CROP MONITORING AND RECOMMENDATION

SYSTEM USING MACHINE LEARNING

TECHNIQUES
A PROJECT REPORT

Submitted by

S KRISHNA PRASAD (2013503514)

B SIVA SREEDHARAN (2013503525)
S JAISHANTH (2013503565)

in partial fulfillment for the award of the degree

of

BACHELOR OF ENGINEERING

IN

COMPUTER SCIENCE AND ENGINEERING

MADRAS INSTITUTE OF TECHNOLOGY, CHENNAI

ANNA UNIVERSITY: CHENNAI 600 025

April 2017
 BONAFIDE CERTIFICATE
Certified that this project report titled “CROP MONITORING AND
RECOMMENDATION SYSTEM USING MACHINE LEARNING
TECHNIQUES” is a bonafide work done by “S.KRISHNA PRASAD
(2013503514), B.SIVASREEDHARAN (2013503525) and S.JAISHANTH
(2013503565)” under my supervision, in partial fulfilment for the award of the
degree of Bachelor of Engineering in Computer Science and Engineering.
Certified further, that to the best of my knowledge the work reported here in
does not form part or full of any other thesis or dissertation on the basis of
which a degree or award was conferred on an earlier occasion to this or any
other candidate.

SIGNATURE SIGNATURE

Dr. P. ANANDHAKUMAR Dr. P.VARALAKSHMI

HEAD OF THE DEPARTMENT SUPERVISOR

Department of Computer Technology, Associate Professor,

Madras Institute of Technology, Department of Computer Technology,

Anna University, Madras Institute of Technology,

Chennai-600 044. Anna University,

Chennai-600 044.

Date:

Place

i
 TABLE OF CONTENTS

CHAPTER TITLE PAGE

NO. NO.
ABSTRACT iii
LIST OF ABBREVIATIONS vii
LIST OF FIGURES viii
LIST OF TABLES ix
1 INTRODUCTION 1
1.1 MACHINE LEARNING 1
1.1.1 Types of Problems and tasks 2
1.1.2 Support Vector Machine 2
1.1.2.1 Advantages 3
1.1.3 Spatial Spectral classification 4
1.1.4 Simple Linear Iterative Clustering 4
1.2 ADABOOST MODEL 6
1.3 MULTIPLE MORPHOLOGICAL COMPONENT 7

ANALYSIS
1.4 OBJECTIVE OF PROPOSED WORK 7

1.5 SCOPE 8

1.8 THESIS OVERVIEW 8

2 LITERATURE SURVEY 9
2.1 IMAGE DENOISING 9
2.2 IMAGE CLASSIFICATION 10
2.3 CROP CLUSTERING 11
2.4 WAVELET TRANSFORMS 13
2.5 PREDICTION ALGORITHM 14

2.6 OVERVIEW 16

iv
3 PROPOSED WORK 17
3.1 PROPOSED SYSTEM ARCHITECTURE 17
3.2 MODULE DESIGN 18
3.2.1 Image pre-processor 18
3.2.2 Image Segmentation – SLIC 18

3.2.3 SVM prefixed by SSSE 18

3.2.4 Image Classifier Evaluator 18

3.2.5 Recommendation 19

3.2.6 Calculating NDVI 19

3.2.7 NDVI VS LAI Relationship 19

3.2.8 Yield Calculation based on NDVI 19

3.3 SYSTEM DEVELOPMENT 19

3.4 ALGORITHM AND WORKING OF EACH MODULE 21

3.4.1 Support Vector Machine 21

3.4.2 Image pre-processor 22

` 3.4.2.1 Algorithm Used: Multiple morphological 21

component analysis
3.4.3 Image Classifier Evaluator 25

3.4.4 Recommendation Model Evaluator 32

4 IMPLEMENTATION AND RESULTS 34

4.1 TOOLS 34
4.1.1 MATLAB 34
4.1.1.1 Features 34
4.1.2 PyCharm 35
4.1.2.1 Features 35

4.1.3 Matplotlib 36
4.2 DEPLOYMENT DETAILS 37

v
 4.3 DATASET 37
4.4 PRPROCESSOR 39
4.5 IMAGE CLASSIFIER 41
4.6 CROP YIELD PREDICTION 43
4.7 RELEVANCE PREDICTION 44
4.8 PERFORMANCE MEASURES 46
4.9 YIELD PREDICTION 47
4.10 CLASSFICATION 49
5 CONCLUSIONS AND FUTURE WORK 53
6 REFERENCES 54

vi
 LIST OF ABBREVIATIONS

Abbreviation Expansion
CNN Convolutional Neural Network
SVM Support Vector Machine
NDVI Normalized difference vegetation index
LAI Leaf Area Index
MMCA Multiple morphological component analysis
FBS Feature Band Extraction
KNN K Nearest neighbour
MODIS Moderate-resolution Imaging Spectroradiometer
RBF Radial Basis Function
SSSE Spatial Spectral Schrodinger Eigenmaps
LE Laplacian Eigenmaps
SLIC Simple Linear iterative clustering
OOC Object oriented classification
CPS Class Pair separability
DMP Differential morphological profiles
SAR Synthetic aperture radar

SLIC Simple Linear Iterative Clustering

MMCA Multiple Morphological Component Analysis
SE Schrodinger Eigen maps
SM Shi – Malik
GB Gilles – Bowles
HZYZ Hou-Zhang- Ye-Zheng
BE Benedetto et al

vii
 LIST OF FIGURES

FIGURE TITLE PAGE NO.

NO.

3.1 Architecture Diagram 17

3.2 Flow Diagram 20

4.1 Pre-processor Module 39

4.2 Original image 40

4.3 Mid-Stretch Filter 40

4.4` Noise removal using grey 41

4.5 Classification Image Results 42

4.6 Classification Accuracy Measures 43

4.7 Training data execution results 44

4.8 Crop Yield Result 44

4.9 Crop soil data set 45

4.10 Relevance prediction for red soil 45

4.11 Training set regression analysis 47

4.12 Testing Regression analysis 47

4.13 True vs Predicted Yield Analysis 48

4.14 Overall Accuracy 50

4.15 Average Specificity 50

4.16 Average Precision 51

4.17 Average Accuracy 51

4.18 Average Sensitivity 52

viii
 LIST OF TABLES

TABLE NO. TITLE PAGE NO.

4.1 DATASET 38

4.2 CLASSIFICATION ACCURACIES 49

4.3 ANALYSIS MATHODS 49

ix
 ABSTRACT

This document proposes a crop recommendation system using Spectral

spatial classification and Support Vector Machine (SVM).The farmer provides

the crop field image as an input to the application. In the pre-processing stage,

Denoising is done using Multiple morphological component analysis (MMCA)

and as a result, filtering the image retaining its necessary portions. SVM

prefixed by Spatial Spectral Schrodinger Eigen Maps (SSSE) is used as a

classification method wherein partial knowledge propagation is leveraged to

improve the classification accuracy.

The classified image along with the Ground truth statistical data

containing the weather, crop yield, state & county wise crops are used to predict

the yield of a particular crop under a particular weather condition. This

predictive model used AdaBoost classifier. Crop recommendation is facilitated

then by collaborative filtering. Further scope of the project would extend to

predictive analytics on the commodity market of the goods grown in the

agricultural fields to predict its waxing and waning.

iii
 ACKNOWLEDGEMENT

We are highly indebted to our respectable Dean, Dr. A. RAJADURAI and to

our reputable Head of the Department Dr. P. ANANDHAKUMAR,

Department of Computer Technology, MIT Campus, Anna University for

providing us with sufficient facilities that contributed to success in this

endeavor.

We would like to express my sincere thanks and deep sense of gratitude to our

Supervisor, Dr. P. VARALAKSHMI for her valuable guidance, suggestions

and constant encouragement which paved way for the successful completion of

this phase of project work.

We would be failing in our duty, if we forget to thank all the teaching and non-

teaching staff of our department, for their constant support throughout the

course of our project work.

