Comparative Study of LSTM-Models to Forecast Millet Production

in India

Bhramara Bar Biswal

, Satyanarayan Sahu

, Satya Narayan Das

, Prahallad Kumar Sahu

Shibani Tripathy

, Agni Tanmaya Behera

, Soumya Ranjan Mishra

, Ashutosh Mallik

and Shobhan Banerjee

Department of CSA, GIET University, Gunupur, Odisha, India

Department of CSE, CUTM, Jatni, Odisha, India

Department of CSE, GIET University, Gunupur, Odisha, India

Department of CSE, Indian Institute of Information Technology - Ranchi, Jharkhand, India

Keywords: Millet Consumption, Time Series Analysis, LSTM, Vanilla LSTM, Bidirectional LSTM, Convolutional

LSTM

Abstract: The production of millets has been extensively emphasized these days. Due to their high intrinsic qualities

and fewer requirements, the government is also trying to promote the production and consumption of millet

in the form of various millet missions. Due to limitations in the size of the data available, it becomes extremely

challenging to fine-tune the model in the context of limited availability. In this paper, we have attempted to

predict the consumption of millet in the Indian market using various variants of Long Short-Term Memory

(LSTM) models and compared their performances to predict the requirements of Jowar, Bajra, Ragi, and other

minor millet to see whether the forecast can meet the overall aggregated requirement or not. Both aggregate

and granular level forecasts have been analyzed to come up with a solution especially where the market is

booming, and data availability is constrained.

1 INTRODUCTION

The cultivation of millet has garnered significant

endorsement from both the central and various state

governments in contemporary times. This advocacy

is attributed to the superior nutritional profile of

millets when juxtaposed with other cereals, including

wheat and rice. While wheat serves as a

commendable source of Vitamin B, it is also

associated with gluten, which may provoke allergic

reactions, gastrointestinal disturbances, and adverse

consequences for gut health. In comparison, rice

exhibits a deficiency in fiber and micronutrients

relative to millet. Millet contains essential minerals

like magnesium, phosphorus, potassium, calcium,

iron, and Vitamin B. Notable advantages of millet

encompass a high fiber concentration, a low glycemic

index, the absence of gluten, and a wealth of

antioxidants.

Governmental bodies are undertaking numerous

initiatives aimed at augmenting millet production,

fostering awareness, facilitating market development,

ensuring sustainability, and formulating relevant

policies for millet agriculture. This endeavor is

congruent with global initiatives focused on

enhancing food security, advancing nutritional

quality, and promoting sustainable agricultural

methodologies.

The multifaceted nature and robustness of millet

significantly contribute to this initiative, considering

their diverse cultivation characteristics. Millets

exhibit a reduced water requirement for cultivation,

thereby enabling growth in regions characterized by

minimal precipitation. Millet demonstrates

remarkable adaptability to soils that are marginal or

of suboptimal quality. Their growth cycle generally

spans from two to four months, facilitating multiple

cropping opportunities throughout the year and

rendering them appropriate for production in areas

experiencing seasonal constraints. In addition to their

water needs, the demand for fertilizers and pesticides

in millet cultivation is notably low due to their

inherent resilience to pests and diseases when

compared to other cereal crops, thereby presenting a

cost-efficient option for farmers facing financial

Biswal, B. B., Sahu, S., Das, S. N., Sahu, P. K., Tripathy, S., Behera, A. T., Mishra, S. T., Mallik, A. and Banerjee, S.

Comparative Study of LSTM-Models to Forecast Millet Production in India.

DOI: 10.5220/0013583000004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 1, pages 631-637

ISBN: 978-989-758-763-4

631

challenges. Consequently, in light of prevailing

climatic conditions, the cultivation of millet is

imperative for the future of sustainable agriculture

and the satisfaction of human nutritional

requirements.

Authors in (Diene, et al. 2024) have addressed the

variability in pearl millet yield based on distance,

using UAV-based proxy sensing and ML. In

(Sankararao, Rajalakshmi, et al. 2022), the authors

have attempted to identify canopy water stress in

pearl millet using a UAV-based HSI sensor,

leveraging five ML-based feature selection

techniques. A blockchain smart contracts-based

method has been used in (Ning, Wang, et al. 2023) to

track millet information in the agricultural supply

chain. Authors in (Diack , et al. 2024) have estimated

the fraction of green cover for millet, using a

framework combining the green cover data Sentinel-

2 images. Machine Learning has been used to study

the reduction of obesity among children using the

nutritional contents of millet.

