1 of 28

Decoding Drought: Embracing Simplicity in�Effective Predictive Models

Akash Poptani

Sayali Lokhande

Rahul Jashvantbhai Pandya�Sridhar Iyer

2 of 28

Understanding Drought Through Machine Learning and Remote Sensing

Connection between rising global temperatures and increased drought occurrences [1]
Importance of early detection for effective mitigation and response strategies
Wide-Ranging Effects of Droughts
Role of Machine Learning (ML) and Artificial Intelligence (AI) in transforming agriculture
Insights provided to farmers: crop health, soil conditions, water usage, and crop patterns
Applications: crop type identification, precision farming, crop monitoring, and yield prediction

[1] C. Cammalleri, G. Naumann, L. Mentaschi, G. Formetta, G. Forzieri, S. Gosling, B. Bisselink, A. De Roo, and L. Feyen, “Global warming and drought impacts in the eu,” Publications Office of the European Union: Luxemborg, no. 29956 EN, 2020. JRC118585.

According to the European Technical Report\cite{cammalleri2020global}, rising global temperatures, linked to heightened levels of greenhouse gases in the atmosphere, play a pivotal role in altering the climate patterns, influencing both, the occurrence and the severity of droughts.

Early detection of drought is crucial for the implementation of effective mitigation and response strategies. It empowers authorities to start water conservation measures, allocate resources for relief efforts, and provide timely assistance to affected communities. This proactive approach is instrumental in minimizing the social and economic impacts of drought, promoting resilience, and ensuring sustainable management of water resources. �drought results in water scarcity, impacting agriculture, economy, and ecosystems on a large scale

Machine Learning (ML) and Artificial Intelligence (AI) play a crucial role in transforming agriculture by providing inventive solutions to diverse challenges. ML can analyze extensive datasets such as, satellite imagery and sensor data, delivering valuable insights to farmers regarding crop health, soil conditions, water usage, and patterns associated with healthy and diseased crops.

3 of 28

Preprocessing - Dataset 1

Advanced Very High Resolution Radiometer (AVHRR) satellites [3]

Available in Star Nesdis website

1981 to 2022, recorded weekly

Star Nesdis

NDVI

Time-shifted labels for VHI data used for predictions

Dataset division for 1 week, 2 weeks, 3 weeks, and 4 weeks forecasting horizons

[3] S. Kalluri, C. Cao, A. Heidinger, A. Ignatov, J. Key, and T. Smith, “The advanced very high resolution radiometer: Contributing to earth observations for over 40 years,” Bulletin of the American Meteorological Society, vol. 102, no. 2, pp. E351 – E366, 2021

4 of 28

Preprocessing - Dataset 1

LandSat 8

LST

Scaling and conversion to K

LST = Land Surface Temperature

5 of 28

Preprocessing - Dataset 1

Star Nesdis

LandSat 8

NDVI

LST

TCI

VCI

VHI

VHI image data at time ‘t’

i is the pixel value

β = 1- α =0.5 (current implementation)

Shape of the images: (192, 256, 1)

Vegetation Condition Index (VCI)

Temperature Condition Index (TCI)

6 of 28

i is the pixel value

β = 1- α

α = β = 0.5

F. Kogan, “Early detection and monitoring droughts from noaa environmental satellites,” NATO Science for Peace and Security Series C: Environmental Security, vol. 97, 10 2011.

Preprocessing - Dataset 1

7 of 28

Preprocessing - Dataset 2

Rainfall�NetCDF Format

Rainfall image data at time ‘t’

IMD Pune

Transformation into 2D raster image dataset
Operations such as resampling, reprojecting, and interpolation for alignment with VHI dataset
Reshaping and padding for further analysis

Shape of the images: (192, 256, 1)

D. Pai, L. Sridhar, M. Rajeevan, O. P. Sreejith, N. Satbhai, and B. Mukhopadhyay, “Development of a new high spatial resolution (0.25° × 0.25°) long period (1901-2010) daily gridded rainfall data set over india and its comparison with existing data sets over the region,” Mausam, vol. 65, pp. 1–18, 01 2014.

The open-source portal of the Climate Research and Services group at the Indian Meteorological Department in Pune, India,

provides precipitation data in gridded NetCDF format with a

spatial resolution of 0.25° × 0.25°, obtained from nearly 7000 rain gauge stations for the period

1981-2021�This data

undergoes transformation into a 2D raster image dataset

through GIS-based operations. The process involves

converting pixel values for each latitude-longitude pair

into a single pixel value to achieve a 2D array for-

mat. Subsequently, operations such as resampling, re-

projecting, and interpolation are applied to align the

spatial resolution with the VHI dataset. Additionally,

the daily rainfall data is converted into a weekly format.

The resulting raster image dataset encompasses key at-

tributes, including latitude, longitude, rainfall, and time,

providing a comprehensive representation of the gridded

precipitation data. Further processing involves reshaping

and padding, resulting in a shape of (192, 256, 1). Two

time-shifted labels, three time-shifted labels, four time-

shifted labels are used for 2 weeks, 3 weeks, 4 weeks

prediction respectively. A min-max scaler is applied to

the samples, followed by normalization.