S.KRISHNA PRASAD
B.SIVA SREEDHARAN
S. JAISHANTH

ii
 Chapter 1
Introduction

This chapter gives a description about machine learning, tasks involved,

models, advantages, disadvantages, objective of the proposed work and thesis
overview.
This document describes a crop yield prediction and recommendation
system using spatial spectral classification and AdaBoost classifier. The classified
crop field images along with the historical weather and yield data are modelled to
obtain the predicted crop yield and recommend suitable crops for a particular field.
This in turn involves machine learning and image processing for classification and
prediction.
1.1 Machine Learning
Machine Learning is a field of Computer Science, where new developments
evolve at recent times, and also helps in automating the evaluation and processing
done by the mankind, thus reducing the burden on the manual human power.
According to techtarget, Machine learning is a type of artificial intelligence (AI)
that provides computers with the ability to learn without being explicitly
programmed.
Machine learning focuses on the development of computer programs that
can change when exposed to new data. Finding out the suitable crops based on the
soil’s appearance becomes tedious for novice farmers. There also exists a need to
prevent the agricultural decay.
Effective utilization of agricultural land is crucial for ensuring food security
of a country. In this document, we propose a crop recommendation system using
spectral spatial classification and AdaBoost meta-algorithm.

1
1.1.1 Types of Problems and tasks

Machine learning tasks are typically classified into three broad categories,
depending on the nature of the learning "signal" or "feedback" available to a
learning system. These are

 Supervised learning: The computer is presented with example

inputs and their desired outputs, given by a "teacher", and the
goal is to learn a general rule that maps inputs to outputs.

 Unsupervised learning: No labels are given to the learning

algorithm, leaving it on its own to find structure in its input.
Unsupervised learning can be a goal in itself (discovering
hidden patterns in data) or a means towards an end (feature
learning).

 Reinforcement learning: A computer program interacts with a

dynamic environment in which it must perform a certain goal
(such as driving a vehicle), without a teacher explicitly telling it
whether it has come close to its goal. Another example is
learning to play a game by playing against an opponent

1.1.2 Support Vector Machine (SVM)

SVM constructs a hyperplane or set of hyperplanes in a high- or infinite-

dimensional space, which can be used for classification, regression, or other
tasks. Intuitively, a good separation is achieved by the hyperplane that has the

2
largest distance to the nearest training-data point of any class, since in general the
larger the margin the lower the generalization error of the classifier.

The computational load should be sensible, the mappings are used by the
SVM scheme to ensure the dot products will be computed in terms of the variable
in the original scope, for that a kernel function k(x,y) selected to get the optimal
computational time.

The higher-dimensional space in the hyper planes is distinct as the set of

points whose dot product with a vector in that space is constant. These vectors in
the hyper planes defining the hyper planes can be chosen to be linear
combinations with parameters of images of feature vectors that occur in the data
base. With this choice of a hyperplane, the points in the feature space that are
mapped into the hyperplane are defined by the relation: The equation of the
output from a linear SVM is

u=w.x–b

Where w is the normal vector of the hyperplane, and x is the input vector.

1.1.2.1 Advantages

Support vector machine is one of the most widely used classification

algorithms due to the advantages it enjoys which are as follows:

 SVMs are helpful in text and hypertext categorization as their application

can significantly reduce the need for labeled training instances in both the
standard inductive and transductive settings.
 Classification of images can also be performed using SVMs. Experimental
results show that SVMs achieve significantly higher search accuracy than
traditional query refinement schemes after just three to four rounds of

3
 relevance feedback. This is also true of image segmentation systems,
including those using a modified version SVM.
1.1.3 Spatial Spectral Classification (SSSE)

Schrodinger Eigenmaps (SE) has recently arose as a powerful graph-based

system for semi-supervised manifold learning and recovery. The spatial spectral is
extended the Laplacian of a graph which is constructed from hyper spectral
imagery to include barrier or cluster potentials, Schrodinger Eigenmaps will
enable machine learning techniques that employ the expert or label the
information provided at a subset of pixels.

The Non - diagonal potentials can be used within the Schrodinger

Eigenmaps framework in a way that allows for the incorporation of spatial and
spectral information in unsupervised manifold learning and recovery. The non -
diagonal potentials scramble spatial proximity then the output is combined with the
spectral proximity information in the original graph which will yields a framework
that is competitive with state-of-the-art spectral and spatial fusion approaches for
clustering and subsequent classification of hyperspectral image data.

1.1.4 Simple Linear Iterative Clustering (SLIC)

Simple Linear Iterative Clustering is a simple and efficient method to

decompose an image in visually homogeneous regions. It is based on a spatially
localized version of k-means clustering. SLIC is similar to mean shift or quick
shift, where each pixel is associated to a feature vector and then k-means clustering
is run on those.

4
 The spatial and appearance components of the feature vectors are imposed a
degree of spatial regularization to the extracted regions are balanced by the
coefficient λ.

The nominal size of the region is regionSize and the strength of the spatial
regularization is regularizer. The image is first divided into a grid with
step regionSize. The center of each grid tile is used to initialize a corresponding k-
means up to a small shift to avoid image edges. By using the Lloyd algorithm, the
k-means centers and clusters are refined which will yield the segmented image

For the restriction and simplification in Simple Linear Iterative Clustering,

during the k-means iterations each pixel can be assigned to only the 2 x 2 centers
corresponding to grid tiles adjacent to the pixel.

The trade-off between clustering appearance and spatial regularization is set

by the parameter regularizer. The coefficient λ is obtained by using

Simple Linear Iterative Clustering optionally removes any segment after the
k-means step whose area is smaller than a threshold minRegionSize by merging
them into larger ones.

K-means uses the standard LLoyd algorithm alternating assigning pixels to

the closest centers a re-estimating the centers as the average of the corresponding
feature vectors of the pixel assigned to them. The only difference compared to
standard k-means is that each pixel can be assigned only to the center originated
from the neighbour tiles. This guarantees that there are exactly four pixel-to-center
comparisons at each round of minimization.
5
 SLIC eliminates any connected region after k-means has converged whose
area is less than minRegionSize pixels. This is done by avidly merging regions to
neighbour ones: the pixels by pixels are scanned in lexicographical order and the
corresponding connected components are visited. If a region has already been
visited, it is skipped; if not, its area is computed and if this is less
than minRegionSize its label is changed to the one of a neighbour region at p that
has already been visited.

1.2 AdaBoost Model

Boosting can be defined as a means for creating a strong classifier using a

set of weaker classifiers. It can be explained as a method in which a second model
is use for rectifying the errors in the first model i.e., the training model. AdaBoost
is used in this context for binary classification. The first model constituting the
training dataset is weighted. The primary weight is modelled to:

weight (yi) = 1/n

Where yi is the ith training instance and n is the total number of instances.

Weak models are sequentially added and trained using the training data as
calculated by the above formula. The process is continued until a predefined
number of weak learning instances have been created or no further enhancement
can be made on the training data set.

Predictions are made by calculation of the weighted averages if the weak

classifiers. The predicted values are weighted by each weak learners stage value.
The prediction for the ensemble model is calculate as a the sum of the weighted

6
prediction values. If the sum is +ve, then the first class is predicted, else the second
class is predicted.

1.3 Multiple Morphological Component Analysis (MMCA)

Remote sensing data has multiple issues related to the dimensionality of the
original set data. This happens due to the large spectral resolution in the
hyperspectral images. This dimensionality reduction uses the minimum noise
fraction (MNF) to reduce the total existing components to a small number of
bands. MMCA aims at decomposing the remote sensing image into texture and
smoothness tuples using feature extraction as a prefix method. These tuples convey
the essence of the original image as a linear combination as shown below:

Y= ya + yb + n

Where n is the approximation image residue value. a and b denote the

smoothness and texture components respectively. Content parameters, coarseness
directional vectors and contrast are the additional parameters to be taken into
account with respect to MMCA.

1.4 Objective of proposed work

This project aims at predicting the crop yield at a particular weather

condition and thereby recommending suitable crop for that field. It involves the
following steps.