In (Suryo, Mustika, et al. 2019), the authors have

compared the RMSE values of LSTM with

the backpropagation algorithm, concluding the

effectiveness and improvement of LSTM over the

latter in the agricultural sector. A data discovery and

visualization tool has been presented in (Dhaliwal,

Galbraith, et al. , 2023) for time-series analysis in

agriculture. A time-series optimization and

forecasting task has been performed using

the Random Forest and ARIMA model in (Banerjee,

Banerjee, et al. , 2022). A comparative analysis

between deterministic and probabilistic time series

approaches has been performed by authors in

(Banerjee, Banerjee, et al. , 2023). LSTM RNNs have

been explained in detail in (Staudemeyer and Morris,

2019).

The challenge with millet consumption is that this

domain is new in the market as of now and sufficient

data is not available. The task is the generate

appropriate forecasts with the limited data available

in hand. In this paper, we have used various variants

of LSTMs namely – Vanilla LSTM, Stacked LSTM,

Bidirectional LSTM, Convolutional LSTM, and

LSTM to study their quality of forecasts with respect

to millet consumption. This is just an approach to

analyze whether the deep learning-based LSTM

models can generate quality forecasts based on

limited data or not.

2 DATA DESCRIPTION

2.1 Data Source

The data has been acquired by the data published by

the Indian Institute of Millets Research (IIMR),

which consists of data individually corresponding to

Finger Millets (Ragi), Pearl Millet (Bajra), Sorghum

(Jowar), and other minor millets. The data has been

made available since 1966-67 up to 2019-20. The

recent data has been acquired from the Agricultural

and Processed Food Products Export Development

Authority (APEDA) up to 2023-24.

2.2 Exploratory Data Analysis

Missing values were imputed with average

values.

Figure 1: Ragi Production

Figure 2: Bajra Production

INCOFT 2025 - International Conference on Futuristic Technology

632

Figure 3: Jowar Production

Figure 4: Minor Millet Production

The data from both sources were merged followed

by which their distributions were plotted. Figs. 1 to 4

show the production of Ragi, Bajra, Jowar, and Minor

Millet over the years. Figure 5 shows the aggregate

production of all the variants of millet over the years.

We see that the overall trend is decreasing as of now,

but ragi production is increasing over the years.

Figure 5: Aggregate Production of millet over the years

Figure 6: Distribution based on production volume

Figure 7: Production & Area wise distribution

The pie chart in Figure 6 represents the overall

production percentages of each variant of millet in

India. We see that the variant widely produced across

India is Jowar. Figure 7 shows the total production

and area distribution for each variant of millet in

consideration. Here we also see that the total area

needed for the cultivation of Jowar is also the highest

and the production of Ragi is the least so far all over

India, even when compared to minor millet

cultivation.

3 LSTM – RECURRENT NEURAL

NETWORKS

Long Short-Term Memory Networks (LSTMs) are a

specialized type of recurrent neural network (RNN)

designed to effectively learn and remember patterns

in sequential data over long periods. Introduced to

Comparative Study of LSTM-Models to Forecast Millet Production in India

633

address the vanishing gradient problem that

traditional RNNs face, LSTMs utilize a unique

architecture featuring memory cells, input gates,

output gates, and forget gates. This structure allows

them to selectively retain or discard information,

making them particularly well-suited for tasks such as

natural language processing, speech recognition, and

time series forecasting. By maintaining a memory of

previous inputs while processing new data, LSTMs

can capture complex dependencies and trends within

sequences, leading to improved performance in

various applications.

Here we’ve used various variants of LSTMs to see

whether or not are they able to generate forecasts with

respect to the production of millet of each type in

consideration.

3.1 Vanilla LSTM

It refers to the standard implementation of Long

Short-Term Memory networks, which serves as the

foundational architecture for many advanced LSTM

variants. It consists of memory cells that can store

information over long sequences, helping to mitigate

issues like the vanishing gradient problem commonly

encountered in traditional recurrent neural networks.

Figure 8 shows the simple Vanilla LSTM

architecture used for our analysis.

3.2 Stacked LSTM

These architectures represent an advancement of the

fundamental Long Short-Term Memory (LSTM)

framework, which entails the superimposition of

multiple LSTM layers to augment the model's

capacity and enhance its proficiency in assimilating

complex representations. In a stacked LSTM

configuration, the output generated by one LSTM

layer is utilized as the input for the subsequent layer,

thereby enabling the network to effectively capture

hierarchical features present within the dataset. This

multi-tiered methodology significantly improves the

model's ability to discern intricate temporal patterns

and dependencies across diverse time scales. By

leveraging multiple layers, these networks can

improve performance on a wide range of applications

while also allowing for greater expressiveness in

modeling sequences. Figure 9 shows the Stacked

Vanilla LSTM as used for our analysis.

