT 929 286 76.46% CT 447 176 71.75% CT 609 270 69.28%
TEST
ZT 236 848 78.23% ZT 286 934 76.56% ZT 153 473 75.56% ZT 266 708 72.69%
79.41% 83.63% 81.39% 76.46% 76.56% 76.51% 74.50% 72.88% 73.66% 69.60% 72.39% 71.07%

3.1.2. Multi-state generalized model


The multi-state generalized model, which was generated with all the training data from
all four states, was applied to the test data set for each state (Table 4). The overall test
accuracies were almost the same as the state-specific models. Except for the Punjab state-
specific model, which was 1.02% better than the multi-state generalized model, the percentage
differences for the other states were within 1%. This indicates that it is possible to create a
multi-state generalized model that could perform as well as state-specific models across the
entire IGP region. Consistent with state-specific models, the overall user and producer
accuracies are all lower in the eastern states. The mean overall classification accuracy for the
IGP is 75.59%.

Table 4. Confusion matrices of the multi-state generalized models on validation data


State Punjab Haryana Uttar Pradesh Bihar GENERALIZED MOD
PREDICT PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE TRUE
GENERALIZED
CT 910 166 84.57% CT 933 282 76.79% CT 479 144 76.89% CT 589 290 67.01% CT 2911 882 76.75%
TEST
ZT 258 826 76.20% ZT 297 923 75.66% ZT 190 436 69.65% ZT 250 724 74.33% ZT 997 2907 74.46%
77.91% 83.27% 80.37% 75.85% 76.60% 76.43% 71.60% 75.17% 73.26% 70.20% 71.40% 70.86% 74.49% 76.72% 75.59%

3.2. Critical time period


We compared the state-specific generalized models with remote sensing data from
different time periods. The previous models used the training data from both the sowing and
peak seasons in 2018. The confusion matrices below show the results of the multi-state
generalized model that only used the data from the sowing season (Table 5), and the model
that used the data from peak season only (Table 6). The model that used sowing season data
has almost the same overall accuracy as the model that used both the sowing and peak seasons
data. Not only does the average accuracy of the planting season data model and the two-
season data model in the whole IGP region remain the same (reduced by 0.08%), but the
accuracy of each state is also very similar to that of the two-season data model, with a
difference of less than 2 percent (±2%).

Table 5. Results of the multi-state generalized model that used the sowing season data only
2018MOD
Punjab Haryana Uttar Pradesh Bihar GENERALIZED MOD
Sow only
PREDICT PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE TRUE
GENERALIZED
CT 916 160 85.13% CT 920 295 75.72% CT 467 156 74.96% CT 558 321 63.48% CT 2861 932 75.43%
TEST
ZT 224 860 79.34% ZT 299 921 75.49% ZT 192 434 69.33% ZT 238 736 75.56% ZT 953 2951 75.59%
80.35% 84.31% 82.22% 75.47% 75.74% 75.61% 70.86% 73.56% 72.14% 70.10% 69.63% 69.83% 75.01% 76.00% 75.51%

The overall accuracies of the multi-state generalized model with only peak season data
dropped by 4% to 21% in the four states from the model with imagery composites from both
seasons (Table 6). In Punjab and Haryana, where the accuracies are high in the previous multi-
state generalized model, the decline in overall accuracies is much greater than that in Uttar
Pradesh and Bihar. The average overall accuracy dropped by 15.7% in the IGP region. The result
indicates that the imagery composites from the peak season may not be as useful as the data
from the sowing season when mapping the ZT adoption.

Table 6. Results of the multi-state generalized model that used the peak season data only
2018MOD
Punjab Haryana Uttar Pradesh Bihar GENERALIZED MOD
Peak only
PREDICT PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE TRUE
GENERALIZED
CT 656 420 60.97% CT 629 586 51.77% CT 384 239 61.64% CT 552 327 62.80% CT 2217 1576 58.45%
TEST
ZT 465 619 57.10% ZT 529 691 56.64% ZT 221 405 64.70% ZT 301 673 69.10% ZT 1511 2393 61.30%
58.52% 59.58% 59.03% 54.32% 54.11% 54.21% 63.47% 62.89% 63.17% 64.71% 67.30% 66.11% 59.47% 60.29% 59.89%

Result: Make a map of ZT/CT for 2018 across the IGP (multi-state generalized model sow only)

