# Summer Training Program 2026. Remote Sensing & GIS with Python Day 12

https://www.youtube.com/watch?v=MyDMItUKB0I

[00:02] Sir am I audible?
[00:03] Yes.
[00:05] Yes sir.
[00:06] So very good evening to everyone.
[00:08] On behalf of the India Space Academy, I warmly welcome all the participants to today's session of the summer training program 2026 skill training and internship under remote sensing and GIS using Python.
[00:20] We are truly honored to have Dr. Deepik Deepak Kumar S with us today.
[00:23] Sir is a distinguished researcher and academician specializing in artificial intelligence, machine learning, deep learning and intelligent energy systems.
[00:35] His research focuses on developing AI-driven solutions for predictive analytics, intelligent monitoring systems, battery health diagnostics and sustainable energy applications.
[00:47] Dr. Dr. Kumar has published numerous research articles in reputed national and international journals and conferences and is also the inventor of multiple patented innovations.
[01:01] His remarkable contributions to research
[01:03] and innovation earned him recognition as young scientist by Anusand.
[01:07] National Research Foundation.
[01:10] ANRF government of India along with prestigious research excellence award from Delhi Technological University.
[01:21] He has also served as visiting research fellow at Mingchi University of Technology where he worked on interdisciplinary projects involving artificial intelligence, advanced materials, energy storage systems and emerging technologies.
[01:38] His current research interests includes green AI, explainable AI, digital twins, edge intelligence and AI enable smart industrial applications.
[01:50] We are privileged to have the opportunity to learn from such an accomplished expert and we are confident that today's session will be highly insightful and inspiring for all the participants.
[02:01] With respect, now I hand over the session to
[02:04] Sir, now.
[02:09] Good evening all.
[02:13] So today, so you.
[02:18] Yes, I'm audible.
[02:22] Hello.
[02:25] Yes, sir. Yes.
[02:30] Okay.
[02:39] Can I start?
[02:42] Yes. So you miss.
[02:51] So I am Deepak Kumar. I'm a researcher in admin in the field of artificial intelligence and uh machine learning. So
[03:07] Today we start machine learning.
[03:11] Or AI role in the geospatial or GIS data.
[03:19] So every day large amount of data sets for geospatial data are generated by using different devices such as satellite drones, GPS devices, GIS data sets generated by uh different laboratories and integrated sensors which are um integrated to uh record and generate or capture the GIS data.
[03:58] So the traditional model of the uh geospatial data sets they are highly uh inefficient and required lot.
[04:10] of energy to process the data set and uh
[04:14] lot of uh computational burden uh
[04:17] required to handle the these type of data sets.
[04:21] So uh here we used uh new emerging area which is uh known as AI artificial intelligence.
[04:33] Uh and uh this artificial intelligence uh we learn the role of uh artificial intelligence in the geo.
[04:44] through the artificial intelligence and machine learning is a powerful devices with tools which are used to extract feature information from the geospatial data sets.
[05:05] So these
[05:10] Uh EI models are efficient in handling the large amount of data set in classification.
[05:19] Uh in uh computationally efficient as compared to traditional methods.
[05:24] So we uh in current scenario use different type of AI, ML which is called machine learning and DL.
[05:37] DL means deep learning models to analyze these data sets for uh getting uh as a outcome insights, prediction, classifications, decision making procedures.
[05:52] So, so uh different kind of work uh done by the uh AI models using the GIS data sets.
[06:07] So what is GOAI?
[06:09] So geo AI is a geospatial
[06:15] GIS remote sensing data sets and artificial intelligence.
[06:18] which integration of geospatial data set.
[06:30] So it enables intelligent analysis.
[06:35] prediction, classifications and decision making process using special data sets.
[06:44] So, uh the AI uh also use different type of application in GIS such as land use and land cover classifications, precision in agriculture or agriculture monitoring of lands, fluid monitorings, deportation detection.
[07:16] urban growth analysis and disaster measure management and climate and environmental monitoring by using the different type of AI models.
[07:29] So what is AI?
[07:33] So AI is a uh computer system that is capable of performing tasks different type of task that typically require human intelligence such as learning, reasoning, decision making, pattern recognition, problem solving.
[07:57] So these different kind of uh uh work or task done by AI uh uh as a human intelligence.
[08:06] So in uh different area AI also uh transform many sectors such as healthc care
[08:17] Sorry, transportations, finance, education and also it implemented in the geospatial science.
[08:29] So what is machine learning?
[08:32] Machine learning is a subset of AI.
[08:36] So it is a branch of AI that enables computers to learn patterns from data set and improve their performance without being explicit program.
[08:50] So it also like uh uh instead of writing rules manually we provide something like input data to the to the machine learning and uh it gives in output like different type of insights such as when we gives like input satellite images, GPS data, weather data and GIS layers.
[09:18] So we uh get different type of outputs.
[09:22] like classifications of land cover maps,
[09:26] fluid predictions, crop classifications and urban growths.
[09:29] So different type of work uh done by the machine learning uh in efficient manner uh uh uh and low energy consumption also.
[09:46] So why machine learning used in the uh current scenario?
[09:53] So the traditional geospatial analysis struggle with massive satellite data set because these type of data set has large amount of um large amount of uh resolution, high volume, high uh it's uh data size and
[10:21] Different, different unsupervised data.
[10:24] Data sets.
[10:27] So the traditional models fails or struggle to learn these type of data set.
[10:31] Uh, data sets, also multi-uh, uh, multi-layer data sets.
[10:44] Sources geospatial.
[10:47] Multi-layer data sets sources geospatial informations.
