Technical Workshop on implementation of YES-Tech

under PMFBY to states and Ics.

15-17 Feb, 2023, MANAGE, Hyderabad

Machine Learning models for crop yield estimation

under YES-TECH

KK Choudhary & Varun Pandey

Agricultural Sciences and Applications Group
RS&GIS Applications Area
National Remote Sensing Centre
ISRO, Hyderabad
E-mail: [email protected]
[email protected]
 Yield estimation methods

❑Semi-physical models

❑Simulation models

❑Statistical/Empirical: Linear/Non-linear
 JawasiyaSolanki
Uni-variate linear model rabi 2017-18

Grain Yield (Q Season Corr.(Yield

Insurance Units ha-1) max )
Rudaheda Sr No Mean (N=10) Mean
1 Amri 46.0 0.78 0.52
2 Ajawada 53.7 0.77 0.87
3 Baledi 42.2 0.80 0.41
4 Bichrod 62.1 0.78 0.78
5 Biharia 54.4 0.78 0.64
6 Chitwalia 39.2 0.76 0.10
7 Ghunaikhalsa 59.9 0.77 0.77
8 JawasiyaSolanki 46.1 0.78 0.87
9 Jhalara2 55.1 0.75 0.74
10 Kakrichand 45.5 0.79 0.52
11 Kaluheda 44.6 0.76 0.89
12 Mungawda 16.5 0.73 0.86
13 Nahariya 56.7 0.79 0.58
14 Ninora 62.3 0.77 0.81
15 Palki 42.9 0.69 0.86
Total Number of Insurance Units 16 Pitlawdia 53.6 0.79 0.63
17 Ramwasha 65.1 0.78 0.41
with CCE = 20 18 Rudaheda 42.0 0.74 0.67
CCE per IU = 10 19 Salakhedi 57.5 0.79 0.79
Total Number of CCE = 200 20 Tarnod 59.8 0.79 0.82
Pooled Data 0.57
Collaborating Institute: KVK, Ujjain
 Limitations of conventional Linear models

Circle NDVI LSWI VH RF RD Yld(Kg/

NDVI LSWI Ha)
C1 0.65 0.62 0.035 1154 75 780
VH
C2 0.71 0.58 0.042 895 79 854
C3 0.68 0.64 0.058 1059 78 1159
Rainfall …… …… …… …… …… …… ……
…… …… …… …… …… …… ……
AKOLA Cn 0.75 0.71 0.049 958 76 942
Yield = w1*ndvi + w2*lswi + w3*vh + w4*rf + w5*rd + k
Countable number of parameters
Assume some data distribution for each feature
Low accuracy in case:
Crop • Non-linear relations
Rainy Days
Yield • Inter correlation among input features
•Dimension of feature space in high.
 Why Machine Learning?
❖Ability of a system to learn
itself from the data. Architecture
Optimization
❖Independent of data
distribution
❖Complex non-linear Output
Input
relations
Training
❖Can handle high data Evaluation
dimensions
❖Independent of multi-
collinearity •Spectral bands
•Crop Map
❖Used globally for large •Vegetation Indices
•Crop Yield
scale crop yield estimation •Weather parameters
❖Used in pilot studies by
several agencies
 Machine Learning: Framework
Input layer (historical
years)
Features
collection Satellite
Bands, ML Models
Task
Indices and
crop map
(Deep Neural Crop Yield
Network, prediction at
Weather Random forest, aggregated
(RF, RD etc) CNN etc) and
disaggregated
level
Environment
(Soil)

Ground
Data
 Popular Machine learning models
ML Models

Genre ML models
Model Selection
Decision trees Classification and Regression tree (CART)
Random forest
Feature Selection

Neural Networks Fully connected, Sparsely connected,

Feature Sequential learning
Engineering

Hyper- Adaptive Models Boosting

parametrization
Kernel based Support Vector machine, KRR, GPR

Testing &
Evaluation
 Random Forest (RF)
ML Models
1000*5
❖ Several subsets of input data (bags)
Model Selection are created Bootstrap Aggravation
Bagging
❖ For each of the bags, a very deep
decision tree is grown
Feature Selection
❖ While making the decision, at every
split point in the decision tree the 750*5 750*5 750*5
learning algorithm looks through only
Feature random subsample of the feature DT DT DT
Engineering
space.
❖ The decision is made by looking the
Hyper- Gini Index value for each feature
parametrization
❖ The final output is the ensembled Output 1 Output 2 Output 3
output of each of the decision tree.
Testing &
Evaluation ❖ Less prone to overfitting Final
Output
 Neural Network (NN)
ML Models
❖ Utilizes the principle of universal
approximation
Yield
❖ Mimic the biological neurons signal
Model Selection transfer (Brain) W2
❖ Comprised of a node layers, (3*1)