3.3. Field data reduction


When we reduce the number of polygons in the training data set, the test overall
accuracy decreased. The overall accuracies of different bootstrap sample sizes are shown in
Figure 3, and the confidence intervals are marked with the error barsError: Reference source
not found. The accuracies of the models that trained with four sizes of training data set are all
significantly different from each other on a significance level of 0.05. The speed of decrease in
the overall accuracy improved with the reduction of training polygons. The overall accuracy of
the model using 75% of the training polygons is 0.71% less than that of using all training
polygons. But when the number of training polygon reduced from 75% to 50%, the overall
accuracy decreased by 0.93%. In a similar way, the overall accuracy decreased even more, when
the number of training polygons is 25% of all the training polygons instead of 50%. It is
remarkable that the overall accuracy decreased almost the same amount when the number of
training polygons decreased from 100% to 50% and when it decreased from 50% to 25%.
77.38%
77.50%
77.00% 76.67%
76.50%

Overall accuracy
76.00% 75.74%
75.50%
75.00%
74.50%
74.02%
74.00%
73.50%
495 (25%) 991 (50%) 1487 (75%) 1982 (100%)
Number of training polygons

Figure 3. Overall accuracies with different number of training polygons in 2018

3.4. Apply the generalized models to previous years


We trained a multi-year generalized model with 70% of all the field polygons from 2016,
2017, and 2018. Then, the multi-state generalized model created with 2018 data and the model
created with multi-year data were applied to the test data sets from 2016 and 2017. Both
models are multi-state generalized models that were created with the sowing season image
composites from the four states. The 2018 multi-state model has a steady performance in
Punjab and Haryana in 2016 with the overall accuracies dropping from ~80% to ~70% (Table 7).
However, the 2018 multi-state model has a very poor performance in Uttar Pradesh and Bihar
in 2017. There are only two tillage types, ZT and CT, in this study. Since the number of training
pixels from each class is almost balanced, the completely random classification accuracy should
be 50 percent. However, the overall accuracy in Bihar is only 2.35 percent higher than 50% and
the performance of the 2018 model is even worse than randomness in Uttar Pradesh (Table 7).
We could not make the statement that the multi-state generalized model in 2018 could be
applied for 2016 and 2017.

Table 7. Confusion matrices for the 2018 model applied to the 2016 and 2017 test data
2018MOD 2018MOD
Punjab Haryana Uttar Pradesh Bihar
2016Data 2017Data
PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE
2018 2018
CT 171 29 85.50% CT 245 115 68.06% CT 44 57 43.56% CT 49 107 31.41%
Data Data
ZT 82 118 59.00% ZT 101 259 71.94% ZT 51 45 46.88% ZT 45 118 72.39%
67.59% 80.27% 72.25% 70.81% 69.25% 70.00% 46.32% 44.12% 45.18% 52.13% 52.44% 52.35%

When using all the training data from 2016, 2017, and 2018 to create the model and
applying it to the previous years, the model performs much better compared to the 2018 multi-
state generalized model. In western IGP, the model has almost the same accuracy in 2016
(Table 8. Confusion matrices for the multi-year generalized model applied to the 2016 and 2017
test dataTable 8) as that in 2018 (Table 5). And more importantly, the multi-year generalized
model has a much higher overall accuracy in eastern IGP than the 2018 multi-state model. The
overall accuracy improved by 8% in Bihar and improved by 21% in Uttar Pradesh. Even though
the accuracies are lower in eastern IGP than in western IGP and the accuracies are lower in
2017 than in 2018, this model has the best performance of the models that trained with the
current field data and method.

Table 8. Confusion matrices for the multi-year generalized model applied to the 2016 and 2017
test data
2018MOD 2018MOD
Punjab Haryana Uttar Pradesh Bihar
2016Data 2017Data
PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE
3-year 3-year
CT 155 45 77.50% CT 296 64 82.22% CT 62 39 61.39% CT 86 70 55.13%
Data Data
ZT 27 173 86.50% ZT 102 258 71.67% ZT 27 69 71.88% ZT 56 107 65.64%
85.16% 79.36% 82.00% 74.37% 80.12% 76.94% 69.66% 63.89% 66.50% 60.56% 60.45% 60.50%