[10:47] Also, uh, complex spatial patterns, like they struggle to recognize the pattern of the uh, uh, captured image or time series data.
[10:57] Uh, data sets H and it's unable to give the information in real time.
[11:05] Like it take uh too many uh time or lot of uh time to process the data or uh, uh, give the output or give the information in.
[11:23] The real time.
[11:27] So as compared to traditional me methods,
[11:29] ML has automated feature extraction,
[11:33] large handling uh efficiently handling large scale data sets,
[11:40] pattern recognition,
[11:41] efficiently high accuracy in prediction,
[11:45] intelligent decision support.
[11:48] So goi review highlight ML is a key technology for managing and exploiting massive geospatial data sets and performing soft uh complex analysis task.
[12:05] So traditional program v machine uh learning program.
[12:10] So in traditional programs we used uh like input rules and get output example like coordination, rooting, navigation then we get the output.
[12:23] But in machine learning we
[12:26] required large of data sets to predict the uh output or predict the uh any type of information.
[12:47] like the procedure input uh data which required the desired output learning algorithm development of ML models or uh the accordingly uh pre-process the data and model development and then we get predictions.
[13:06] So here you can uh see we uh uh uh we u uh input uh we give the model to the as a input like satellite data.
[13:15] So it also um uh uh analyze the uh data uh such as
[13:27] in human interpretation like time consuming h subjective difficult for large data regions or recognition.
[13:37] But in machine learning learn from samples or pixels automatically classified the millions of pixels produces accurate and thermal maps.
[13:51] So these uh uh uh uh type of work efficiently handled by the machine learning h.
[13:58] So here the application like already we discussed in the previous slides.
[14:04] So land use and land cover mapping, crop monitoring, float prediction, forest deforestation, weather forecasting.
[14:14] So here so so in next slide uh
[14:38] okay so AI machine learning and deep learning
[14:42] So AI so machine learning is a subset of uh ML and DL in uh deep learning is a subset of ML.
[14:52] So like AI here you can AI it's ML and it's a subset of DL.
[15:01] So like you can understand here artificial intelligence.
[15:11] So it's a broad area or uh focused on the creating intelligence system that can mimic mimic like human intelligence or decision making and problem solving.
[15:27] But yeah, machine learning is a subset of AI where system learn from the patterns or data for improving the performance
[15:39] Automatically.
[15:43] But in the uh uh deep learning is a subset of ML that uses artificial neural networks with multiple layers to learn complex data or complex pattern directly from the large data set.
[15:57] So deep learning is a very powerful uh uh tool to or uh models that are used to handle handle uh large of data set which has high volume and many different type of supervised and unsupervised data sets.
[16:23] So here uh the uh simple comparison uh between the AI uh ML and deep learning models.
[16:31] So what is goal like AI AI goal is simulate intelligence uh machine learning learn.
[16:40] From the data and deep learning uh learn complex patterns.
[16:44] So data requirement low and moderate data for machine learning moderate data sets.
[16:50] But for deep learning it require large data set to train the models.
[16:59] Feature engineering like it uh uh use the manual feature engineering or feature extraction to feed into the AI models.
[17:11] Sometimes in ML used manual or also automatic but in deep learning always use automatic uh feature engineering to feed as a input for DL models in computational efficiency.
[17:30] So it low it moderate but it high because it uh has handle or capability to handle large amount of
[17:42] Data set in efficient manner.
[17:46] So like you can uh see the some model example like export system random forest.
[17:52] CNN and transformer models.
[17:57] So uh here the uh machine learning models.
[18:04] Why deep learning is important for geospatial analysis.
[18:07] So deep learning can automatically learn features from the different type of data sets such as satellite images.
[18:17] UAB images.
[18:20] LDR, SR and video streaming.
[18:22] So these uh type of uh DL models used in different type of applications like road extraction, building detection, crop classification, fluid mapping and forest monitoring.
[18:37] So you can uh uh take a
[18:45] Example such as like the ML model.
[18:49] So ML model is a manual feature engineering extraction like use classification for building mapping.
[18:54] But in the uh DL also you use image but large amount of uh uh data set as a image.
[19:08] So here directly feed uh images into the model and map the building where the uh industrial where uh the building are uh residential or uh where the uh uh commercial buildings.
[19:30] So already uh we uh discussed in the previous slides like latic uh like what is goi?
[19:35] So goi is a uh is integration of multiple uh data set from the uh go uh gis yeah geospatial plus remote sensing.
[19:49] And AI.
[19:54] So why uh goi important?
[19:57] So it's generating numerous amount of geospatial data set every day from different type of sources.
[20:04] So like earth, UAB, GPS, mobile, IoT, GIS data sets.
[20:08] So why we need so here time uh traditional models are time consuming h limited scalability or uh less implemented in realtime applications or difficult to process uh large amount of data sets.
[20:25] So GOI based analysis are uh automated processing, high accuracy, faster decision making, realtime monitoring and large scale analysis or easy to implement in the hardware in real time.
[20:42] So what is the component of GOI?
[20:45] So GOI is a uh first one is a geospatial data.
[20:51] in which such a satellite image
[20:55] use land use maps, GPS and weather data
[20:59] and GIS. So it is uh geographic
[21:03] information system. It it is a tool that
[21:07] used to store the and manage and analyze
[21:10] and visualize geospatial data sets such
[21:14] as the software QGIS and ArcGIS used to
[21:19] generate or uh manage or identified or
[21:23] visualize the uh different type of uh
[21:27] GIS data set by using QGIS and RKGIS
[21:32] software. here remote sensing. So remote
[21:35] sensing
[21:38] is a uh acquiring information about the
[21:42] earth surface without physical contact
[21:45] by using lot of sensors. So here example
[21:49] like uh land data set
[21:53] 2 data set and modis
[21:56] and fourth one is a artificial
[21:59] intelligence. So here we use different
[22:02] type of AI models or ML models or DL
[22:07] models in GOAI like random forest, SBM,
[22:11] uh neural networks, you can also ML uh
[22:15] also DL like CNN. So different type of
[22:18] models used in AI.