 Collect satellite images for agricultural crop monitoring.

7
  Classify the image based on Soil type, moisture content, weather conditions,
pH value, organic nitrogen etc.

 Perform satellite image processing with respect to tex tural and spatial
features.

 Analyze crop patterns with the help of past records and map them with
calculated data.

 Monitor crop yield and find ways for increasing it.

 Recommend profitable crops for each land type.

1.5 Scope
Agriculture is the backbone for a developing economy like India and there is
an enormous need to maintain the agricultural sustainability. Hence it is a
significant contribution towards the economic and agricultural welfare of the
countries across the world.

1.6 Thesis overview

The rest of the thesis is organized as follows. Chapter 2 gives an overview of
related work. Chapter 3 describes the proposed architecture and details of the
machine learning & image processing models used for evaluation. The
implementation and experimental results are given in Chapter 4. Chapter 5
summarizes the overall work and future work.

8
 Chapter 2
Literature Survey

2.1 Image denoising

Xiang Xu et.al. (2016) present a new technique based on multiple
morphological component analysis (MMCA) that exploits multiple textural
features for decomposition of remote sensing images. The proposed MMCA
framework separates a given image into multiple pairs of morphological
components (MCs) based on different textural features, with the ultimate goal of
improving the signal-to-noise level and the data separability. A distinguishing
feature of our proposed approach is the possibility to retrieve detailed image
texture information, rather than using a single spatial characteristic of the texture.
In this paper, four textural features: content, coarseness, contrast, and directionality
(including horizontal and vertical), are considered for generating the MCs.

Hemant Kumar et. al. (2016) state how Hyperspectral unmixing is the
process of estimating constituent endmembers and their fractional abundances
present at each pixel in a hyperspectral image. A hyperspectral image is often
corrupted by several kinds of noise. This work addresses the hyperspectral
unmixing problem in a general scenario that considers the presence of mixed noise.
The unmixing model explicitly takes into account both Gaussian noise and sparse
noise. The unmixing problem has been formulated to exploit joint-sparsity of
abundance maps. A total-variation-based regularization has also been utilized for
modeling smoothness of abundance maps. The split-Bregman technique has been
utilized to derive an algorithm for solving resulting optimization problem. Detailed
experimental results on both synthetic and real hyperspectral images demonstrate
the advantages of proposed technique.

9
 Gabriela Ghimpe teanu et.al. (2016) explain the image decomposition model
that provides a novel framework for image denoising. The model computes the
components of the image to be processed in a moving frame that encodes its local
geometry (directions of gradients and level lines). Then, the strategy we develop is
to denoise the components of the image in the moving frame in order to preserve
its local geometry, which would have been more affected if processing the image
directly. Experiments on a whole image database tested with several denoising
methods show that this framework can provide better results than denoising the
image directly, the index metrics is similar in terms of peak signal-to-noise ratio.

2.2 Image classification

Xia Zhang et.al.(2016) introduced the new crop classification method which
involves the construction of the vegetation feature band set (FBS) and optimization
object-oriented classification (OOC). The model also computes additional 20
spectral indices sensitive as parameters for feature band set to distinguish specific
vegetation. For reducing the data redundancy Class pairs separability (CPS) to
improve the separability between class pairs. The proposed algorithm shows that
the crop classification accuracy will improve significantly, reduce edge effects and
textural features combined with spectral indices that give sensitive to the
chlorophyll, carotenoid and Anthocyanin indicators which also contribute
significantly to improve the crop classification. Therefore, it is an effective
approach for classifying crop species, monitoring invasive species, as well as
precision agriculture related applications.

Xin Huang et. al. (2016) lists the differential morphological profiles (DMPs)
that are widely used for the spatial/structural feature extraction and classification
of remote sensing images. The DMPs shows the response of the image structures

10
that are related to different scales and sizes of the structural elements
(SEs).Traditional DMPs will ignore the discriminative information for features that
are across the scales in the profiles. The proposed model will scale span with
differential profiles to obtain the entire differential profiles. GDMPs used to obtain
the complete shape spectrum and measure the difference between arbitrary scales.
Since the random forest is used to interpret GDMPs for high dimensionality data
and ability of evaluating the importance of variables. The random forest errors are
used to quantify the importance of each channel of GDMPs and also discriminative
information for the entire profiles.

Kunshan Huang et.al.(2016) proposed a novel spectral–spatial hyper spectral

image classification method based on K nearest neighbor (KNN). The proposed
method consists of support vector machine which is used to obtain the initial
classification probability maps. The obtained pixel-wise probability maps are
refined with the proposed KNN filtering algorithm that is based on matching and
averaging non local neighborhoods. The model does not need any particular
segmentation and optimization strategies which use the nonlocal principle of real
images by using KNN and gives the competitive classification with fast
computation. Several experiments performed on the hyper spectral data sets in
which the result show that the classification results obtained by the proposed
method are comparable to several recently proposed hyper spectral image
classification methods.

2.3 Crop clustering

Wei Yao et.al. applies evaluates a modified Gaussian-test-based
hierarchical clustering method for high resolution satellite images. The purpose of
the model to obtain homogeneous clusters within each hierarchy level which later

11
allow the classification and annotation of image data ranging from single scenes up
to large satellite data archives. After cutting a given image into small patches and
feature extraction from each patch, k-means are used to split sets of extracted
image feature vectors to create a hierarchical structure. As image feature vectors
usually fall into a high-dimensional feature space, we test different distance
metrics, to tackle the “curse of dimensionality” problem. By using three different
synthetic aperture radar (SAR) and optical image datasets, Gabor texture and Bag-
of-Words features are extracted, and the clustering results are analyzed via visual
and quantitative evaluations.

Michael D. Johnson et.al developed crop yield forecast models for barley,
canola and spring wheat grown on the Canadian Prairies for vegetation indices
which is derived from satellite data and machine learning methods. The model use
hierarchical clustering to group the crop yield data from Census Agricultural
Regions (CARs) into several larger regions for building the forecast models. The
Moderate-resolution Imaging Spectro radiometer (MODIS) derived Enhanced
Vegetation Index (EVI), Advanced Very High Resolution Radiometer (AVHRR)
derived Normalized Difference Vegetation Index (NDVI) and these model are
considered as predictors for crop yields. Multiple linear regression(MLR) and
machine learning models – Bayesian neural networks (BNN) and model-based
recursive partitioning (MOB) – were used to forecast crop yields, with various
combinations of MODIS-NDVI, MODIS-EVI and NOAA-NDVI as predictors.

R.B. Arango et.al. (2016) focuses on a methodology for the automatic

delimitation of cultivable land by means of machine learning algorithms and
satellite data. The method uses a partition clustering algorithm called Partitioning
Around Medoids and considers the quality of the clusters obtained for each

12
satellite band in order to evaluate which one better identifies cultivable land. The
proposed method was tested with vineyards using as input the spectral and thermal
bands of the Landsat 8 satellite. The experimental results show the great potential
of this method for cultivable land monitoring from remote-sensed multispectral
imagery.

2.4 Wavelet transforms

M. Krishna Satya Varma et.al.(2016) state and create thematic maps of the
land wrap present in an image, the classified data thus obtained may then be used.
Classification includes influential an appropriate classification system, selecting,
training sample data, image pre-processing, extracting features, selecting
appropriate categorization techniques, progression after categorization and
precision validation. Aim of this study is to assess Support Vector Machine for
efficiency and prediction for pixel-based image categorization as a contemporary
reckoning intellectual technique. Support Vector Machine is a classification
procedure estimated on core approaches that was demonstrated on very effectual in
solving intricate classification issues in lots of dissimilar appropriated fields. The
latest generation of remote Sensing data analyzes by the Support Vector Machines
exposed to efficient classifiers which are having amid the most ample patterns.