Back Propagation
containing an input layer, one or

Feed Forward
Feature Selection more hidden layers, and an output h1/ h1/ h3/ Hidden
layer. a1 a2 a3
layer
❖ Each node has an associated
Feature weight and threshold.
Engineering ❖ If the output of any individual node W1
is above the specified threshold (5*3)
value, that node is activated,
Hyper- sending data to the next layer of the
parametrization x1 x2 x3 x4 x5
network
❖ Learning algorithms to optimize the
weights and bias of the model: Input layer
Testing &
Evaluation Gradient descent, ADAM etc
 Which ML model to select?
ML Models

Model Selection ❖ Tree based or neural network

Feature Selection
❖ Fully connected or sparsely connected

❖ Performance of multiple models needs

Feature
Engineering to be compared

Hyper- ❖ Model that works well for Tabular data

parametrization

Testing &
Evaluation
 Category Features Source
Satellite based Reflectance bands Sentinel-2, Landsat-8, MODIS
ML Models
Vegetation Indices-Greenness Sentinel-2, Landsat-8, MODIS
(NDVI, EVI, Red edge index)
Vegetation Indices-wetness Sentinel-2, Landsat-8, MODIS
(NDWI, LSWI, etc)
Model Selection
Radar backscatter (VH, VV, RVI Sentinel-1, EOS-4
etc)
Meteorological Rainfall, Rainy Days IMD gridded, CHIRPS, any other gridded/reanalysis
data
Feature Selection
Dry-spell/Wet-spell IMD gridded, CHIRPS, any other gridded/reanalysis
data

Feature Temperature IMD gridded, any other gridded/reanalysis data

Engineering
Growing degree days IMD gridded, CHIRPS, any other gridded/reanalysis
data
Heat wave/cold wave IMD gridded, CHIRPS, any other gridded/reanalysis
Hyper- data
parametrization
Bio-physical FAPAR PROVA/Sentinel-3, MODIS, Sentinel-2

LAI PROVA/Sentinel-3, MODIS, Sentinel-2

Testing & Edaphic Soil (texture, depth, AWC, etc) NBSSLUP (1: 250K, 1:50K)
Evaluation
DEM SRTM, DEM
Soil moisture SMAP, AMSR-E
 Satellite based crop condition indicators
ML Models Increasing Wavelength
Canopy Temperature, ET

Canopy Biomass
Model Selection Canopy wetness IR-Thermal
Chlorophyll, LAI
(LST, LST_EVI GPP) Microwave
SWIR (Radar
Visble-NIR (NDWI, LSWI) Backscatter)
Feature Selection (NDVI)

Feature
Engineering

Hyper-
parametrization

Testing &
Evaluation
 Crop mapping
ML Models
K = 0.81

❖ Features extraction for

Model Selection only specific crop pixels

❖ Multiple years crop maps

to be prepared
Feature Selection

❖ Only medium resolution

satellite data need to be
Feature K = 0.88 K = 0.93
Engineering
exploited for mapping

❖ Classification accuracies
Hyper- should be higher at
parametrization disaggregated level

Testing &
Evaluation
 Crop window
ML Models

Model Selection

Feature Selection

Feature
Engineering
Sep-2FN Oct-1FN Oct-2FN Nov-1FN Nov-2FN Dec-1FN Dec-2FN Jan-1FN Jan-2FN Feb-1FN Feb-2FN Mar-1FN Mar-2FN Apr-1FN Apr-2FN
❖ SOS and EOS needs to Early-wheat
Hyper-
parametrization be derived at Mid-wheat
disaggregated level Late-wheat
❖ Features should Gram
Testing & represent the crop Garlic
Evaluation
growing window Potato
 Need for feature engineering
ML Models GP: Guskara (m)
Dist: Purba Bardhaman

Model Selection

Feature Selection

Feature
Engineering

Hyper-
parametrization

Testing &
Evaluation
 Feature engineering: Example
ML Models

1 3 4 5
2
Maximum Red Maximum Max. Minimum
Model Selection Max. p o s i t i v e
NDVI Negative NDVI
Slope (average)
Bare soil at sowing (average o
f3 max.)
Slope Harvest crop or
preparation Fastest growth Maximum Fastest non green
of vegetation
biomass reduction of residue
Feature Selection greeness

Feature
Engineering

Hyper-
parametrization

Testing &
Evaluation
 Feature engineering: Example
ML Models

Model Selection Feature Extracted

-13
Feature Selection
1. Season Maximum VH
-15
(Smax VH)