[22:22] So
[22:24] workflow of the uh GOI how we um
[22:30] implement or how to we make the GOI. So
[22:35] first we uh collect the data sets.
[22:44] Collect the data set. Here you can see
[22:48] these are data sets. So we collect the
[22:51] data set from different sources and
[22:55] then we pro pre-process the data set. H
[23:00] in next slide we uh discuss about the uh
[23:03] different type of ML uh uh
[23:06] classifications or uh data set how we
[23:10] supervised data set unsupervised data
[23:12] set H so we pro pre-process the data set
[23:18] feed into the machine learning models or
[23:20] you can also directly feed into the
[23:22] machine learning models and then we
[23:25] predict and map and decision make based
[23:28] on the prediction.
[23:30] So here the realtime application of GOI
[23:34] like we use in agriculture like crop
[23:37] classification, ill prediction,
[23:39] environmental monitoring also
[23:43] uh used in the uh
[23:51] uh used in the dis uh disaster
[23:54] management like flu uh mapping or
[23:58] landslides prediction. s also used in
[24:01] urban planning water resources. So here
[24:05] you can uh seen a uh simple example uh
[24:09] for food mapping by using the GOI. So
[24:14] directly you can capture the uh images
[24:18] satellite images by using the
[24:20] satellites. uh so different uh uh
[24:25] regions uh uh uh satellite images by
[24:30] using the satellites and directly
[24:34] accept the feature of these
[24:38] uh satellite images and feed into
[24:43] model ML model as a input and predict.
[24:50] So what predict? So where you can
[24:53] predict see in which region fluid is
[24:57] high or not non-fluded area uh and
[25:01] fluided area. So you can uh take a
[25:03] decision uh based on this predictions
[25:08] like fluid risk mapping. This allow uh
[25:11] government and disaster agencies to
[25:14] respond quickly and efficiently to
[25:17] inform people or their
[25:21] units. So GOI GOI is where so geography
[25:26] meets with artificial intelligence also
[25:31] transform raw
[25:33] geospatial data set into external
[25:38] knowledge or insights. It is one of the
[25:43] f fastest growing fields in the GIS
[25:47] remote sensing and geo spatial data
[25:50] science.
[25:53] So here machine learning workflow for go
[25:58] AI applications. So already we discuss
[26:02] in the previous slides but here we
[26:05] discuss in the uh uh broad manner. So
[26:10] how the machine learning work like end
[26:13] to end uh geospatial
[26:16] uh ML workflow. So first uh one is a
[26:21] data acquisition or data collection.
[26:24] Then in second steps we pre-process the
[26:27] data sets. In third step feature
[26:31] engineering or feature extraction of uh
[26:34] pre-process data sets. In
[26:37] these
[26:39] feature extraction uh data sets or
[26:44] information
[26:46] divided into
[26:48] two categories. First one is training or
[26:53] second one is
[26:56] testing data sets. So here we use uh
[27:00] feature engineered uh data sets or
[27:03] information as a input by splitting into
[27:06] two categories like first one is a
[27:09] training
[27:11] training and second one is a testing
[27:13] data sets. So we use 70 sorry uh
[27:16] standard is a uh 80% of data sets for
[27:20] training purpose and 20% data set of for
[27:24] testing purpose.
[27:35] So so uh and next we uh feed the
[27:38] training data sets or uh and testing
[27:41] data sets for model training. So model
[27:45] training is a process to learn the uh
[27:48] model from the uh
[27:52] input data sets. So and uh after
[27:56] learning it the six stage is a
[27:59] prediction or classification based on
[28:01] the uh learning efficiency of the model
[28:08] by using the data sets. And seventh step
[28:13] is accuracy assessment and eighth one is
[28:16] a mapping generation. So here data
[28:20] acquisition
[28:23] steps we u collect the data from
[28:28] different uh
[28:31] sources like satellite GPS uh uh drones
[28:36] h GPS weather GIS data set and LDR in
[28:41] data prep-processing so in data
[28:44] prep-processing basically we remove the
[28:47] uh noise, missing values, cloud in
[28:51] satellite images, geometric distortion
[28:54] and out uh liners and other um like uh
[28:59] in the data set uh NA and uh zero and
[29:04] many different type of uh uh noise or uh
[29:10] content present in the tested data set.
[29:14] So in pre-process
[29:16] u pre-processing uh section or steps we
[29:20] remembering these type of
[29:23] uh raw data contained like noising
[29:26] missing value clouds in satellites
[29:28] images geometric so like uh some uh task
[29:35] performed uh
[29:38] in uh during the pre-processing like
[29:40] data cleaning cloud removal image age
[29:43] corrections, data normalization.
[29:47] So feature extraction feature extraction
[29:51] is a measurable information that
[29:54] [clears throat] help the model
[29:56] distinguish between classes such as for
[30:01] example you can see here data types and
[30:04] features. So in
[30:08] satellite
[30:09] images. So here present different type
[30:12] of features such as red color like
[30:17] represent the uh different type of uh
[30:20] mappings land mappings or uh red,
[30:25] green, blue and nir in them data set
[30:30] like elevation, slopes, vegetation like
[30:35] NDVI and in weather uh type of data sets
[30:39] such as
[30:44] rainfall and temperatures.