4. Discussion
Our study assessed the random forest models for mapping ZT areas in the smallholder
farming systems with multiple field data combinations and remote sensing image composites.
We found that we can generate a multi-state generalized model that has the same accuracy as
the state-specific models in 2018. We also successfully created a model that could be applied to
identify the ZT areas for multiple years in the IGP of India. These results demonstrated that the
high spatiotemporal resolution satellite imagery, Sentinel-2, could be used to map ZT adoptions
in the smallholder farming systems in a large spatial and temporal scale.
In our study, the state-specific models and the multi-state model have similar
classification accuracies for all the states in the IGP in 2018 which indicates that it is possible to
create a high-accuracy generalized model across states for the entire IGP. The different
approaches in farming are the greatest challenge of making a generalized model for such a
large area, especially the variance in the sowing date. This variance causes a trade-off between
the size of the composite time window and the amount of remote sensing data availability. To
cover the entire range of sowing dates from all the four states in a consistent time window, the
size of the time window tends to be large. But a large time window will also reduce the amount
of remote sensing data when the images within the time window are all combined into one
composite. Due to this dilemma, the use of composite image percentiles is an efficient way to
simultaneously increase the amount of input data and capture the crop phenological
characteristics. In addition, the feasibility of a model across states could also effectively reduce
the computing effort. We don’t need to create multiple separate models for particular states in
the IGP when mapping the ZT areas. A single model can be created and applied to all the four
states in the IGP as long as part of the data from each state is included in the training data set.
When detecting the ZT adoption, the time period around the sowing date is more critical
than other time periods in the growing season. The composite imagery percentile method,
which captures the sowing dates in the sowing season composite by the 0 th percentile of
vegetation indices and 100th percentile of reflectance bands, ensures that the information on
tillage which happened right before the sowing date is captured. To be specific, the minimum
NDTI and minimum CRC from the sowing season composite are more beneficial to the multi-
year random forest model than other vegetation indices when distinguishing ZT from CT. This
result is consistent with previous studies in residue mapping (Jin et al., 2015; Zheng et al.,
2013). In contrast, the lack of additional information from the peak season composite indicates
that the vegetation indices of wheat from ripening to harvesting in ZT fields and CT fields are
not important. However, most of the previous research found that ZT adoption is able to
improve crop yield (Erenstein et al., 2008; Kumar et al., 2019; Nandan et al., 2018). The reason
for the decrease in the usefulness of the peak season composite might be that most of the
vegetation indices used in the classification model are for residue mapping rather than biomass
measurement.
One of the advantages of using remote sensing data only from the sowing season is that
it can reduce the amount of input data by half. Shortening the time span of remote sensing
imagery requirements will save a substantial amount of data pre-processing time which is the
most time-consuming but essential part of this study. The image pre-processing is the first
requirement for all the images within the time span, even though the remote sensing images
would be combined into two composites with the percentile method. There are 4744 Sentinel-2
images from October 15th, 2018 to April 15th, 2019 on GEE, so not including the peak season in
the study could reduce the pre-processing for around 2000 images and that’s just for one of the
three years in this study.
Data from previous years is vital to the classification model if the objective is to create a
model that could be applied for multiple years. The huge drop in accuracy when applying the
multi-state generalized model trained with the 2018 field data to the 2016 and 2017 test data
set confirms the inter-annual differences in the sowing season. For example, the sowing date
might vary from year to year. Thus, including data from multiple years could take the inter-
annual variation into consideration and improve model compatibility. However, this finding also
means researchers have to collect ZT and CT data in all the years they want to apply the model,
which cannot ease the field team’s workload. But the helpful result from our study for the field
team is that the number of field polygons that required in one year is reducible. The multi-state
generalized model that used 50% of the training data in 2018 has a comparable accuracy as the
model that used all the training polygons. Even though the accuracies significantly decreased
when reducing the number of training polygons, the absolute changes in overall accuracies are
small, which only dropped less than 4% from the model that used 1982 polygons for training to
the model that used 495 polygons for training.
The model has a better performance in western IGP for several reasons. First, the field
data quality is higher in western IGP than that in eastern IGP. The number of field polygons in
Punjab and Haryana is greater than the number of polygons in Uttar Pradesh and Bihar, thus
capturing the diversity in the fields of western IGP. The field size in western IGP is also larger.
The fields in eastern IGP are more fragmented with diverse crop types in addition to wheat.
Second, there is more crop residue in the ZT fields in western IGP. Many studies have used
remote sensing images to map the residue quantity in the field, but in this study, the amount of
residues on the ground was not taken into account. The more residue there is in the ZT fields,
the easier it is to separate them from the CT fields. The field team confirmed that there is more
surface residue in the ZT fields in Punjab and Haryana. In addition, the ‘Happy Seeder’, which is
a machine that helps to plant the wheat seed among the rice straw, are widely used in
northwest IGP. The usage of ‘Happy Seeder’ from a side indicates that the volume of rice
residues are large in Punjab and Haryana.
Further work could consider whether other vegetation indices and satellite products
could provide more information to map ZT adoption. First, including more vegetation indices
that could measure crop production during the peak season. Since the effect of ZT practice is
reflected in crop production, calculating the indices like crop dry mass (CDM), leaf area index
(LAI), and green area index (GAI) may provide more features that could help identify the ZT
adopted areas. Second, using radar imagery like Sentinel-1, which could provide the difference
of roughness and texture in the field, in the model may help improve the classification accuracy.