8 of 28

VHI+Rainfall Dataset

Star Nesdis & LandSat8

VHI Dataset

VHI Values

Padding, Normalization and Scaling Operations

Conversion to weekly format

Transformation into 2D raster image dataset

Rainfall data from IMD Pune

NDVI, LST, VCI and TCI Calculation

Reshaping, Padding and Normalization Operations

Alignment with VHI dataset using resampling, reprojecting and interpolation

Rainfall Dataset

9 of 28

Integrated system model: Utilizing VHI and rainfall data

Shape of the merged dataset: (192, 256, 2)

LR

VHI

Prediction outcome

Rainfall�NetCDF Format

VHI image data at time ‘t’

Rainfall image data at time ‘t’

……...

Flattened VHI data

Flattened rainfall data

Output

Other Models

Prediction outcome

Output

merge

……...

Flattened VHI+Rainfall data

10 of 28

Merging The Datasets

Shapes of VHI and Rainfall datasets for Rajasthan region - (192, 256, 1)
Merged to create a dataset with images having shape of (192, 256, 2) except for MLR
For MLR, we input separate VHI and Rainfall values
One time-shifted labels - one week prediction, two time-shifted labels - two week prediction, three time-shifted labels - three week prediction and so on

11 of 28

Models

Multivariate Linear Regression (MLR)
Random Forest (RF)
Multi-Layer Perceptron (MLP)
K-Nearest Neighbors (KNN)
Multi-Output Support Vector Regression (SVR)
1D Convolutional Neural Network (1DCNN)

12 of 28

Multivariate Linear Regression (MLR)

Mathematical expression for MLR:

For dataset 1 (VHI),

x₁ → VHI Image Data

For dataset 2 (VHI+Rainfall),

x₁ → VHI Image Data, x₂ → Rainfall Image Data

13 of 28

K-Nearest Neighbours (KNN)

Calculating the Euclidean distance between two data points

Identifying the K instances with the smallest Euclidean distances
Target variable is predicted by averaging the values of its

k-nearest neighbors

14 of 28

Multilayer Perceptron (MLP)

For dataset 1,

VHI_pred = σ_out(W ^(p)_out·ReLU ^(p)(. . . ReLU ⁽²⁾(W ⁽²⁾·ReLU ⁽¹⁾(W ⁽¹⁾·VHI + b ⁽¹⁾) + b ⁽²⁾) . . .) + b_out

For dataset 2,

VHI_pred = σ_out(W ^(p)_out·ReLU ^(p)(. . . ReLU ⁽²⁾(W ⁽²⁾·ReLU ⁽¹⁾(W₁⁽¹⁾·VHI + W₂⁽¹⁾·Rainfall + b ⁽¹⁾) + b ⁽²⁾) . . .) + b_ou_t

W_out and b_out→ Weight matrix and bias for the output layer
W and b → Weight matrix and bias for the hidden layer
p → Number of hidden layers
W₁⁽¹⁾ and W₂⁽¹⁾ → Weight matrices for the input layer corresponding to VHI and Rainfall

15 of 28

Multi-output Support Vector Regression (SVR)

Mathematical formulation of SVR:

Beta → Coefficient matrix containing the weights assigned to support vectors
H → Kernel matrix representing the similarity between each pair of input data points
u → Prediction Error which is calculated using Euclidean norm of the error matrix
L → Slack Variable Penalty Term, C → Regularisation Parameter, Lp → Cost

function

16 of 28

1DCNN

Sequential application of layers in the 1DCNN architecture:

Here, Y represents the predicted output, and X is the input data
For dataset 1, VHI values are used as input
For dataset 2, merged VHI and Rainfall values are given as input

17 of 28

Random Forest (RF)

Mathematical formulation of RF is shown as:

Here, Y represents the predicted output, and X is the input data
For dataset 1, VHI values are used as input
For dataset 2, merged VHI and Rainfall values are given as input

18 of 28

Output for Dataset 1

Models	MLR			RF			MLP
Duration	R2 score	MAE	MSE	R2 score	MAE	MSE	R2 score	MAE	MSE
1 week	0.965	0.026	0.002	0.746	0.073	0.017	0.622	0.091	0.021
2 weeks	0.961	0.027	0.003	0.779	0.067	0.014	0.677	0.083	0.018
3 weeks	0.921	0.040	0.005	0.722	0.075	0.019	0.722	0.077	0.015
4 weeks	0.929	0.039	0.005	0.765	0.069	0.016	0.753	0.071	0.013

19 of 28

Output for Dataset 1

Models	KNN			SVR			1D-CNN
Duration	R2 score	MAE	MSE	R2 score	MAE	MSE	R2 score	MAE	MSE
1 week	0.941	0.033	0.004	0.915	0.041	0.006	0.829	0.053	0.009
2 weeks	0.927	0.038	0.005	0.914	0.041	0.006	0.831	0.054	0.009
3 weeks	0.905	0.044	0.007	0.906	0.044	0.007	0.846	0.050	0.008
4 weeks	0.891	0.048	0.008	0.896	0.047	0.008	0.830	0.053	0.009