Olivier Regniers et.al.(2016) explore the potentialities of using wavelet-

based multivariate models for the classification of very high resolution optical
images. A strategy is proposed to apply these models in a supervised classification
framework. This strategy includes a content-based image retrieval analysis applied
on a texture database prior to the classification in order to identify which
multivariate model performs the best in the context of application. Once identified,
the best models are further applied in a supervised classification procedure by

13
extracting texture features from a learning database and from regions obtained by a
pre segmentation of the image to classify. The classification is then operated
according to the decision rules of the chosen classifier. The use of the proposed
strategy is illustrated in two real case applications using Pléiades panchromatic
images: the detection of vineyards and the detection of cultivated oyster fields. In
both cases, at least one of the tested multivariate models displays higher
classification accuracies than gray-level co-occurrence matrix descriptors.

N. Prabhu et.al. (2016 have presented a series of experiments to investigate

the effectiveness of some wavelet based feature extraction of hyper spectral data.
The model introduces the three types of wavelets have been used which are Haar,
Daubechies and Coiflets wavelets and the quality of reduced hyper spectral data
has been assessed by determining the accuracy of classification of reduced data
using Support Vector Machines classifier. The hyper spectral data has been
reduced up to four decomposition levels. Among the wavelets used for feature
extraction Daubechies wavelet gives consistently better accuracy than that
produced from Coiflets wavelet. The two -level decomposition is capable of
preserving more useful information from the hyper spectral data. Furthermore,
two-level decomposition takes less time to extract features from the hyper spectral
data than one-level decomposition.

2.5 Prediction algorithm

Monali Paul et.al.(2016) analysis the Soil Behaviour and Predicted the
Crop Yield using Data Mining approach. To predict the proper selection of crops
various yield prediction algorithm used to help farmers. Traditional yield
prediction was performed by considering the farmer's experience on a particular
field and crop. The proposed system uses data mining techniques in order to

14
predict the category of the analyzed soil datasets. The category which is predicted
will indicate the yielding of crops. The problem of predicting the crop yield is
formalized as a classification rule, where Naive Bayes and K-Nearest Neighbor
methods are used.

Yvette Everingham et.al.(2016) predict the accurate prediction of sugarcane

yield using a random forest algorithm. A data mining method like random forests
can cope with generating a prediction model when the search space of predictor
variables is large. Researcher investigated that the accuracy of random forests to
explain annual variation in sugarcane productivity and the suitability of predictor
variables generated from crop models coupled with observed climate and seasonal
climate prediction indices is limited. Simulated biomass from the APSIM
(Agricultural Production Systems sIMulator) sugar-cane crop model, seasonal
climate prediction indices and observed rainfall, maximum and minimum
temperature, and radiation were supplied as inputs to a random forest classifier and
a random forest regression model to explain annual variation in regional sugarcane
yields.

Benjamin Dumont et.al.(2016) predict the Assessing the potential of an

algorithm based on mean climatic data to predict wheat yield .This paper presents
a methodology that addresses the problem of unknown future weather by using a
daily mean climatic database, based exclusively on available past measurements. It
involves building climate matrix ensembles, combining different time ranges of
projected mean climate data and real measured weather data originating from the
historical database or from real-time measurements performed in the field. Used as
an input for the STICS crop model, the datasets thus computed were used to
perform statistical within-season biomass and yield prediction. This work
15
demonstrated that a reliable predictive delay of 3–4 weeks could be obtained. In
combination with a local micrometeorological station that monitors climate data in
real-time, the approach also enabled us to (i) predict potential yield at the local
level, (ii) detect stress occurrence and (iii) quantify yield loss (or gain) drawing on
real monitored climatic conditions of the previous few days.

2.6 Overview
Hence, wavelet algorithms can be leveraged for denoising the image by
means of dimensionality reduction. Crop classification involves calculating the
NDVI and LAI as a means for training the model for existing greenery percentage.
K Nearest neighbour is also of prime importance for graph edge construction in
case of fusion between locational and intensity components. Classification is
improved by means of taking into account class pair separability for object
oriented classification. The above mentioned aspects are kept in mind for building
our proposed architecture of the project thereby eradicating the limitations of
algorithms mentioned above.

16
 CHAPTER 3
PROPOSED WORK
This chapter gives a detailed description about the architecture and
algorithms used in the proposed work.
3.1 Proposed system architecture
The overall block diagram of the system is shown in the figure 3.1. The
image pre-processor and the image segmentation does the same task of segmenting
the image to get a vivid soil portion from the image, as it may contain unwanted
portions which may make the system to work with decreased efficiency. The SSSE
module takes the pre-processed image as input and finds the type of the soil and
returns the soil class, which is then given to SVM along with latitude and longitude
values.

Figure 3.1 Architecture diagram

17
 The system eventually aims at predicting the yield of crops based on the set
of whether and yield data including geographical parameters.

3.2 Module design

The working of the recommendation system incorporates the following
modules for orderly functioning.
3.2.1 Image preprocessing
The pre-processor uses multiple morphological component analysis
(MMCA) for dimensionality reduction by means of content, coarseness, directional
vectors and contrast features. This works by finding the possible contours with the
specification to identify the soil areas.
3.2.2 Image Segmentation - SLIC

This module leverages the creation of superpixels using Simple linear

iterative clustering (SLIC). In this method, the preprocessed images are taken as
input and segmented crop field image is produced as the result.
3.2.3 SVM prefixed by SSSE
This module is the primary analysis module of the project. The SVM
classifier is the backend classifier which takes the input of the Spectral spatial
Schrodinger Eigen maps. This involves considering both spectral and spatial
parameters in addition to textural features for partial knowledge and knowledge
propagation.
3.2.4 Image classifier evaluator
This module takes in the pre-processed input image at the end and uses the
pickled historical data sets of whether and crop yield parameters to predict the soil
class and yield values.

18
3.2.5 Recommendation model evaluator
This module takes in the soil class and GPS parameters as input and also
uses the Recommendation Model to recommend the crop suitable for the image
and historical data set provided.

3.2.6 Calculating NDVI

The normalized difference vegetation index (NDVI) is a simple graphical
indicator that can be used to analyze remote sensing measurements and assess
whether the target being observed contains live green vegetation or not.

3.2.7 NDVI vs LAI relationship calculation

Accurate estimation of LAI is important for monitoring vegetation
dynamics, and LAI information is essentially required for the prediction of
microclimate and various biophysical processes within and below canopy.

(LAI = 0.128 * exp (NDVI/0.311))

3.2.8 Yield calculation based on NDVI

The relationship between NDVI and yield of the data analyzed, indicates the
possibility of considering agrometeorological conditions to obtain accuracy in
yield estimation.

3.3 System development

The workflow of the system is shown in figure 3.2. The module descriptions,
algorithms used with analysis methods and tools used to achieve the result are as
follows
.

19
 Figure 3.2 Flow diagram

This system has been developed using WinPython 2.7 which includes the
packages such as: NumPy, SciPy, Pandas, Theano , Sklearn, Matplotlib and
OpenCV.

The overview of the algorithm for the system is as given below.

Initially Img is assigned the input image, Lon is assigned with longitude and
Lat is assigned with latitude. This is fed to SVM prefixed by SSSE classifier

20
module which in turn invokes the AdaBoost and Collaborative filtering module
modules to predict the crop yield and preferable crop respectively. Then the soil
class together with Lat and Lon are used to predict the suitable crop. Algorithm 1
explains the functioning of the entire recommendation system.

Img ←Input Soil Image

Lon ←Longitude
Lat ←Latitude
PREDICT(Img, Lon, Lat):
1. soil class ←IMAGE CLASSIFIER EVALUATOR(Img)
2. suitable crop ←RECOMMENDATION MODEL
EVALUATOR(soil class, Lon, Lat)

Algorithm 1 : Overall Working

3.4 Algorithms and working of each module

The algorithms used under each module are explained using pseudocode
representations and performance analysis measures as follows.

3.4.1 Support vector machine

SVM finds its place in this work for training the Recommendation system
with training set. It is additionally used after the classification using SSSE.