[30:52] Okay. So in next steps we train the u
[30:57] data preparation like training data
[31:00] preparation. So create sample like for
[31:03] example so we use the pixel like P1 P2
[31:07] and P3. So feature uh engineering like
[31:12] RGB plus NDBI
[31:16] uh which is uh indicate for poorest and
[31:19] RGB and DBI which P2 represent the water
[31:24] and P3 like RGB and DBI represent the
[31:28] urban. So such as we u data uh prepared
[31:33] for training. So
[31:37] here one of the most thing to uh keep in
[31:41] mind like
[31:44] the quality of training data directly
[31:47] affect the model performance.
[31:55] well qualified
[31:58] model performance you can get the model
[32:03] prediction up to uh 99% or near to 100%.
[32:09] So in uh model training uh in fifth
[32:13] section we uh we
[32:17] uh feed the training data sets to the
[32:23] model as a input. So here so we uh
[32:29] make or we develop different type of
[32:32] algorithm to uh learn the
[32:37] data
[32:39] from training data sets. So here like
[32:43] algorithm learns relationship between
[32:45] feature and classes levels. Like here we
[32:50] class the different type of pixels or
[32:53] you can say image
[33:00] like P1, P1, P2, P3 uh which is
[33:06] image like P1, P1, P2, P3 uh which is
[33:06] like P3 is
[33:09] uh uh uh represents uh urban classes, P2
[33:14] represent water and P1 represent forest.
[33:18] So here these things learn by
[33:26] develop models. H so it ability to
[33:31] classify
[33:33] the uh different pixels or different
[33:36] data sets into different classes
[33:40] like here.
[33:42] So decision tree random forest SB neural
[33:46] network models like the red training
[33:48] data set learning and training models.
[33:51] So in prediction or classification the
[33:54] trend model is applied to new data sets
[33:58] here satellite images model and
[34:01] classified maps. So your model able to
[34:07] uh classify or learn or predict like
[34:13] for classes like P for classes if it
[34:22] are related to the P1 data set. So it
[34:27] gives out
[34:30] what
[34:34] like so accuracy assessment. So in this
[34:40] how is the model like your model of your
[34:46] model?
[34:49] Excuse
[34:55] me sir.
[35:03] So am I
[35:08] Your voice is very low. Students are not
[35:09] able to hear you.
[35:12] Okay. One.
[35:17] Are you
[35:35] audible now?
[35:39] Yes, sir. But your voice is very
[35:43] I'm trying to
[36:16] So uh in the accuracy assessment we uh
[36:24] test the model how it
[36:28] predict like it overall accuracy in
[36:31] precision, recall, FN score and by using
[36:36] the confusion matrix like how it predict
[36:40] where vious ground. Okay. So generate or
[36:45] uh make the decision supports for
[36:47] mapping like landsc
[36:50] uh flu trees mapping crop uh health
[36:54] mapping or urban growth mapping. So
[36:59] uh
[37:03] so different type of ML learning models.
[37:07] So here
[37:10] it classified into
[37:14] in four learning models like first one
[37:18] is a supervised learning, unsupervised
[37:20] learning, semi-supervised learning and
[37:23] reinforcement
[37:25] learning. So their Tamil system are
[37:29] classified into
[37:31] supervised, unsupervised,
[37:33] semi-supervised and reinforce learning.
[37:38] So
[37:40] how do machine learn like by uh
[37:45] supervised learning to learn from label
[37:48] data set? Here
[37:53] very important things in supervised
[37:56] learning we give the model
[38:01] as a input sorry uh we give the data set
[38:05] uh to model as a label data set.
[38:09] So like here you can see uh like input
[38:13] data set from satellites like already
[38:18] classified data into different category
[38:21] to uh make the data supervised like
[38:26] forest, water, urban. So already
[38:29] classified data before feed into
[38:34] model. So in you can say like
[38:38] different type applications
[38:41] used uh RF decision SPM and networks. So
[38:46] in unsupervised machine learning
[38:50] learn from the unlimited data sets right
[38:54] it has make
[38:56] like different classes and random
[38:59] classes and
[39:02] random
[39:04] like
[39:06] images and different type of data
[39:14] used for supervised learning means
[39:21] lamb and meth.
[39:37] Excuse me sir, sorry to interrupt.
[39:46] So am I audible?
[39:51] So your voice is very low.
[40:13] Sir,
[40:14] am I audible?
[40:19] Uh, sir, your voice is very low. Can you
[40:22] please modify it?
[40:25] Yes. Uh, I'm going to connect with my uh
[40:28] PC.
[40:30] So, can you give me one uh second?
[40:37] Hello.
[40:40] Hello.
[40:44] Is voice is okay?
[40:46] Uh sir, it is uh still not coming. It is
[40:49] very low.
[41:00] Okay. So uh in the unsupervised uh
[41:05] learning
[41:08] uh so it the uh uh learn the model from
[41:11] the unlabelled data set. So we uh give
[41:13] the uh information to uh uh data uh such
[41:19] as un super wise.
[41:30] So
[41:31] uh this uh type of data sets like uh uh
[41:37] there are no predefined output label or
[41:40] targeted value during the training. So
[41:45] here are the different type of algorithm
[41:46] which are used to uh
[41:50] uh learn the uh uh common algorithm like
[41:54] k means uh db s cn and hac clustering.
