Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 39

DEFORESTATION

JOEL T MASHAWI PREDICTION FROM


B190144A AERIAL IMAGERY
USING MACHINE
Supervisor:
Mr ChAKA
LEARNING
(Convolutional Neural
Network)
Introduction
• Deforestation is one of the most concerning problems that must be handled on an urgent basis by
finding an efficient solution, as it has affected biodiversity, habitat loss, and climate change
exponentially, causing a massive loss to natural elements.
• Regardless of its negative impact, most nations don't have itemized statistics on the level of
deforestation. The Zimbabwe government funder and non-governmental organizations have all
aggressively encouraged the creation of cooperative woodlots in Zimbabwe’s communal regions for
decades.
• Today, the deforestation agent (actors responsible for the process) can be national or international,
since the trade of timber and agricultural products are responsible for a substantial amount of tropical
deforestation emissions.
• There is a strong growing scientific and social movement that tries to analyze and improve the efforts
that are preventing the constant destruction of ecosystem( Alencaret al.,2004,Scholz,2006). These
lesions are microaneurysms (MA), haemorrhages (HM), soft and hard exudates (EX).
• However, since there is a direction connection between deforestation and economic gain(Grieg-
Gran,2006),more than just an ecological approach needs to be taken to plan and prevent deforestation
Background Of The Study
• Regardless of the increasing awareness of deforestation and its consequences,
deforestation and water reserves changes are continuing
• Permanent destruction of forests or water reserves to make the land available for
other uses is increasing rapidly form last several years
• Consequences are related to increased emissions of greenhouse gases, pollution of
water, and loss of biodiversity
• The use of remote sensed images to track deforestation is more is very important
in helping researchers to have a better understanding of where, how and why
deforestation happens and also how to respond to it
• The developments in satellite imagery technology may have a better led in
deforestation detection and can improve in predictions to have a faster, more
convenient and more accurate results
Problem Statement
• Deforestation is a major problem facing our country and the world as it results
rising global temperature, which are being attributed by the abundance of
greenhouse gasses and massive desertification.
• Desertification and greenhouse gases are serious environmental and
socioeconomic problems encompassing huge territories at a global, and regional
and local scale.
• In addition to the mentioned problems, once deforestation has happened the
result causes soil erosion because without forest, the soil erodes and washes
away, causing farmers to move on and perpetuates the cycle.
• The barren which is left behind in the work of these unsustainable agricultural
practices is then more susceptible to flooding, specifically in coastal regions
Objectives
• To design and implement a system which uses CNN to
predict deforestation using remote sensed images
• To analyse different machine learning algorithms used
in image recognition, classification and processing
• To analyse the effectiveness of using CNN in
processing relevant images to predict deforestation
Research Questions
• Q1: How is the author going to analyze different
machine learning algorithms?
• Q2: How is the system going to predict
deforestation?
• Q3: How is the author going to evaluate the
effectiveness of using CNN in the prediction of
deforestation?
Research propositions/ hypothesis
• H0: The system will be able to detect and Predict
deforestation from Satellite Images.
• H1: The system will not detect and Predict
deforestation
Performance Comparison of The Reviewed Classification
Models
Author Title Dataset DL Method Results
Odette Pantoja K., (1999) Forecasting Amazon Brazilian Amazon CNN (AlexNet, VggNet16, custom Accuracy-79%
Rain-Forest Reinforest Degradation CNN) Sensitivity-60% Specificity-
Deforestation Using a 1999-2010 83%
Hybrid Machine Area Under
Learning Model Amazon forest

Prof.Deepti Vijay Neural Network Models STARE (20), DRIVE (40) CNN (ResNet50, InceptionV3,
Chandran.,(2000) for Prediction of and CHASE_DB1 (28) InceptionResNetV2, Xception Accuracy-0.9194
Deforestation and DenseNets) Sensitivity-0.8315
Specificity -0.9858