Input: Training set containing suitable crops for given soil class and
Latitude and longitude parameter from SSSE module.

21
 Output: Recommendation Model Evaluator.

The working algorithm of SVM is explained below as Algorithm - 2

1. DS ←dataset of soil images with soil classes as labels

2. SVM[i] ←one vs rest Support Vector Machine for class i
3. for row in D
4. SVM[row.class].train(row)

Algorithm 2 : SVM
3.4.2 Image pre – processer
Preprocessing plays a crucial role in denoising the crop field image by
dimensionality reduction of components by means of wavelet algorithms. The
proposed MMCA framework is elucidated in Algorithm 3.
Input: Crop field Image

Output: Processed soil image

1. I ←input image
2. Features[] ← Extract{content,coursness,direction,contrast}
3. Proc image ←MMCA (I, Fearures[])
4. Return Proc image

Algorithm 3: Preprocessing

3.4.2.1 Algorithm Used: Multiple morphological component analysis

Morphological component analysis allows to separate features contained in
an image when those tuples represent different morphological aspects. Considering
N
N as the number of pixels such that y ∈ R , the aim is to generate the sparsest
solution for the noise removal problem as follows:

22
 x = arg min||x||i
K
where x ∈ R denotes the sparse coefficients of the MC. K represents the number
of atoms in the dictionary of pixels (typically K>N). The resultant is the separation
fo textural and content components as a result of decomposition by MCA.

Smoothness and texture component represent the original image under a

linear combination as follows:

y= ys + yt + n
The limitation of this representation is the ignorance in including the spatial
components which are in abundance in the remote sensing images. MMCA aims at
exploring the entire scope of spatial textural features to aid in noise removal ie
denoising. Hence, as a result, 4 new textural features: content, coarseness, contrast
and directionality are added in this linear combination. Algorithm 4 clearly
expounds the proposed MMCA algorithm.

1. Randomly choose several portions from the image

for initializing the smoothness and texture
component dictionaries.
2. Build dictionaries for ys and yt based on some
transitions on the image snippets chosen.
3. Perform sparse coding to learn the Morphological

Algorithm 4: MMCA

Issues in the existing MCA algorithm

The construction of dictionaries for the 2 components creates a large over
heard in memory and time.

23
Enhancement made to accommodate the limitations without overhead
In addition to the existing algorithm components like smoothness and
texture, the below mentioned 4 components make the entirety of the MMCA
dimensionality reduction procedure.

Content tuple: Anisotropic structures, smooth curves and varied length

edges are extracted using wavelet transforms and local ridgelet transforms. To
build the dictionary from the randomly chosen mage snippets, cosine transform is
made use of.
Coarseness tuple: It’s a Gaussian distribution based dictionary construction
mechanism wherein filtering replaces the intensity value at each pixel of an image
with the weighted average values of intensity with respect to the neighbouring
pixels. The radiometric differences between pixels are also taken into account
which makes possible procedural looping and a s a result regulating the weights to
the nearby pixels accordingly.
Contrast tuple: The variance of the grey scale distribution wherein high and
low contrast values form the rapid and slow intensity anomalies form the crux f the
contrast tuple. The dictionary construction mechanism involves achieving both
high and low contrast dictionaries by means of anisotropic diffusion.
Directionality tuple: It involves the orientation of the textural features into
horizontal and vertical vectors. Stationary wavelet transform is leverages to
complete the task of decomposing the texture into different direction components.

Line likeliness, regularity, roughness and scattering ability are the other
textural features which are in the future scope of the denoising algorithm.

24
 Considering DS and DT to be the dictionaries for the smoothness and texture
component and xS and xT be the sparse coefficients for the same. Therefore
according to the linear combination of the dual components: in Equation 1

Y = yS + yT + n = DSxS + DTxT + n (1)

yS and yT are calculated as a result of the optimization problem as follows in

Equation 2 :

< , > = 𝐚𝐫𝐠 𝐢 , ,

|| − − || + 𝝀 ||𝑻 || + 𝝀 ||𝑻𝑻 𝑻 || (2)

where λ1 and λ2 are the regularization parameters. DS and DT are serially updated
at each iteration using total variation and regularization threshold.

Issues in the proposed work

To fully retrieve the image texture information, the inability to incorporate
varied textural features is a marked overhead since a single spatial feature is not
always enough in case of different kinds of remote sensed imagery. Selection of an
optimal feature among the multiple tuples for a particular image is a possibility to
be analyzed in future.

3.4.3 Image classifier evaluator

This module evaluates the processed soil image to identify the crop class
for the particular field.
Input: Processed Soil Image

25
 Output: Soil class of the image

1. proc image ←preprocessed image

2. SSSE ←Spectral + Spatial barrier parameters
3. N ←number of potentials in the column of
barriers
4. YiSSSE ←Classification of SSSE for the class i
5. Yi ← Prediction SSSE for the class i
6. Yi ←1/N (ΣSSSE YiLE)
7. return argmaxi(Yi)

Algorithm 5: SSSE

Existing algorithm: Laplacian Eignenmaps

It’s a non-linear dimensionality reduction algorithm which preserves

computational efficiency and locality preservation properties. The affinity to
clustering algorithms is evident in the flow of the functioning of the algorithm.

Steps involved

1. Build an undirected graph G = (X, E) where the vertices are the set of piints
in X and edges E defined based on the spatial proximity of the vertices. For
this proximity calculation, ∈ - neighborhoods and K nearest neighbor search
is made use of.

2. The weights for each edge are defined in E. The heat kernel based on the
Euclidean distance between the pixels is the common method for defining
weights as explained in Equation 3.

|| − ||
, = 𝐞𝐱𝐩 (− ) (3)
𝛔

26
 if the edge exists or Wi,j = 0 otherwise.

3. The smallest m+1 eigenvalues an eigenvectors are calculated using the

conventional eigenvector problem Lf = λDf where D is the diagonal
weighted matrix by Di,j = Ʃ jWi,j and L= D – W is the laplacian matrix.
The eigenvectors found f0,…..,fm are serially arranged in order so that

= 𝝀 ≤ 𝝀 ≤ .…… ≤ 𝝀

Then the points y1T , y2T , ..., y3T are defined to be the rows of

𝑭= [ , ,……, ]

Schroedinger Eigenmaps

It can be elucidated as a variance of the LE algorithm in incorporate a

potential matrix V. The eigenvector problem is modified as in Equation 4:

(L + αV)f = λDf (4)

Where α is the weight parameter to prioritize the contributions of the

laplacian and potential matrices. Barrier and cluster potentials are both
analyzed in this mechanism. A non-negative diagonal matrix is used as
barrier potentials wherein the +ve values are forcing the points in y towards
the origin. Cluster potentials are defined as a entirety of non-diagonal
matrices using equation 5.

, , ∈{ , , , }
,
= { − , , ∈{ , , , } (5)
,

27
 The crucial advantages of SE over LE is that semi supervised
clustering is made possible by the potential matrix V. K means algorithm is
invoked as the ending classifier model.

Spectral Spatial fusion in Schrodinger algorithm

A combination of the spatial (location) and spectral (intensity) is

incorporated in this dimensionality reduction method. This happens by
p f
concatenating both the spatial location xi and spectral xi .

xiT = [ xif, xip ]

Analysis of multiple fusion methods

The analysis differs in the nature of how edge weights are defined in each
method.

Shi - Malik

1. The graph is constructed based on the spatial neighborhood

ie €- neighborhood between edges in such a way that an edge is

defined between xi and xj if ||xi – xj ||2 < €.
p p

2. Define edge weights as calculated in Equation 6.

|| − || || − ||
{ 𝐞𝐱𝐩 − − ,( , 𝝐𝜺
, = 𝛔 𝛔 (6)
{ ,

3. Proceed with the final step in the LE algorithm.

28
Gilles – Bowles

This method extended the Shi Malik approach towards penalty on

the differences in the directionality of the spectral information. The edgr
weight is based on Equation 7.