[42:00] So in this semi-supervised
[42:03] uh learning uh uh different type of uh
[42:09] uh
[42:13] some data are like labelled data sets
[42:17] and some data are like unlabelled data
[42:20] sets. So few labelled data samples and
[42:23] many unlabelled data samples to make the
[42:28] data set semi-supervised learning. So
[42:31] better learning why it is useful because
[42:35] reduce the labeling cost. So in the uh
[42:39] uh uh uh supervised uh uh learning so
[42:44] there are many many time or lot of hours
[42:48] required to label the data set before
[42:51] feeding as a input to the model.
[42:55] So
[42:57] uh
[42:59] [clears throat]
[43:00] like uh here uh it is used for large
[43:04] satellite data sets help when the labels
[43:07] are scarce like you can you cannot label
[43:10] the data sets. So many examples like
[43:14] satellite data sets land cover labeling
[43:17] data and building footprints generation.
[43:20] So fourth one is a reinforcement
[43:25] learning. So learn through the
[43:26] interaction with an environment like so
[43:31] it first this type of machine learning
[43:34] models create an environment which is
[43:38] learned with different environments like
[43:41] here
[43:43] uh you can see. So uh first we create
[43:47] the environment and second we reward our
[43:52] penalty and then we learn the policy and
[43:57] take the action. So in the uh uh um uh
[44:04] re
[44:06] enforce re enforcement learning
[44:12] different examples such as like the uh
[44:16] uh autonomous drones, robots, smart
[44:20] traffic and root optimization. So in
[44:24] which like we feed an agent like uh
[44:28] these machine learning uh uh integrate
[44:32] with an agents learns to make the
[44:35] decision by inter interacting with the
[44:38] environment or some receiving rewards or
[44:43] penalties based on its action. So
[44:48] this type of learning
[44:52] it maximize the commulative rewards over
[44:56] the time such as like uh agent take
[45:00] action environment provide the feedback
[45:04] and reward penalty like learning such as
[45:07] wrong and uh uh true and improve
[45:11] decision making process. So like uh it
[45:15] has key uh component of the uh um
[45:21] uh reinforcement learning like agent,
[45:25] environment, different states, action,
[45:28] rewards, policy and these type of steps
[45:33] used to uh make the reinforcement
[45:37] learning. So like in the uh uh uh
[45:43] process like how the agent observe the
[45:46] current state like and agent select the
[45:50] action and environment respond to the
[45:53] action like SNA
[45:56] and agent receive the reward or
[45:58] penalities from the environment and then
[46:00] agent update its policy and then process
[46:04] are repeated until optimal behavior.
[46:07] here of the machine is learned.
[46:13] So different type of uh reinforcement
[46:17] learning also uh uh uh present like
[46:21] positive reinforcement learning,
[46:23] negative uh reinforcement learning. So
[46:26] comparison learning uh types like
[46:30] learning types, supervised learning. So
[46:33] it required label data sets also goal
[46:37] defined like prediction or
[46:39] classifications
[46:41] unsupervised uh learning like no label
[46:44] data set pattern discovery
[46:46] semi-supervised learning like partial
[46:49] and fewle data set data set and many
[46:53] unlevel data sets. So it improved the
[46:56] learning with fewle data sets h
[47:01] reinforcement learning like non-lel data
[47:04] sets and learn through the xable or by
[47:08] agent.
[47:09] So here in the uh geospatial examples
[47:13] where these type of learning uh
[47:16] algorithm used
[47:19] like for the land cover classification
[47:22] many times supervised learning used
[47:27] crop type mapping supervised learning
[47:30] used spatial clustering unsupervised
[47:33] learning
[47:35] change detection unsupervised learning
[47:39] and building extraction and fewer levels
[47:42] for semi supervised learning and for the
[47:46] drone path learning reinforcement
[47:49] learning are used. So which type will we
[47:54] focus on today? So here we only focus on
[47:58] two type of
[48:01] learning methods
[48:03] unsupervised sorry supervised
[48:05] classification and unsupervised
[48:08] classifications.
[48:12] So supervised classifications so what is
[48:14] the supervised classification? So in
[48:19] supervised classification is a machine
[48:21] learning approach where the model learn
[48:24] from the label uh training data and
[48:27] predict the class of unknown samples
[48:32] like training data set feature
[48:33] extraction and labeling of the data set.
[48:36] These data feed into the uh machine
[48:38] learning models and we get the
[48:40] predictions. So here the real time
[48:43] examples is already we discussed like
[48:46] dense, vegetation, river, buildings,
[48:49] forms
[48:51] for
[48:53] classified the different type of
[48:55] classes. So
[48:58] why it is uh called supervised? Like the
[49:03] data
[49:05] also
[49:07] learn in the learn in a pattern like
[49:11] teacher provide the correct answer as a
[49:13] label and the student learn pattern from
[49:16] the examples and exam predict the answer
[49:20] for the unseen questions.
[49:23] supervised classification workflow like
[49:26] images, training samples, feature
[49:28] extraction, model training,
[49:30] classification, accuracy assessment and
[49:33] final mapping. So here the example of uh
[49:38] un sorry supervised learning. So uh the
[49:41] input features like we give the input
[49:44] label data sets uh like red one is uh
[49:48] band green band uh
[49:52] sorry red band green band blue band n
[49:54] band and
[49:56] NDBI band. So here green for the uh
[50:00] forest, blue for the uh water, black for
[50:04] the urban and yellow for for the
[50:10] agriculture. So output you can see in
[50:13] here they these model indicate where the
[50:17] forest like green area like where is the
[50:21] water like here and where is the urban
[50:25] urban is represent the black like here
[50:28] here and agriculture area like
[50:33] so popular uh supervised algorithm in
[50:36] goi like decision tree,
[50:41] random forest and SBM models. So these
[50:45] three models uh commonly used in
[50:50] supervised
[50:54] classifications.