5. Conclusions
This study shows that the high spatiotemporal resolution satellite imagery, Sentinel-2,
could be used to map ZT adoptions accurately in the smallholder farming systems in the IGP of
India. The multi-state generalized model has an average overall accuracy of 75.51% for all the
test data sets. Amongst all the states, Punjab and Haryana have higher accuracy (~80%) than
eastern Uttar Pradesh and Bihar (~70%). The satellite data used in the model were composite
image percentiles within a 3-month time window around the wheat sowing date which is the
most essential time period for mapping ZT. Reducing the number of training polygons does not
have much effect on absolute overall accuracy. We also trained a multi-year generalized model
that can be used across years and states with three years’ ZT and CT field data in the IGP. The
results of this study demonstrate the potential of using remote sensing imagery with a high
spatiotemporal resolution to map the ZT area across a large region.

Acknowledgments
Reference
Alexandratos, N., 1995. World agriculture: towards 2010: an FAO study .FAO. Rome.
Azzari, G. et al., 2019. Satellite mapping of tillage practices in the North Central US region from
2005 to 2016. Remote Sens. Environ. 221, 417–429.
Breiman, L., 2001. Random Forests. Mach. Learn. 45(1), 5–32.
Carlson, T. N., Ripley, D. A., 1997. On the relation between NDVI, fractional vegetation cover,
and leaf area index. Remote Sens. Environ. 62(3), 241–252.
DAC&FW, 2015. All India Report On Agriculture Census 2010-11.
Davidson, R., MacKinnon, J. G., 2000. Bootstrap tests: How many bootstraps? Econom. Rev.
19(1), 55–68.
de Freitas, P. L., Landers, J. N., 2014. The Transformation of Agriculture in Brazil Through
Development and Adoption of Zero Tillage Conservation Agriculture. Int. Soil Water
Conserv. Res. 2(1), 35–46.
Efron, B., Tibshirani, R. J., 1993. An Introduction to the Bootstrap. In An Introduction to the
Bootstrap.
Erenstein, O. et al., 2008. On-farm impacts of zero tillage wheat in South Asia’s rice-wheat
systems. F. Crop. Res. 105(3), 240–252.
Erenstein, O., 2009. Zero Tillage in the Rice-Wheat Systems of the Indo-Gangetic Plains: A
review of impacts and sustainability implications .International Food Policy Research
Institute.
Erenstein, O., Laxmi, V., 2008. Zero tillage impacts in India’s rice-wheat systems: A review. In
Soil and Tillage Research.
ESA, 2015. Sentinel-2 User Handbook (pp. 1–62).
Gebhardt, M. R. et al., 1985. Conservation tillage. Science (80-. ). 230(4726), 625–630.
Gitelson, A. A. et al., 2003. Remote estimation of leaf area index and green leaf biomass in
maize canopies. Geophys. Res. Lett. 30(5), 1248.
Gorelick, N. et al., 2017. Google Earth Engine: Planetary-scale geospatial analysis for everyone.
Remote Sens. Environ.
Hagen, S. et al., 2016. Operational Tillage Information System ( OpTIS ): A Pilot Demonstration
Project Mapping Tillage Practice and Winter Cover Crops Annually Across the State of
Indiana Between 2006 and 2015 .Applied GeoSolutions and the Conservation Technology
Information Center.
Horowitz, J. et al., 2010. “No-till” farming is a growing practice. Econ. Inf. Bull. 70, 1–28.
Jin, X. et al., 2015. Estimation of maize residue cover using Landsat-8 OLI image spectral
information and textural features. Remote Sens. 7(11), 14559–14575.
Kuhn, M., 2008. caret Package. J. Stat. Softw.
Kumar, V. et al., 2019. Impact of tillage and crop establishment methods on crop yields,
profitability and soil physical properties in rice–wheat system of Indo-Gangetic Plains of
India. Soil Use Manag. 35(2), 303–313.
Liaw, A., Wiener, M., 2002. Classification and Regression by randomForest. R News.