< , > || − ||
−
(− ( )− ) ,( , 𝝐𝜺
, = { || ||.|| || 𝛔 (7)
,

Hou – Zhang –Ye - Zheng

This model proposes a slightly different approach to spectral spatial

fusion. In the above mentioned methods, graph edges were based solely on

spatial information and weights were based on spectral –spatial fused

information. On the other hand, this method uses the fused information in
the step of defining the graph edges in addition to using binary weights.

1. Construct G so that edges E are defined based on the k nearest neighbor

algorithm ie define an edge if it’s in the k nearest neighbors of each other
according to the measure mentioned in Equation 8.

|| − || || − ||
( , = − 𝐞𝐱𝐩 − ). − − ) (8)
𝛔 𝛔

2. Define binary edge weights by calculating based on Equation 9.

, ( , ∈ 𝜺
, ={ (9)
,

29
Benedetto et al

This method proposes a variety of ways to fuse spectral and spatial

information into the LE based algorithm that is used in integration with
linear discriminant analysis (LDA) for classification. The metric used in
Equation 10 is used for the same.

|| − || || − ||
𝜷( , = √𝜷 + −𝜷 (10)
𝛔 𝛔

Where 0≤ ≤ 1. d0 measures the scaled eucliedan distance based purely on

the spatial components and d1 measures the scaled euclidean distance vased
purely on spectral components.

Further 𝛽 is the graph constructed so that the edges 𝜀 𝛽 are defined based
no the k – nearest neighbor algorithm. The weight matrix 𝛽 is defined
using equation 11.

𝜷 (− 𝜷 , , , 𝝐 𝜺𝜷
, = { (11)
,

Laplacian matrix 𝛽 = 𝛽 − 𝛽

Modified Spectral Spatial Schrodinger Eigenmaps

This involves spatial spectral fusion based on Schrodinger Eigenmaps using

cluster potentials and spatial proximity.

30
 Steps involved

1. Construct an undirected graph G= (X,E) with vertices in X and edges E

defined based in spatial proximity.

2. Define weight for the edges between points in X as in equation 12.

−|| − ||
(12)
𝛔

3. Define the cluster potential matrix V representing the spatial proximity

between the vertices as in Equation 3.

|| −
,
= ∑ ∑𝑥 𝜖
𝑝 . , . (− ) (13)
𝜖 𝑥 𝛔
=

𝑝
Where 𝜖 is the set of points in X whose spatial components are in the

𝜖 neighborhood of the spatial components of xi

4. The smallest m+1 eigenvalues an eigenvectors are calculated using the

conventional eigenvector problem Lf = λDf where D is the diagonal
weighted matrix by Di,j = Ʃ jWi,j and L= D – W is the laplacian matrix.
The eigenvectors found f0,…..,fm are serially arranged in order so that

= 𝝀 ≤ 𝝀 ≤ .…… ≤ 𝝀

Then the points y1T , y2T , ..., y3T are defined to be the rows of

𝑭= [ , ,……, ]

31
 Knowledge propogation:

Situation arise when cluster potentials which are extracted from small
set of manually provided labels may have a crucial impact on the
dimensionality reduction method. This can be achieved by replacing the
cluster potential matrix M as follows in Equation 14.

− ,
= ∑ 𝜂 ,
. (14)
𝑥 ,𝑥 𝜖

3.4.4 Recommendation model evaluator

Evaluate the soil class obtained from the image classifier and data and
use the historical data to recommend suitable crops for each crop field.

Input: Feature vector containing features like soil class, terrain,

slope, elevation, etc

Output: a list of recommended crops

1. X ←input feature vector

2. AdaBoost ←one vs rest Booster for
class i
3. return argmaxi(AdaBoost .predict(X))

Algorithm: 6

Technical description
Given the training data , ,………… ,
∈ {−1, +1 } , ∈ is the object or instance, yi is the classification.
For t=1,……,T

32
Form a distribution DT on {1,…..,m}
Select weaker classifier with smallest error ∈ on 𝑡

∈=𝑃 𝑡 [ℎ𝑡 ≠ ]
ℎ𝑡 : → {−1, +1}
Output single classifier Hfinal(x)

33
 CHAPTER 4
IMPLEMENTATION AND RESULTS
This chapter gives a detailed description about the experimental setup, tools
used for crop prediction, the implementation of proposed work, its analysis and
output results.
4.1 Tools
This part gives a detailed introduction and feature analysis of th tools and libraries
used to achieve the results.
4.1.1 MATLAB
MATLAB (matrix laboratory) is multi-paradigm mathematical computing
background and fourth-generation programming language. A patented
programming language established by MathWorks, MATLAB performs matrix
operations, plotting of functions and data, execution of algorithms, formation of
user interfaces, and interfacing with programs written in other languages, including
C, C++, C#, Java, Fortran and Python. Although MATLAB is envisioned chiefly
for numerical computing, an noncompulsory toolbox uses the MuPAD symbolic
engine, allowing admission to symbolic computing capabilities. An added package,
Simulink, adds graphical multi-domain simulation and model-based design for
dynamic and embedded systems

4.1.1.1 Features
Matlab provides the following important characteristic features of usage.

 It offers functions for assimilating MATLAB based algorithms with

external applications and languages such as Java, .NET, C and MS Excel.

34
  It provides implements for constructing applications with custom graphical
interfaces.

 It provides built-in graphics for envisioning data and tools for creating
custom plots.

 It is a high-level language for numerical computation, visualization and

application development.

 MATLAB's programming interface gives development tools for refining

code quality maintainability and maximizing performance.

 It also provides an interactive environment for iterative exploration, design

and problem solving.

 It provides huge library of mathematical functions for linear algebra,

statistics, Fourier analysis, solving ordinary differential equations,
numerical integration, filtering and optimization.

4.1.2 PyCharm
PyCharm is an Integrated Development Environment (IDE) used for
programming in Python. It provides code analysis, a graphical debugger, an
integrated unit tester, integration with version control systems (VCSes), and
supports web development with Django.

4.1.2.1 Features
PyCharm provides the following important characteristic features of usage
 Coding Assistance and Analysis, with code completion, syntax and error
highlighting, linter integration, and quick fixes

35
  Project and Code Navigation: specialized project views, file structure views
and quick jumping between files, classes, methods and usages
 Python Refactoring: including rename, extract method, introduce variable,
introduce constant, pull up, push down and others
 Support for web frameworks: Django, web2py and Flask
 Integrated Python Debugger
 Integrated Unit Testing, with line-by-line coverage
 Google App Engine Python Development
 Version Control Integration: unified user interface for Mercurial, Git,
Subversion, Perforce and CVS with changelists and merge

4.1.3 Matplotlib
Matplotlib is a plotting library for the Python programming language and its
numerical mathematics extension NumPy. Matplotlib is a library for making 2D
plots of arrays in Python. It provides an object-oriented API for embedding plots
into applications using general-purpose GUI toolkits like wxPython, Qt, or GTK+.
The Matplotlib code is conceptually divided into three parts: the pylab interface is
the set of functions provided by Matplotlib.
The Matplotlib frontend or Matplotlib API is the set of classes that do the
heavy lifting, creating and managing figures, text, lines, plots and so on. This is an
abstract interface that knows nothing about output. The backends are device-
dependent drawing devices, aka renderers, that transform the frontend
representation to hardcopy or a display device

36
4.2 Deployment details
The deployment of the system requires Windows 10 (or) 8.1 operating
system. The system must also be installed with Python 2.7 (or) 3.4. Any IDE like
Pycharm can be used to deploy the system successfully.

4.3 Dataset

The following are the datasets used for training and testing the prediction
model built.
The weather data is extracted using Darksky API to get instantaneous data
regarding weather , humidity details etc. Hyperspectral images are obtained from
Hyperspectral Remote sensing scenes.