[50:57] So application of supervised
[50:59] classifications like it used in
[51:01] agriculture like crop type mapping, wind
[51:04] prediction in environment like forest
[51:07] monitoring, deforestation detection like
[51:10] this disaster management, urban
[51:12] planning. So what are the advantages of
[51:15] the um supervised um classification like
[51:19] it has high accuracy know the
[51:24] output classes
[51:26] easy
[51:28] of interpretation reliable performance
[51:31] and widely used in GIS and remote
[51:34] sensing. So what are the limitation of
[51:36] these type of uh supervised learning
[51:39] like required label data set time
[51:42] consuming in data collection and
[51:44] labeling
[51:46] also accuracy depends on the training
[51:49] and quality of the data sets also class
[51:52] imbalance can affect the performance.
[51:58] So unsupervised classification here see
[52:03] unsupervised classification is a machine
[52:06] learning approach where the algorithm
[52:08] learn from the unlabelled data and
[52:11] automatically group similar observation
[52:14] into clusters.
[52:16] It's like
[52:18] non- label data set pattern discovery
[52:21] and clustering like blue one side yellow
[52:25] one side red one side light yellow one
[52:28] side and blue one side. So unlike
[52:30] supervised learning the algorithm does
[52:32] not know the correct answer
[52:36] of the
[52:38] foreign.
[52:39] So
[52:41] why we use unsupervised classification
[52:45] here? Sometimes we do not have ground
[52:49] two data, training samples and class
[52:52] label. So such case unsupervised
[52:56] learning very useful.
[52:59] So like such examples, newly acquired
[53:02] satellite images from a remote sensing
[53:05] region where no field survey or data
[53:09] exist like unseen data sets. How does it
[53:13] work? So the algorithm uh groups pixel
[53:16] based on the similarity such as
[53:18] clustering, clustering two, cluster
[53:20] three, cluster four. So these later uh
[53:26] analyst interprets the cluster into
[53:28] interpretation
[53:30] like cluster one represent the water,
[53:33] cluster two represent the forest,
[53:35] cluster three represent the urban and
[53:37] cluster four represent the agriculture.
[53:41] So clustering concept what is the goal
[53:43] of clustering uh concept? So similar
[53:46] picture like same cluster and different
[53:51] pixels like different colors.
[53:54] So popular unsupervised algorithms.
[54:04] So K means uh clustering algorithm
[54:07] mostly most widely used in the
[54:09] clustering technique like working
[54:11] principle. Select the number of clusters
[54:14] K. Here they assign the uh pixel to
[54:17] nearest cluster and update the cluster
[54:20] centers and repeat it until stable. It's
[54:26] so here the different type of data or
[54:29] clusters. So here the uh K means uh
[54:34] algorithm and it classified the
[54:37] different classes and density based
[54:41] clustering like group data based on the
[54:44] its density. So what are its advantages
[54:48] like detect the irregular sets handling
[54:52] noisy data set suitable for the large
[54:55] geospacial data sets. So applications
[54:58] like land cover uh exploration,
[55:03] chain detection, image segmentation and
[55:06] horse analysis like crime
[55:09] diseases, traffic, horse and
[55:12] additifications.
[55:15] So like you can use in the unsupervised
[55:18] learning in land cover clustering like
[55:21] input satellite data set pre-processed
[55:24] by G means clustering like different
[55:26] type of clustering and you can get like
[55:30] your images as a output like no training
[55:34] label data set are required in
[55:37] unsupervised
[55:38] type of learning methods. So what are
[55:41] the advantages of this technique like no
[55:44] label data required useful for uh
[55:49] exploratory analysis and also fast
[55:53] implementation cost effective and it's
[55:57] uh can reveal the hidden uh hidden
[56:00] pattern in the data sets. So limitation
[56:04] lower accuracy
[56:06] then the supervised method
[56:08] interpretation required difficult to
[56:10] determine optimal number of cluster and
[56:13] result may vary depending on the
[56:16] algorithm settings.
[56:18] So
[56:20] difference between the uh supervised
[56:24] learning and unsupervised learning like
[56:27] non-le data sets, train model and
[56:29] predict classes. Here no label data set
[56:33] find pattern and create clusters.
[56:40] So this is already discussed in the uh
[56:43] previous slides. So like what is the
[56:46] workflow for training labels machine
[56:48] learning predict classes unlabelled data
[56:51] sets and like the geospatial examples
[56:54] already we discussed in the previous
[56:56] slides. So training data preparation. So
[57:01] why training data important? So in
[57:03] machine learning the quality of the
[57:05] training data directly affected the
[57:07] performance of the models such as so
[57:10] poor training data poor model
[57:12] performance good training data and
[57:15] accurate and best predictions. So in a
[57:18] machine learning model can only learn
[57:21] from the information required during the
[57:23] training. So what of data you you feed
[57:27] into the models? the model pretty uh
[57:29] learn from these type of data and their
[57:32] prediction depend on the
[57:35] training data or input data. So what is
[57:38] training data? Training data consist of
[57:40] sample with known the class levels used
[57:43] to tease them model like you can such as
[57:47] the example sample one NBDI like 0.8 and
[57:52] its uh elevation identified at 350 m. So
[57:58] it represent forest class to like you
[58:01] can these type of data sets like label
[58:03] data sets. So it uh create the
[58:06] difference or learn the difference
[58:08] relationship between features and
[58:10] classes. So different type of sources
[58:13] training data sets like field survey h
[58:16] existing GISS data sets remote sensing
[58:20] training data collection workflow
[58:23] already we discussed in the previous
[58:24] slide like satellite images and if
[58:27] identifying sample area assign classes
[58:31] levels create training data sets then
[58:33] quality check of the data sets. So data
[58:37] set splitting strategy. So here the uh
[58:41] evaluate model performance the data uh
[58:45] uh frequently or commonly divided into
[58:47] the two parts like training data set and
[58:51] testing data set. So the
[58:54] amount of data sets uh um largely used
[58:59] for training like uh 70 to 80%.