Mcnairn, H., Protz, R., 1993. Mapping corn residue cover on agricultural fields in oxford county,
ontario, using thematic mapper. Can. J. Remote Sens. 19(2), 152–159.
Murphy, S., 2018. Atmospherically Corrected Time Series using Google Earth Engine. GitHub
Repos.
Nandan, R. et al., 2018. Crop establishment with conservation tillage and crop residue retention
in rice-based cropping systems of Eastern India: yield advantage and economic benefit.
Paddy Water Environ. 16, 477–492.
Pal, M., 2005. Random forest classifier for remote sensing classification. Int. J. Remote Sens.
26(1), 217–222.
Pandit, M. C. et al., 2010. Economic analysis of adoption of zero tillage technology in wheat
cultivation in trans-gangetic plains of India. Indian J. Agric. Sci. 80(12), 1092–1094.
Pathak, H. et al., 2003. Trends of climatic potential and on-farm yields of rice and wheat in the
Indo-Gangetic Plains. F. Crop. Res. 80(3), 223–234.
Prasad, S. K. et al., 2018. Zero tillage impacts on economics of wheat production in far western
Nepal. Farming Manag. 3(2), 93–99.
R Development Core Team, 2019. R: A language and environment for statistical computing .R
Foundation for Statistical Computing. Vienna, Austria.
Sullivan, D. G. et al., 2006. Evaluating techniques for determining tillage regime in the
Southeastern Coastal Plain and piedmont. Agron. J. 98(5), 1236.
Timsina, J., Connor, D. J., 2001. Productivity and management of rice-wheat cropping systems:
Issues and challenges. In Field Crops Research.
Tripathi, R. S. et al., 2013. Impact of zero tillage on economics of wheat production in Haryana.
Agric. Econ. Res. Rev. 26(1), 101–108.
Van Deventer, A. P. et al., 1997. Using thematic mapper data to identify contrasting soil plains
and tillage practices. Photogramm. Eng. Remote Sensing 63(1), 87–93.
Wilson, R. T., 2013. Py6S: A Python interface to the 6S radiative transfer model. Comput.
Geosci.
Zheng, B. et al., 2013. Multitemporal remote sensing of crop residue cover and tillage practices:
A validation of the minNDTI strategy in the United States. J. Soil Water Conserv. 68(2),
120–131.
Supplement
#1 Applied the three-year multi-state generalized model to the 2018 data
Comparing the results of the 3-year model with that of the 2018 model, the overall
accuracies are very similar as well. The differences are all within one percent (±1%), which
means that the additional input training data from other years didn’t add much information to
the model when mapping the ZT areas in 2018.

Confusion matrices of the multi-year generalized model that used the sowing season data from
2016, 2017, and 2018
2018MOD
Punjab Haryana Uttar Pradesh Bihar GENERALIZED MOD
3year sow data
PREDICT PREDICT PREDICT PREDICT PREDICT
CT ZT CT ZT CT ZT CT ZT CT ZT
TRUE TRUE TRUE TRUE TRUE
GENERALIZED
CT 922 154 85.69% CT 931 284 76.63% CT 461 162 74.00% CT 551 328 62.68% CT 2867 926 75.59%
TEST
ZT 220 864 79.70% ZT 297 923 75.66% ZT 193 433 69.17% ZT 233 741 76.08% ZT 945 2959 75.79%
80.74% 84.87% 82.69% 75.81% 76.47% 76.14% 70.49% 72.77% 71.58% 70.28% 69.32% 69.72% 75.21% 76.16% 75.69%

#2 Variable importance of the 2018 generalized model


The variable importance in the random forest model can be assessed by the mean
accuracy decreasing when not including a certain variable in the model. The variable
importance plot of the multi-year generalized model showed that the minimum NDTI is the
most important feature in mapping the ZT areas in the IGP. After that, the maximum red edge
4, minimum blue, minimum CRC, and minimum SWIR 1 are also very important when classifying
ZT and CT.

Variable importance in the random forest model

You might also like