2. Dynamic range of VH
VH

Feature -17
Engineering 332 Kg/Ha
(Range_VH)
-19 544 Kg/Ha

Hyper-
1025 Kg/Ha 3. Area under the VH
parametrization -21 1173 Kg/Ha curve (AUC)
1293 Kg/Ha
-23
Testing &
Evaluation
 Tuning of Hyper-parameters
ML Models
ML Model Hyper-parametrs
Random Forest •Number of trees
Model Selection •No. of features selected at each node
•Minimum Leaf size

Deep Neural Network ▪Network Architecture:

Feature Selection ▪Activation function {Sigmoid, Tanh, ReLu, L-ReLU): convert
input into non-linear
▪Learning Algorithm {GD, Momentum, Adam}:
Feature
Engineering
▪Learning rate {between 0-1}
▪Loss Function {RMSE, Cross-Entropy}
▪Regularization criteria {dropout, batch normalization,
Hyper- early stopping}
parametrization
▪Number of training epochs (number of cycle for model
tuning)
▪Batch size
Testing &
Evaluation
 Training and Validation approach
ML Models

❖ Model accuracy need to be tested based on MAE/RMSE/ NRMSE

Model Selection
❖ For model training and validation, stratified random sampling
approach from different yield classes need to be followed.
Feature Selection
❖ From each yield class range at least 30% randomly selected dataset
should be used for model validation and the rest for training.
Feature
Engineering ❖ A good quality training yield data (CCE/IU averaged) as vetted by
the state agricultural department from historical crop season should
be used.
Hyper-
parametrization
❖ The validated model needs to be run for the current season and the
final deliverables will be the modelled yield of insurance units for the
current and historical (at least 5 years) years.
Testing &
Evaluation
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
❖ All the major soybean growing districts of Unique clusters
Maharashtra are divided into 12 clusters based
upon AESR zones
❖ For each group, a separate ML model was trained
by using the crop yield data from 2017 onwards.
Dataset used
Datasets Source Feature derived
Temporal VH Sentinel 1 Monthly max VH from June to October Av. soybean yield
NDVI Sentinel-2 Season max NDVI from June to October between 2017-2020
LSWI Sentinel-2 Season max LSWI from June to October
Rainfall State Agril Monthly rainfall during June to October
Dept.
Rainy Days Derived from Monthly rainy days during June to October
rainfall
Profile NBSSLUP Profile Available water capacity
Available water
capacity
Soybean crop State Agril Soybean frop yield at circle level from 2017
yield Dept. onwards.
Soybean crop map derived from Sentinel-1 & Sentinel-2 data for all the years
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
Methodology

Temporal VH
Model Training
Weather data
(RF&RD) Data ML model training &
preprocessing validation
(gap filling, outlier •Model parameter optimization
• 2017-2020 soybean yield data
Smax NDVI removal, data
normalization etc)

Smax LSWI
Model deployment
PAWC Current Year Trained ML Soybean crop yield
datamatrix model estimate for current year
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
Model Hyper-parameters
DNN 18-36-18-9-1
Architecture
Activation Leaky ReLU
function
Based upon the accuracies
Learning Adam observed during model training
Algorithm process, these architecture were
Learning rate 0.001 further modified for each clusters
Loss Function RMSE
Regularization Dropout of 0.2

training epochs 500

Batch size 8
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
Results Model Performance
Cluster Training RMSE Validation
(Kg/Ha) RMSE (Kg/Ha)
Akola-Washim 228 283
Amravati-Yavatmal 165 240
Nagpur,Wardha 203 240
Jalgaon,Buldana 214 287
Parbhani-Jalna-Hingoli 228 259
Latur-Nanded 187 280
Osmanabad-Solapur 140 246
Beed,Ahmednagar-Pune 231 244

Satara-Sangli-Kolhapur 311 430

Nashik,Aurangabad 308 388
Nandurbar-Dhule 346 374
Chandarpur-Bhandara 293 305
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
Results
CCE reported soybean yield 2021 Ml predicted soybean yield
 Case Study
Machine Learning based Soybean yield estimation in Maharashtra
Results (2021)
Akola Amravati Latur

Nagpur
Osmanabad
 To Summarize
❖ Machine learning models can captures the non-linear relationships
between the yield and the features influencing the crop yield

❖ Proper workflow should be carried out which includes pre-processing,

feature design, splitting data into training and validation sets, selecting
machine learning algorithm, training, optimization, evaluations and testing

❖ Datasets/features must represent crop growth influencing factors

❖ Proper feature engineering should be carried out

❖ Regularization criteria must be followed to check the overfitting of the

model. Optimization function should be reflected as an output to visualize
the learning rate, batch size etc

❖ A good quality training yield data must be used