[59:04] But it depends on the uh volume of the
[59:09] data set. It's highly depends on the uh
[59:13] number of data sets or volume of the
[59:15] data sets. So testing it's like 30 to
[59:19] 20% or 10% or you can give the 10%. Like
[59:25] these data set are used to train the
[59:27] model and these data sets used to
[59:30] evaluate the model performance. So
[59:33] characteristics of the good training
[59:35] data like it representative, accurate,
[59:38] balanced, diverse and sufficient
[59:41] quality. So example of balance and
[59:45] imbalanced data like balanced data sets
[59:49] like forest 500 samples, H water 500
[59:54] samples, urban 500, agriculture 500. But
[59:59] in the unbalanced data sets like you uh
[01:00:03] your the data volume of the forests is
[01:00:06] highly greater as compared to other data
[01:00:08] sets. So the model learn highly from
[01:00:12] these type of data sets not learn these
[01:00:14] type of data sets. So your prediction uh
[01:00:18] can be different or uh can be very um uh
[01:00:25] um prediction based on the your data
[01:00:29] sets. So common challenge in the
[01:00:32] training data set uh preparation like
[01:00:36] misleading
[01:00:37] mixed pixels like different type of mix
[01:00:42] uh pixels used in the one class
[01:00:45] insufficient samples class imbalance or
[01:00:48] temporal mismatch.
[01:00:52] So geospatial examples like
[01:01:00] training samples like forest, polygons,
[01:01:04] water bodies, urban areas and
[01:01:07] agriculture fields. So these samples are
[01:01:10] used to train your classifiers such as
[01:01:13] random forest or SBM. So these are
[01:01:16] supervised learning method. So best
[01:01:20] practices like collect representative
[01:01:23] samples, verify levels carefully
[01:01:26] before fing into the model. Include all
[01:01:29] the learn cover classes, maintain class
[01:01:33] balance, use the independent string
[01:01:37] data sets, update the data sets.
[01:01:41] So feature engineering or selection and
[01:01:48] of the ML model. So what is feature? So
[01:01:51] feature is a measurable characteristics
[01:01:53] or attributes used by the machine
[01:01:56] learning model to make predictions like
[01:01:59] it uh it uh work to identify vegetation
[01:02:04] from the satellite imagery
[01:02:07] model H like red band, green band, blue
[01:02:12] band and
[01:02:14] I band H. So these variables called
[01:02:18] features like satellite data has
[01:02:21] different type of features already
[01:02:23] exist. So we uh uh classify them them or
[01:02:30] we identify them by using feature
[01:02:33] engineering or feature selections.
[01:02:37] So why uh are the feature important?
[01:02:41] So feature provide the information that
[01:02:43] help the model distingute or difference
[01:02:47] between uh uh
[01:02:50] uh different classes like here low n
[01:02:55] reflectance represents water, high NDVI
[01:02:59] represents forex, urban represents the
[01:03:03] high reflectance in visual bands and
[01:03:05] agriculture
[01:03:07] um represents the seasonal spectrum
[01:03:10] patterns. So without these useful
[01:03:13] features the model cannot learn
[01:03:15] effectively. H so your uh prediction um
[01:03:20] are affected
[01:03:23] type of geospatial features like
[01:03:26] spectral
[01:03:28] spatial and top uh topographic features
[01:03:32] and temporal features. So in spectrum
[01:03:35] like you can use the uh different type
[01:03:38] of satellite imagery like red, green,
[01:03:40] blue and error and sir
[01:03:43] also in spatial features like
[01:03:46] neighborhood characteristics like
[01:03:48] textures shapes age information object
[01:03:51] size in topographic features like uh
[01:03:55] derived from the DM data like elevation
[01:03:58] slopes aspects in temporal like change
[01:04:02] over the time like NDBI, time series,
[01:04:06] seasonal variation and crop growth
[01:04:09] pattern. So common feature used in the
[01:04:14] uh GOI like NDBI, vegetation monitoring,
[01:04:18] elevation like terrain analysis, slope
[01:04:22] like landslide study, temperature like
[01:04:25] climate analysis, rainfall
[01:04:28] used for fluid prediction and land
[01:04:32] surface and temperature used for urban
[01:04:35] heat detection. So feature selection. So
[01:04:40] here
[01:04:42] feature selection is the process of
[01:04:44] choosing most relevant feature from the
[01:04:48] sorry for the models. So the feature
[01:04:51] selection like what type of uh uh we uh
[01:04:57] want the prediction or classification
[01:05:00] like water
[01:05:07] forest.
[01:05:11] you can uh classify or uh identify most
[01:05:16] relevant uh features for the model
[01:05:21] training like objective all features
[01:05:24] select important features and train the
[01:05:28] better models. So why perform feature
[01:05:30] selections? It's already advantages
[01:05:33] improve model accuracy reduction in
[01:05:36] computational cost faster training
[01:05:38] reduction in overfitting and simplified
[01:05:41] model interpretation like example you
[01:05:44] can vegetation classifications. So here
[01:05:48] without NDBI like feature red, green and
[01:05:52] blue H here and with NDBI red, green,
[01:05:57] blue and NDBI. So without NDBI you can
[01:06:02] seen here like this but with NDBI you
[01:06:06] can so most relevant more accurate
[01:06:09] vegetation detection
[01:06:12] because NDBI directly capture
[01:06:16] the plant health. So these bands are
[01:06:21] represents the different type of
[01:06:24] classes. So what are the feature
[01:06:27] engineering? So feature engine creating
[01:06:30] new features from the exit ad existing
[01:06:33] data. For example, create NBI from the
[01:06:38] red band and NI bands. So write mixer of
[01:06:41] red band and N I band
[01:06:46] create ND VI bands. So this type of uh
[01:06:51] feature engineering uh mainly used in
[01:06:54] fluid mapping like elevation, flops,
[01:06:58] distance and rainfall and lands cover.