David A.Coomes., Using Deep Freely available, global Custom Convolutional neural Accuracy-75%
(2006) convolutional neural dataset Network(CNN) Sensitivity-30% Specificity-
networks to forecast 95%
spatial patterns of
Amazonian deforesation.
Methodology-Software & Hardware Requirements

• Windows 10/11 operating system • Core i7+ CPU


• Apache or Tomcat Server • Keyboard
• Jupyter Notebook
• Mouse
• SPYDER (Scientific Python Development
Environment) • Monitor
• Tensorflow
• Keras
• Google Chrome Browser
• Python 3.9
• Anaconda Python IDE
• Streamlit
Prototyping Development Model
• is a software development model in which a prototype is built, tested,
and reworked until an acceptable prototype is achieved. It also
creates base to produce the final system or software.
• 1. Requirement Identification
• 2. Design Stage
• 3. Build the Initial Prototype
• 4. Review of the Prototype
• 5. Iteration and Enhancement of Prototype
System Design-DFD
• Training DFD
Flowchart
• Flowchart
Dataset
• The satellite images where obtained from the region of Mutare.
• All of the images are already saved into their respective folders
according to the categories.
• The dataset have 3 categories deforestation 2000 images, mild
deforestation 1700 images and No deforestation 3500 images
Image/Data Pre-Processing
• During preprocessing, the images were cropped, augmented and
resampled to balance the classes.
Image resizing
Image cropping
Gaussian Blur
• images also can contain different types of noise, especially because of
the source (camera sensor).
• Image Smoothing techniques help in reducing the noise.
Data Augmentation

• refers to the techniques that synthetically expand a dataset by


applying transformations on the existing examples, thus augmenting
the amount of available training data.
• To mitigate the consequences of over / under-representation, data
augmentation is used.
• By adjusting specific parameters, applying random changes to the
original training images.
• These random changes are applied to each epoch, which means that
the model will train on” different” images at each iteration.
Dropout Regularization

• is a technique where randomly selected neurons are ignored during training


• They are “dropped-out” randomly.
• As a neural network learns, neuron weights settle into their context within the
network.
• Weights of neurons are tuned for specific features providing some
specialization.
• Neighboring neurons become to rely on this specialization, which if taken too
far can result in a fragile model too specialized to the training data
• The effect is that the network becomes less sensitive to the specific weights
of neurons.
Training model
from tensorflow.keras.applications import EfficientNetB7
base_model = EfficientNetB7(
include_top=False,
weights='imagenet',
input_shape=(*IMAGE_SHAPE, 3),
classes=4)

# Freeze the base_model


base_model.trainable = False

# append my own layers on the top of the model for Transfer Learning
x = base_model.output

# 1st conv block


x = layers.Conv2D(256, 3, padding='same')(x)
x = layers.BatchNormalization()(x)
x = layers.Activation('relu')(x)
x = layers.GlobalAveragePooling2D(keepdims = True)(x)

# 2nd conv block


x = layers.Conv2D(128, 3, padding='same')(x)
RESULTS
• The matrices utilized to determine the efficiency and efficacy of the
produced solution are accuracy, recall, specificity, sensitivity, weighted
F1 score, prevalence, error rate/misclassification rate.
Output of Feature Extraction
Black Box Testing
Confusion Matrix

• True Positive (TP)


• The actual value was positive and the model predicted a positive value
• The image have deforestation and the model was able to detect first and predicts the
outcome
• True Negative (TN)
• This is the case where the actual value is false and the predicted value is also false.
• In other words, the image have deforestation and our model predicted there is mild
deforestation.
• False Positive (FP)
• It refers to the number of predictions where the classifier incorrectly predicts the
negative class as positive.
• The imge does not contain deforestation and our model says it does.
Metrics Of The Confusion Matrix
• False Negative (FN)
• It refers to the number of predictions where the classifier incorrectly
predicts the positive class as negative.
• The model predicted no deforestation, but the image actually had
deforestation .
Evaluation metrics
Evaluation metric Formula