3.3 Bidirectional LSTM

Bidirectional LSTMs represent a sophisticated

modification of the conventional Vanilla LSTM

architecture, significantly augmenting the model's

proficiency in assimilating contextual information

from both antecedent and subsequent sequences. In

contrast to traditional LSTMs, which typically

process data in a unilateral direction (generally from

antecedent to subsequent), bidirectional LSTMs are

comprised of two distinct LSTM layers: one layer

processes the input sequence in a forward manner,

while the other layer undertakes the processing in a

reverse manner. This dualistic methodology enables

the network to acquire a holistic comprehension of

the temporal dynamics inherent in the data, as it is

capable of integrating information from both

temporal directions.

3.4 CNN LSTM

The integration of Convolutional Neural Networks

(CNNs) and Long Short-Term Memory networks

(LSTMs) in a CNN-LSTM architecture leverages the

distinct advantages presented by each model to

proficiently process and analyze spatio-temporal

datasets. The resultant output produced by the CNN

is subsequently utilized as the input for the LSTM,

which adeptly captures the temporal dependencies

inherent in the sequence of feature maps generated by

the CNN. This synergy allows the model to leverage

CNN's ability to identify spatial hierarchies while the

LSTM handles the sequential relationships over time.

Figure 11 shows the CNN LSTM used for our

analysis.

3.5 Convolutional LSTM

Convolutional LSTMs (ConvLSTMs) are a

specialized variant of CNN LSTMs, where the

convolutional layers replace the fully connected

Figure 8: Vanilla LSTM Architecture

INCOFT 2025 - International Conference on Futuristic Technology

634

Figure 9: Stacked Vanilla LSTM Architecture

Figure 10: Bidirectional LSTM Architecture

Figure 11: CNN LSTM Architecture

layers found in standard LSTMs, enabling the model

to process input data as multi-dimensional arrays

rather than one-dimensional sequences. Figure 12

shows the Convolutional LSTM Architecture as used

for our analysis.

Figure 12: Convolutional LSTM Architecture

4 IMPLEMENTATION

The data acquired up to 2018-19 from IIMR has been

used to train the model. The data acquired from

APEDE has been used to test the model. This creates

a train set of 53 data points and a test set of 5 data

points corresponding the which the forecast will be

validated.

The Adaptive Moment Estimation (Adam)

optimizer has been used to train the model, since it

gives the benefits of both the RMSProp and

Momentum optimizers, hence adaptively adjusting

the learning rate. The mean-squared error has been

kept as the loss metric, which we are trying to

minimize during the training process. The Rectified

Linear Unit (ReLU) activation function has been used

across all the models. Since the data is not so big,

hence the batch size used in one pass has been set to

1, but the number of data points passed at a time will

be varied corresponding to which the RMSE values

will be calculated. The one that corresponds to the

lowest RMSE value will be finalized. The

hyperparameters along with their values have been

summarized in Table 1 below:

Table 1: Hyperparameters & their values

Paramete

Value

activation ‘relu’

timize

‘adam’

loss ‘mse’

features 1

n_seq 1

Comparative Study of LSTM-Models to Forecast Millet Production in India

635

Table 2: n_step parameter values

Variant LSTM n

ste

s RMSE

Ragi

Vanilla 11 125.75

Stacke

12 151.39

Bidirectional 10 141.00

CNN 12 154.89

Convolutional 13 169.21

Bajra

Vanilla 14 720.08

Stacke

15 743.24

Bidirectional 10 736.07

CNN 20 856.97

Convolutional 20 614.41

Jowar

Vanilla 7 239.02

Stacke

7 251.46

Bidirectional 13 276.84

CNN 19 241.43

Convolutional 14 324.56

Minor

Millet

Vanilla 17 40.23

Stacke

20 37.92

Bidirectional 9 32.40

CNN 3 49.65

Convolutional 4 42.95

For each variant, for each LSTM, the

hyperparameter – n_steps were varied from 3 to 20

and the value corresponding to which the lowest

RMSE value was acquired had been chosen for the

final training of the model. Apart from the

Convolutional LSTM which was trained for 500

epochs due to its complex nature, the LSTMs were

trained for 200 epochs. The n_step parameter values

corresponding to the lowest RMSE have been

mentioned in Table 2.

From Table 2, we can see that the Vanilla LSTM

even though the simplest one, gave us the least value

of RMSE for Ragi, Bajra, and Jowar. Bidirectional

LSTM gave us the least RMSE corresponding to

minor millet. Hence, we proceed with the final

forecasts using these values of n_steps, training the

Vanilla and Bidirectional LSTMs for 200 epochs.

5 RESULTS AND DISCUSSION

After running the forecasts corresponding to each

variant for 5 years, their aggregate sum was

calculated. Figure 13 shows the aggregate forecasts

along with the actual values from 2019-20 to 2023-

24.