URL:
www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes

County and crop yield data sets are obtained from fao.org

URL:

www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/harmonized-world-
soil-database-v12/en/

Weather
weather_all.csv Instantaneous weather data

Yield
yield_proc_data.csv Yield forecast text data

37
Pickle Feed
PickleFeed.csv County wise crop data

Location
county_lat_lng.csv Locational data for each county

Images
PaviaU.mat Hyperspectral images
KSC.mat Hyperspectral images
Salinas.mat Hyperspectral images
Salinas_corrected.mat Hyperspectral images
Botswana.mat Hyperspectral images
Indian_pines_g.mat Hyperspectral images

Table 4.1 Dataset

For Training the model, years 2010 – 2014 have been considered and the
subsequent 2 years till 2016 are taken into account for testing. Training images are
taken as Pavia, Salinas, Botswana region and testing: Indian Pines. The
hyperspectral images extracted from the above mentioned URL are for the United
States of America. The crop yield and crop recommendation modules are
implemented for 4 main states in the USA with multiple counties within the state.
Hence our area under consideration is Barley and it’s variances of crop growth.

38
4.4 Preprocessor

The preprocessor module performs variety fo classifiers and segmentation

mehods on the training data. Figure 4.1 shows the user interface for the
preprocessor module.

Figure 4.1 Preprocessor module

39
 ‘’
Figure 4.2 Original Image

Figure 4.3 Mid-Stretch filter

40
 As shown above, multiple filters like erode, midstretch, median, grey
dialate and sharpen filter are applied to the training image for noise removal using
Multiple morphological component analysis in figures 4.2 ad 4.3.

Figure 4.4 Noise removal using grey- dialate filter

4.5 Image Classifier

For the image classifier, the test data consists of images belonging to
different soil class collected manually from the internet. The input to the image
classifier is a soil image. It predicts the soil class for a given image using a spatial
spectral fusion methodology which is fed into the SVM along with latitude and
longitude values The input to the SVM is a tuple containing soil class, latitude and
longitude values using which it recommends a crop

41
 The following figures represent the output of the classification method for
Spatial Spectral Schrodinger Eigenmaps. The performance measures are
simultaneously calculates using the standards like overall accuracy, average
accuracy, average precision, average sensitivity and average specificity.

Figure 4.5 Classification image results

42
 Figure 4.5 displays the image classification results and 4.6 shows the
accuracy and time elapsed for the same process. This involves time to setup the
unconstrained and ML constrained method of classification.

Figure 4.6 Classification Accuracy measures

4.6 Crop yield prediction

The adaBoost classifier model is used to predict the crop yield per
state or country for whichever weather condition that is available at the
darksky server.

43
 Figure 4.7 Training data execution results

Figure 4.8 Crop yield result

The above figures 4.7 and 4.8 show the results of the crop yield prediction
along with the training module root mean square error values.

4.7 Relevance prediction of crops

This model involve the comparison between truth and predicted crop for that
particular latitude and longitude as evident in Figure 4.10

44
 Figure 4.9 Crop soil data set

Figure 4.10 Relevance prediction for red soil

45
 4.8 Performance measures

A binary classifier foresees all data instances of a testing information

as +ve or -ve. This gives four main results – true positive, true negative, false
positive and false negative.
True Positive (TP): correct positive prediction
False Positive (FP): incorrect positive prediction
True Negative (TN): correct negative prediction
False Negative (FN): incorrect negative prediction

𝑃+ 𝑃+
𝐴𝑐𝑐 𝑎𝑐 = =
𝑃+ + + 𝑃 𝑃+

𝑃 𝑃
𝑖𝑖 𝑖 = =
𝑃+ 𝑃

𝑐𝑖 𝑖𝑐𝑖 = =
+ 𝑃

𝑃
𝑃 𝑐𝑖 𝑖 =
𝑃+ 𝑃

1 ^
= √ ∑( −
=

Where n is the total number of tuples and yi are the data points.

46
 4.9 Yield Prediction

Figure 4.9 deals with training model analysis of regression, The geaph is set
upon the relation between each data point and the yield in bushels/acre.

Figure 4.11 Training set regression analysis

After the training part of the model is completed for the years 2010 to 2014,
the testing phase starts for the subsequent years till 2016.Regression analysis of the
testing mechanism results in the following figure 4.10.

Figure 4.12 Testing regression analysis

47
 Figure 4.11 shows the handoff between the existing yield values and the
predicted one and how far they are correlated in context with the data sets..

Figure 4.13 True vs Predicted yield analysis

The performance analysis measures for the comparison between different
classification mechanism are as follows

48
 4.10 Classification

=> SM SSSE GB BM BE HZYZ

=> UN ML UN ML UN ML UN ML UN ML UN ML
Overall 0.9689 0.9786 0.9760 0.9831 0.9799 0.9800 0.9669 0.9768 0.9714 0.9752 0.9742 0.9739
Accuracy:
Average 0.9961 0.9973 0.9975 0.9979 0.9966 0.9975 0.9959 0.9971 0.9964 0.9969 0.9968 0.9967
Accuracy:
Average 0.9112 0.9124 0.9812 0.9815 0.9341 0.9188 0.9051 0.9188 0.9329 0.9185 0.9767 0.9680
Precision:
Average 0.9800 0.9808 0.9851 0.9863 0.9842 0.9831 0.9760 0.9810 0.9818 0.9797 0.9767 0.9751
Sensitivity:
Average 0.9978 0.9985 0.9981 0.9988 0.9986 0.9986 0.9977 0.9984 0.9980 0.9983 0.9982 0.9982
Specificity:

Table 4.2 Classification Accuracies

SM – Shi Malik
GB – Gillis Bowles
SSSE – Spatial Spectral Schrodinger Eigenmaps
BM - Benedetto-M: Fused Metric
BE - Benedetto-E: Fused Eigenvectors
HZYZ - Hou-Zhang-Ye-Zheng
Table 4.3 Analysis methods

49
 0.985

0.98

0.975
Overall Accuracy

0.97

0.965

0.96

0.955
SM GB SSSE BM BE HZYZ
UN 0.9689 0.9797 0.976 0.9669 0.9714 0.9742
ML 0.9786 0.9805 0.9831 0.9768 0.9752 0.9739

Figure 4.14 Overall Accuracy

0.988

0.986

0.984

0.982
Average Specificity

0.98

0.978

0.976

0.974

0.972

0.97

0.968
SM GB SSSE BM BE HZYZ
UN 0.98 0.9842 0.9851 0.976 0.9818 0.9767
ML 0.9808 0.9831 0.9863 0.981 0.9797 0.9751

Figure 4.15 Average Specificity

50
 0.988

0.986

0.984

0.982
Average Precision

0.98

0.978

0.976

0.974

0.972

0.97

0.968
SM GB SSSE BM BE HZYZ
UN 0.98 0.9842 0.9851 0.976 0.9818 0.9767
ML 0.9808 0.9831 0.9863 0.981 0.9797 0.9751

Figure 4.16 Average Precision

0.988

0.986

0.984

0.982
Average Accuracy

0.98

0.978

0.976

0.974

0.972

0.97

0.968
SM GB SSSE BM BE HZYZ
UN 0.98 0.9842 0.9851 0.976 0.9818 0.9767
ML 0.9808 0.9831 0.9863 0.981 0.9797 0.9751

Figure 4.17 Average Accuracy

51
 0.988

0.986
Average Sensitivity 0.984

0.982

0.98

0.978

0.976

0.974

0.972

0.97

0.968
SM GB SSSE BM BE HZYZ
UN 0.98 0.9842 0.9851 0.976 0.9818 0.9767
ML 0.9808 0.9831 0.9863 0.981 0.9797 0.9751

Figure 4.18 Average Sensitivity

The above Figures from 4.11 to 4.15 depict the performance measures of the
classification done by analysis of 6 methods mentioned in table 4.2. Hence the ML
constrained performance is upgraded in every measure in case of Spectral spatial
Schrodinger Eigenmaps.