20 of 28

Output for Dataset 2

Models	MLR			RF			MLP
Duration	R2 score	MAE	MSE	R2 score	MAE	MSE	R2 score	MAE	MSE
1 week	0.946	0.032	0.004	0.708	0.078	0.019	0.584	0.097	0.024
2 weeks	0.922	0.039	0.005	0.693	0.081	0.020	0.587	0.097	0.024
3 weeks	0.891	0.047	0.007	0.721	0.077	0.018	0.584	0.097	0.024
4 weeks	0.876	0.050	0.008	0.698	0.080	0.020	0.583	0.097	0.024

21 of 28

Models	KNN			SVR			1D-CNN
Duration	R2 score	MAE	MSE	R2 score	MAE	MSE	R2 score	MAE	MSE
1 week	0.961	0.027	0.003	0.897	0.045	0.007	0.827	0.053	0.009
2 weeks	0.944	0.032	0.004	0.895	0.046	0.007	0.804	0.059	0.011
3 weeks	0.904	0.043	0.007	0.894	0.046	0.007	0.802	0.059	0.011
4 weeks	0.904	0.043	0.007	0.892	0.046	0.007	0.803	0.059	0.011

Output for Dataset 2

22 of 28

Comparison Of Models

	MLR		RF		MLP		KNN		SVR		1DCNN
Dataset	1	2	1	2	1	2	1	2	1	2	1	2
Mean R2	0.944	0.909	0.753	0.705	0.693	0.584	0.916	0.928	0.908	0.895	0.834	0.809
Mean MSE	0.004	0.006	0.016	0.019	0.017	0.024	0.006	0.005	0.006	0.007	0.009	0.011
Mean MAE	0.033	0.042	0.071	0.079	0.081	0.097	0.041	0.036	0.043	0.046	0.052	0.058
Computational Time (mins)	10	12	240	260	4	6	<1	<1	6	8	13	16

23 of 28

Predicted Drought Maps for Dataset 1

MLR

KNN

MLP

RF

SVR

1D-CNN

24 of 28

Predicted Drought Maps for Dataset 2

MLR

RF

MLP

KNN

SVR

1D-CNN

25 of 28

Analysis

RF, MLP models - Complex architecture with numerous parameters

RF → Multiple decision trees, resulting in potential computational intensity when training a large number of trees
MLP → Characterized by multiple layers and neurons, introduces an abundance of weights and biases

MLR, KNN, SVR, 1DCNN models - Simple architecture with enhancing computational efficiency

MLR → Straightforward relationship between input variables and the output
KNN → Reliance on nearby points aligns with the frequency patterns observed in droughts
SVR → Model architecture involves dot products and addition operations
1DCNN → Maintains simplicity

with a single convolutional layer

26 of 28

Conclusion

Presented a comprehensive analysis of drought prediction models, focusing on Rajasthan region using two distinct datasets: VHI and VHI + Rainfall
Simpler models, such as MLR, KNN, SVR, and 1DCNN, outperform complex models like RF and MLP across various timescales (1 to 4 weeks)
MLR attains the highest average R2 score of 0.944, and lowest average MSE and MAE of 0.004 and 0.033 respectively for the VHI dataset
KNN exhibits an average R2 score of 0.93, with least MSE and MAE of 0.005 and 0.036 respectively for VHI + Rainfall Dataset

27 of 28

References

[1] C. Cammalleri, G. Naumann, L. Mentaschi, G. Formetta, G. Forzieri, S. Gosling, B. Bisselink, A. De Roo, and L. Feyen, “Global warming and drought impacts in the eu,” Publications Office of the European Union: Luxemborg, no. 29956 EN, 2020. JRC118585.

[2] N. P. S. S. A. S. Dipanwita Dutta, Arnab Kundu, “Assessment of agricultural drought in rajasthan (india) using remote sensing derived vegetation condition index (vci) and standardized precipitation index (spi),” The Egyptian Journal of Remote Sensing and Space Science, vol. 18, no. 1, pp. 53–63

[3] S. Kalluri, C. Cao, A. Heidinger, A. Ignatov, J. Key, and T. Smith, “The advanced very high resolution radiometer: Contributing to earth observations for over 40 years,” Bulletin of the American Meteorological Society, vol. 102, no. 2, pp. E351 – E366, 2021

[4] D. Pai, L. Sridhar, M. Rajeevan, O. P. Sreejith, N. Satbhai, and B. Mukhopadhyay, “Development of a new high spatial resolution (0.25° × 0.25°) long period (1901-2010) daily gridded rainfall data set over india and its comparison with existing data sets over the region,” Mausam, vol. 65, pp. 1–18, 01 2014.

28 of 28

THANK YOU