Accuracy (TP+TN)/(TP+TN+FP+FN)

Precision TP/(TP+FP)

Recall TP/(TP+FN)

Specificity TN/(TN+FP)

Fall-out FP/(FP+TN)

False omission rate FN/(FN+TN)

False discovery rate FP/(FP+TP)

F-score 2 * (Precision * Recall)/(Precision + Recall)

Common evaluation metrics extracted from the confusion


matrix
Confusion Matrix values/Metrics

TP FP FN TN

Mild Deforestation 9 1 1 46

Deforestation 4 4 1 42

No Deforestation 8 1 1 46
Accuracy

• It gives the overall accuracy of the model, meaning the fraction of the
total samples that were correctly classified by the classifier.
• Accuracy calculation as adopted from Karl Pearson (1904)
• Accuracy = (TP+TN)/(TP+TN+FP+FN)
The average accuracy attained was 94.4%
Misclassification Rate/ Error Rate

• Overall, how often is it wrong?


• It tells you what fraction of predictions were incorrect. It is also
known as Classification Error.
• This formula is adopted from Kuha (2005)
• Error rate = (FP+FN)/(TP+TN+FP+FN) or 1-Accuracy()
Misclassification Rate/ Error Rate(cont)
Sensitivity/Recall/True Positive Rate

• When it's actually yes, how often does it predict yes?


• It tells what fraction of all positive samples were correctly predicted
as positive by the classifier. It is also known as True Positive Rate
(TPR), Sensitivity, Probability of Detection.
• Adopted from Powers(2011)
• Recall = TP/(TP+FN)
Precision

• When it predicts yes, how often is it correct?


• Adopted from Selvik (2007)
• Precision = TP/(TP+FP)
Specificity/True Negative Rate

• When it's actually no, how often does it predict no?


• It tells what fraction of all negative samples are correctly predicted as
negative by the classifier. It is also known as True Negative Rate (TNR).
• Adopted from Parikh(2008)
• Specificity = TN/(TN+FP) or 1-FP rate
• Equivalent to 1 minus False Positive Rate
F1-Score/F1 Measure

• It combines precision and recall into a single measure.


• F1-score=2 x (Precision x Recall/ Precision + Recall)
• =2TP/(2TP+FP+FN)
The average specificity was at 94.4% meaning that 94.4% of all negative samples are correctly
predicted as negative by the model.
• The model’s error rate was calculated and it reached
0.036% for the Mild Deforestation class, 0.098%, and
0.053% for Deforestation ,and No Deforestation
classes respectively.
• This brings the overall average error rate at 0.06%.
This means that from the test samples, an average of
0.06% of predictions were incorrect.
Aims & Objectives Realization
• To achieve the goals of the study, the researcher created a model that
makes use of EfficientNet CNN to identify and categorize the various stages
of deforestation as well as predict if it will occur in the near future.
• EfficientNet, a deep CNN architecture was introduced, subjects were
classified into there categories which were No Deforestation, Mild
Deforestation , and Deforestation
• The researcher analysed vast literature relating to the subject matter, did
comparisons and contrasting of literature, discovered limitations and gaps
and addressed these.
• The model solution was tested & evaluated using Blackbox, Whitebox
testing & the confusion matrix.
• A average accuracy of 94.4%, specificity and error rate of 80.4%,95%
and 0.06% respectively were achieved.
• Therefore, providing an improvement over the results obtained by
other deep CNN architectures (Pratt et al., 2016), (Dekhil et al., 2019)
by almost 14% as well as SVM Based classifiers (Priya and Aruna,
2012) with an accuracy improvement of 7% and 3% respectively.
• The gained results are significant and confirm the efficiency of the
EfficientNet CNN neural network in detecting and classifying DR. This,
therefore shows that the objectives mentioned in chapter 1 were
achieved.
• Automated screening technologies drastically shorten the time it
takes to determine a diagnosis, saving ophthalmologists time and
money while also allowing patients to receive treatment sooner.

You might also like