52
 Chapter 5
Conclusion and future work
Machine learning is a type of artificial intelligence (AI) that provides
computers with the ability to learn without being explicitly programmed . The
proposed machine recommends the suitable crop given an image of the soil and the
parameters like latitude and longitude, with classification of the soil class
intermediate. The system builds up an Image Classifier Model, using SSSE and
SVM, which acts as an image classifier builder. The Image Pre-processor used to
remove the noise from the image (unwanted area). This could then ease the work
of Image Classifier Evaluator, to predict the soil class with improved accuracy.
This model also uses the super pixel (collection of pixel) instead of individual
pixel.
The system can be extended to the mobile application to help the farmers by
uploading the image of agriculture area. The efficiency of the pre-processing is
limited by the amount of unwanted information (like leaves, grass and other stuffs)
present in it. Due to this undesirable information present in the input image, both
during training and classification, the pre-processor fails to identify the exact
contours, thus failing to perform with improved efficiency. The parameter for the
image like climatic factor, moisture and past dataset can be used to predict the
yield of the crop. Collection of more valid details of soil class, latitude, longitude
and suitable crop can greatly accelerate the efficiency of work. The pre-processing
unit could hence be improved and a lot more features can be extended, thus
significantly contributing towards the agricultural welfare worldwide.

53
 Chapter 6
References

[1] Xiang Xu, Jun Li Mauro Dalla Mura- Multiple Morphological Component
Analysis Based Decomposition for Remote Sensing Image Classification-IEEE
Transactions on GeoScience and Remote Sensing. Pages 3083-3102,Jan-2015.

[2] X.E. Pantazi D. Moshou T. Alexandridis b , R.L. Wheton M Mouazen Wheat

yield prediction using machine learning and advanced sensing techniques
Computers and Electronics and Agriculture Pages 57–65,June-2016.

[3] Xia Zhang, Yanli Sun, Kun Shang, Lifu Zhang, Senior Member, IEEE,
and Shudong Wang “IEEE JOURNAL OF SELECTED TOPICS IN APPLIED
EARTH OBSERVATIONS AND REMOTE SENSING” Pages 01,Fevruary-2016.

[4] Wei Yao, Otmar Loffeld - Application and Evaluation of a Hierarchical Patch
Clustering Method for Remote Sensing Images, VOL. 9, NO. 6, JUNE 2016 2279
– 2289.

[5] Michael Johnson, William Hsieha, AlexJ. Cannonb, Andrew Davidsonc,

Frédéric Bédardd -Crop yield forecasting on the Canadian Prairies by remotely
sensed vegetation indices and machine learning methods on Agricultural and
Forest Meteorology 218–219 (2016).

[6] Z. Xue, J. Li, L. Cheng, and P. Du, “Spectral–spatial classification of

hyperspectral data via morphological component analysis-based image
separation,” IEEE Transaction, vol. 53, no. 1, pp. 70–84, Jan. 2015.

54
[7] J. L. Starck, M. Elad, and D. L. Donoho, “Image decomposition via the
combination of sparse representations and a variational approach,” IEEE Trans.
Image Process., vol. 14, no. 10, pp. 1570–1582, Oct. 2015.

[8] X. Jia, B. C. Kuo, and M. Crawford, “Feature mining for hyperspectral image
classification,” Proc. IEEE, vol. 101, no. 3, pp. 676–697, Mar. 2015.

[9] Imaging Spectroscopy: Earth and Planetary Remote Sensing with the USGS
Tetracorder and Expert Systems Roger N. Clark, Gregg A. Swayze, K. Eric Livo,
Raymond F. Kokaly, Steve J. Sutley, Journal of Geophysical Research, 2012.

[10]J. Li, J. M. Bioucas-Dias, and A. Plaza, “Semi-supervised discriminative

random field for hyperspectral image classification,” in Proc. 4th WHISPERS, pp.
4/ 2015.

[11] J. L. Starck, M. Elad, and D. L. Donoho, “Image decomposition via the

combination of sparse representations and a variational approach,” IEEE Trans.
Image Process., vol. 14, no. 10, pp. 1570–1582, Oct. 2015.

[12] M. Fauvel, J. Chanussot, and J. A. Benediktsson, “A spatial–spectral kernel-

based approach for the classification of remote-sensing images,” Pattern Recognit.,
vol. 45, no. 1, pp. 381–392, Jan. 2012.

[13] J. C. Nunes, Y. Bouaoune, E. Delechelle, O. Niang, and P. Bunel, “Image

analysis by bidimensional empirical mode decomposition,” Image Vis. Comput.,
vol. 21, no. 12, pp. 1019–1026, Nov. 2013.

[14] B. Demir and S. Erturk, “Empirical mode decomposition of hyperspectral

images for support vector machine classification,” IEEE Trans. Geosci. Remote
Sens., vol. 48, no. 11, pp. 4071–4084, Nov. 2010.

55
[15] Y. Tang, Y. Lu, and H. Yuan, “Hyperspectral image classification based on
three-dimensional scattering wavelet transform,” IEEE Trans. Geosci. Remote
Sens., vol. 53, no. 5, pp. 2467–2480, May 2015.

[16] J. M. Bioucas-Dias et al., “Hyperspectral remote sensing data analysis and

future challenges,” IEEE Geosci. Remote Sens. Mag., vol. 1, no. 2, pp. 6–36, Jun.
2013.

[17] M. Fauvel, Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton,

“Advances in spectral–spatial classification of hyperspectral images,” Proc. IEEE,
vol. 101, no. 3, pp. 652–675, Mar. 2013.

[18] A. Plaza et al., “Recent advances in techniques for hyperspectral image

processing,” Remote Sens. Environ., vol. 113, no. S1, pp. 110–122, Sep. 2014.

[19] Y. Tarabalka, J. A. Benediktsson, and J. Chanussot, “Spectral–spatial

classificationof hyperspectral imagery based on partitional clustering techniques,”
IEEE Trans. Geosci. Remote Sens., vol. 47, no. 8, pp. 2973–2987, Aug. 2009.

[20] Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and

classification of hyperspectral images using watershed transformation,” Pattern
Recognit., vol. 43, no. 7, pp. 2367–2379, Jul. 2010.

[21] Benedetto, J., Czaja, W., Dobrosotskaya, J, Doster, T, Duke, K, and Gillis,
“Integration of heterogeneous data for classification in hyperspectral satellite
imagery,” vol. 21, no. 12, pp. 1019–1026, Nov. 2

56
[22] Cahill, N. D, Czaja, W, and Messinger, “Schroedinger eigenmaps with
nondiagonal potentials for spatial-spectral clustering of hyperspectral imagery”,
vol. 48, no. 11, pp. 4071–4084, Nov. 2010.

[23] Belkin, M. and Niyogi, “Laplacian eigenmaps for dimensionality reduction

and data representation,” vol. 45, no. 1, pp. 381–392, Jan. 2012.

[24] Czaja, W. and Ehler, “Schroedinger eigenmaps for the analysis of

biomedical data,” IEEE Transactions on Pattern Analysis and Machine Intelligence
vol 35,pp 1274–1280 (May 2013).

[25] Wagstaff, Cardie, Rogers, and Schrodl, “Constrained k-means clustering

with background knowledge,” in Proceedings of the Eighteenth International
Conference on Machine Learning, ICML ’01,pp 577–584 2015.

[26] Benedetto, Czaja, Dobrosotskaya, and Gillis, “Integration of hetero-

geneous data for classification in hyperspectral satellite imagery,” Vol. 8390,
839027–1–839027–12 June 2014.

[27] Bachmann, Ainsworth, and Fusina, “Exploiting manifold geometry in

hyperspectral imagery,” Geoscience and Remote Sensing, IEEE Transactions on
43, 441–454 March 2015.

[28] Fauvel, Chanussot, and Benediktsson, “Kernel principal component analysis

for the classification of hyperspectral remote sensing data of urban areas,”
EURASIP Journal on Advances in Signal Processing ,vol 14 ,pp 1–14 2014.

[29] Prasad, S. and Bruce, “Limitations of principal components analysis for

hyperspectral target recogntion,” Geoscience and Remote Sensing Letters, IEEE 5,
625–629., and Zheng 2013.

57