A Deep Learning Framework for Smart Health Farming

  • Slides: 22
Download presentation
A Deep Learning Framework for Smart -Health Farming Presented by: Biswadeep Khan (IIIT, Delhi)

A Deep Learning Framework for Smart -Health Farming Presented by: Biswadeep Khan (IIIT, Delhi) Chinmayee Mallick (TERI School of Advanced Studies, Delhi) Kusuma Gollapudi (JNTU, Anantapuram) 1

Contents ØNeural Network ØRecurrent Neural Network ØUnrolling an RNN for Sequential Modeling ØLong Short-Term

Contents ØNeural Network ØRecurrent Neural Network ØUnrolling an RNN for Sequential Modeling ØLong Short-Term Memory (LSTM) ØApplication of RNN ØApplication of Deep Learning to design smart crop growth/ crop pathology prediction ØDSSAT model ØExpected outcome of the work ØInputs for DSSAT simulation ØOutputs of DSSAT simulation ØOutput Parameters and Plots of Simulation results ØResults ØChallenges and future Scope 2

Neural Networks • Generally there are two kinds of neural networks: Ø Feed forward

Neural Networks • Generally there are two kinds of neural networks: Ø Feed forward Neural Networks: ü connections between the units do not form a cycle Ø Recurrent Neural Network: ü connections between units form cyclic paths 3

Recurrent Neural Network (RNN) • Recurrent since they receive inputs, update the hidden states

Recurrent Neural Network (RNN) • Recurrent since they receive inputs, update the hidden states depending on the previous computations, and make predictions for every element of a sequence. • RNNs are a neural network with memory. • RNNs are very powerful dynamic system for sequence tasks such as speech recognition or handwritten recognition since they maintain a state vector that implicitly contains information about the history of all the past elements of a sequence. 4

Unrolling an RNN for Sequential Modeling • 5

Unrolling an RNN for Sequential Modeling • 5

LONG SHORT-TERM MEMORY (LSTM) • An LSTM is a special kind of RNN architecture,

LONG SHORT-TERM MEMORY (LSTM) • An LSTM is a special kind of RNN architecture, capable of learning long-term dependencies. • LSTM networks introduce a new structure called a memory cell. • Each memory cell contains four main elements: § Input gate § Forget gate § Output gate § Neuron with a self-recurrent These gates allow the cells to keep and access information over long periods of time. LSTM Memory Cell 6

Applications of RNN ØSpeech Reorganization ØQuestion and Answering ØGoogle Search ØSentiment Analysis ØProtein Function

Applications of RNN ØSpeech Reorganization ØQuestion and Answering ØGoogle Search ØSentiment Analysis ØProtein Function Prediction Our experiment : Sentiment Analysis Problem Statement: Given a movie review in IMDB. Can the model predict whether it is positive or negative? ? Methodology: RNN using LSTM model Input: Movie Reviews of 500 words(text mapped to integer) Input Size: 8300 Output Size: 2700 Accuracy: 88. 38% on test data 7

Application of Deep Learning to design smart heath crop growth prediction /crop pathology DSSAT

Application of Deep Learning to design smart heath crop growth prediction /crop pathology DSSAT Output Images of Crop Pathology Time Series Data for Crop Growth Stages Application of Convolution Neural Network(CNN) Application of RNN Smart Health Crop Growth/ Crop Pathology Prediction 8

DSSAT MODEL • The Decision Support System for Agro technology Transfer is a software

DSSAT MODEL • The Decision Support System for Agro technology Transfer is a software application program that comprises crop simulation model for over 40 crops. • DSSAT is supported by database management programs for soil, weather and crop management data. • It simulates growth, development and yield as a function of the soil-plant-atmosphere dynamics. 9

Expected Outcome of this Project HEALTH SMART FARMING Precise Farm Monitoring Crop growing Conditions,

Expected Outcome of this Project HEALTH SMART FARMING Precise Farm Monitoring Crop growing Conditions, Weather and soil Model-Predictive Plant Health Control input/output wastage, pests and diseases 10

Inputs for DSSAT Simulation Crop Management Data Module Environment Fields Soil Analysis Management Treatment

Inputs for DSSAT Simulation Crop Management Data Module Environment Fields Soil Analysis Management Treatment Simulation Option Cultivators Planting Irrigation Initial Conditions Environmental Modification Fertilizer Harvest Tillage Chemical Application Organic Amendment 11

Inputs for Soil Inputs for Weather 12

Inputs for Soil Inputs for Weather 12

13

13

Output after Simulation 14

Output after Simulation 14

Output Parameters After Simulation 15

Output Parameters After Simulation 15

Plots after Simulation 16

Plots after Simulation 16

Classification of Crop Growth Stages: Labels: 0 -5 Total Number of classes: 6 For

Classification of Crop Growth Stages: Labels: 0 -5 Total Number of classes: 6 For this Experiment: 0 -34(Day) : Label 0 35 -59(Day): Label 1 60 -66(Day) : Label 2 67 -100(Day): Label 3 101 -113(Day): Label 4 114 -160(Day): Label 5 17

18

18

19

19

Results ( RNN for Crop Growth Prediction) ØExperiment 1: Methodology: RNN model using LSTM

Results ( RNN for Crop Growth Prediction) ØExperiment 1: Methodology: RNN model using LSTM Model Description: 3 layers of Sequential LSTM followed by 2 layers of dense network activation = Softmax loss =Categorical Cross Entropy Input Size: 4950 Output Size: 1485 Batch Size: 99 Training Accuracy: 88. 05% Testing Accuracy: 85. 38% ØExperiment 2: Methodology: RNN model using LSTM Model Description: 3 layers of Sequential LSTM followed by 2 layers of dense network activation = Softmax loss =Categorical Cross Entropy Input Size: 9940 Output Size: 2520 Batch Size: 140 Training Accuracy: 93. 66% Testing Accuracy: 90. 07% 20

Results( CNN for Crop Pathology Prediction) ØExperiment 1: Methodology: CNN model for classification of

Results( CNN for Crop Pathology Prediction) ØExperiment 1: Methodology: CNN model for classification of 38 different classes of crop pathology Model Description: 6 layers of convolution layers 3 layers of Maxpool, 3 layers of dense network. Image Source: Literature from Stanford (Plant Village Database) Parameters Used: Learning Rate -0. 005, Image Size: 256 Pixels, Model-Stochastic Gradient Descent, Decay Rate-0. 0005, Momentum-0. 9 Input Size: 10, 840 Output Size: 3, 200 Testing Accuracy: 12. 06% (Paper reported only an accuracy of 2. 86% for random experiment and 98% accuracy on training size of 54, 038 images. ) 21

Challenges and Future Scope Challenges: ØData preprocessing is required. ØExperiment were conducted on synthetic

Challenges and Future Scope Challenges: ØData preprocessing is required. ØExperiment were conducted on synthetic data simulated by DSSAT model crop simulator. But for real life prediction to get the input data from different sources and integrating them is a challenging task. Future Scope of work: ØMulticlassification of crop growth predicting both crop types and stages of growth. ØImprovement of CNN image prediction accuracy. 22