Figure 13: Aggregate Forecast

From Figure 13, we can clearly see that the exact

and predicted values have a lot of differences. But if

we visually analyze the trend of the aggregate curve,

after 1990-91 there has been a constant decrease in

the production and our forecasts lie in line with the

decreasing slope. This shows that based on the past

trends of the univariate data distribution, the

generated forecasts were up to the mark.

6 CONCLUSION AND FUTURE

SCOPE

Even after the generation of good-quality forecasts,

we see that there has been a good difference between

the forecast and the actual values. The increase in

production as depicted from the actual values might

be a consequence of the campaigns being run by the

government to enhance the consumption and hence

the production of millets.

This is where multivariate analysis comes into

the picture where there’s a need to analyze various

other factors such as awareness, marketing, subsidies,

etc. based on the availability of the data. This problem

will be addressed in our future works in a sequel to

this paper.

ACKNOWLEDGEMENTS

This work is inspired by the valuable insights and

guidance given by Mr. Shiv Charan Banerjee, Joint

Director – IT, National Informatics Centre, MeitY –

Jharkhand.

INCOFT 2025 - International Conference on Futuristic Technology

636

REFERENCES

S. Mansour Diene et al., "Improving Pearl Millet Yield

Estimation From UAV Imagery in the Semiarid

Agroforestry System of Senegal Through Textural

Indices and Reflectance Normalization," in IEEE

Access, vol. 12, pp. 132626-132643, 2024, doi:

10.1109/ACCESS.2024.3460107.

A. U. G. Sankararao, P. Rajalakshmi, S. Kaliamoorthy and

S. Choudhary, "Water Stress Detection in Pearl Millet

Canopy with Selected Wavebands using UAV Based

Hyperspectral Imaging and Machine Learning," 2022

IEEE Sensors Applications Symposium (SAS),

Sundsvall, Sweden, 2022, pp. 1-6, doi:

10.1109/SAS54819.2022.9881337.

Y. Ning, T. Wang, T. Liu, Z. Gao, L. Sun and H. Zhang,

"The Traceability of Millet Based on Blockchain Smart

Contracts in Agricultural Supply Chain," 2023 2nd

International Conference on Artificial Intelligence and

Blockchain Technology (AIBT), Zibo, China, 2023, pp.

65-70, doi: 10.1109/AIBT57480.2023.00020.

I. Diack et al., "Combining UAV and Sentinel-2 Imagery

for Estimating Millet FCover in a Heterogeneous

Agricultural Landscape of Senegal," in IEEE Journal of

Selected Topics in Applied Earth Observations and

Remote Sensing, vol. 17, pp. 7305-7322, 2024, doi:

10.1109/JSTARS.2024.3373508.

M. Birundadevi, G. Premalatha, M. Nalini, C. Iyyanar and

V. Arul, "A Machine Learning Strategy for Reducing

Childhood Obesity Using Millet," 2023 9th

International Conference on Smart Structures and

Systems (ICSSS), CHENNAI, India, 2023, pp. 1-6, doi:

10.1109/ICSSS58085.2023.10407542.

P. S. Budi Cahyo Suryo, I. Wayan Mustika, O.

Wahyunggoro and H. S. Wasisto, "Improved Time

Series Prediction Using LSTM Neural Network for

Smart Agriculture Application," 2019 5th International

Conference on Science and Technology (ICST),

Yogyakarta, Indonesia, 2019, pp. 1-4, doi:

10.1109/ICST47872.2019.9166401.

J. K. Dhaliwal, M. E. Galbraith, C. K. Leung and D. Tan,

"A Data Discovery and Visualization Tool for Visual

Analytics of Time Series in Digital Agriculture," 2023

27th International Conference Information

Visualisation (IV), Tampere, Finland, 2023, pp. 268-

271, doi: 10.1109/IV60283.2023.00053.

S. C. Banerjee, S. Banerjee and P. Rai, "Vaccine Supply

Optimization and Forecasting using Random Forest and

ARIMA Models," 2022 IEEE 3rd Global Conference

for Advancement in Technology (GCAT), Bangalore,

India, 2022, pp. 1-5, doi:

10.1109/GCAT55367.2022.9972154.

S. C. Banerjee, S. Banerjee and R. K. Jain, "Vaccine Supply

Forecasting and Optimization using Deterministic and

Probabilistic Approaches," 2023 2nd International

Conference for Innovation in Technology (INOCON),

Bangalore, India, 2023, pp. 1-5, doi:

10.1109/INOCON57975.2023.10101255.

Staudemeyer, R. C., & Morris, E. R. (2019, September 12).

Understanding LSTM -- a tutorial into Long Short-

Term Memory Recurrent Neural Networks. arXiv.org.

https://arxiv.org/abs/1909.09586.

Comparative Study of LSTM-Models to Forecast Millet Production in India

637