[01:07:03] So challenges during the feature
[01:07:05] selection
[01:07:07] and feature engineering too many
[01:07:09] features increase complexity high
[01:07:11] computation because it large data sets
[01:07:16] irrelevant feature like reduction
[01:07:18] accuracy increased noises redundant
[01:07:21] features like unwanted feature duplicate
[01:07:24] features unnecessary pre-processing
[01:07:27] temporal mismatching like training data
[01:07:29] and imaginary acquire different type
[01:07:35] times or different
[01:07:38] models to train this data set.
[01:07:42] So machine learning workflow for geo
[01:07:46] classification in one terms like into
[01:07:49] like here you can uh seen the data set
[01:07:52] reprocess training feature selection
[01:07:56] model training classification
[01:07:59] and here accuracy assessment and finally
[01:08:02] you can gener generate the map. So this
[01:08:06] is the total workflow already we
[01:08:08] discussed in the previous slides. So
[01:08:11] like data acquisition, pre-process
[01:08:13] training, feature engineering, model
[01:08:15] training, classification, accuracy and
[01:08:19] map generation. So application of uh
[01:08:22] geospatial domain like already we
[01:08:25] discussed in the uh uh application in
[01:08:29] different slides.
[01:08:32] So common machine learning algorithms
[01:08:35] for the geospatial classifications.
[01:08:38] So
[01:08:42] uh already we discussed this slide like
[01:08:44] unsupervised
[01:08:46] unsupervised unsupervised and deep
[01:08:49] learning.
[01:08:54] So here you can see in the decision tree
[01:08:57] for the decision tree used for the uh
[01:09:00] land cover classification, soil mapping,
[01:09:03] fluid um successibility mapping and
[01:09:07] random forest like different type of uh
[01:09:11] combines decision tree use voting for
[01:09:14] final classifications like final output
[01:09:17] such as forest. So it has high accuracy
[01:09:20] robust handling large uh data sets
[01:09:23] reduce overlifting. So in GOI
[01:09:27] application like landsc
[01:09:31] mapping crop classification
[01:09:33] environmental and change detection. So
[01:09:36] SBM SBM is the find the optimal boundary
[01:09:40] hyperplanner
[01:09:42] hyper plane that uh separate different
[01:09:45] classes. to different classes and it uh
[01:09:51] uh make a boundary between two classes.
[01:09:54] So it advantage excellent performance
[01:09:57] with limited uh training samples
[01:10:00] effective in high dimensional data sets.
[01:10:03] Widely used in um remote sensing and
[01:10:08] geospatial
[01:10:10] um monitoring like changing detection
[01:10:14] sorry uh vegetation urban area etc. And
[01:10:19] here is one thing is a anon. So anon is
[01:10:23] a mimic of the structure of human brain
[01:10:27] like they used neural network to make
[01:10:31] the models. So what are the advantages
[01:10:34] of these type of models? Learn complex
[01:10:36] nonlinear relationships handle large
[01:10:40] data set good prediction ability. So
[01:10:44] their application in GOI like
[01:10:46] environmental assessment,
[01:10:49] LCL classification, climate analysis,
[01:10:54] K means clustering already be discussed.
[01:10:57] DB
[01:10:59] yes CN already. So CNN, CNN is a
[01:11:03] convolution neural network which used to
[01:11:06] classify or extract the features from
[01:11:08] the images. So this uh algorithm
[01:11:13] uh frequently used in when the data set
[01:11:19] is
[01:11:21] in image format because it handle uh
[01:11:25] efficiently image data sets. So you and
[01:11:29] uh net deep learning models. So here the
[01:11:32] special uh CNR architecture for image
[01:11:35] segmentation like it has the encoder and
[01:11:38] decoder uh path
[01:11:41] and segment the satellite images.
[01:11:48] So practical demonstration uh for the uh
[01:11:52] land use and land cover uh
[01:11:56] classification using the deep learning
[01:12:13] So
[01:12:16] I will show my code
[01:12:19] how to implement
[01:12:22] uh this uh data uh set into the
[01:12:30] CNN model
[01:12:34] by using different type of image or
[01:12:37] different type of classes.
[01:13:11] I request all the participants for
[01:13:12] attention. Due to some technical issues,
[01:13:15] sir is on a 5 minutes break. He'll be
[01:13:17] right back by then. It's a 5 minute
[01:13:20] breaks for you all. Due to some
[01:13:21] technical issues, he'll be right back.
[01:13:58] Okay.
[01:14:00] So,
[01:14:03] is my screen visible?
[01:14:08] Hello.
[01:14:12] Hello, ma'am. Okay. Okay. Okay. Okay.
[01:14:19] So here uh
[01:14:26] we used
[01:14:28] Python or a Jupiter
[01:14:33] notebook platform to implement
[01:14:37] different type of satellite images
[01:14:44] to predict
[01:14:46] classic dictations. So
[01:14:52] in this
[01:15:00] class
[01:15:03] in this class
[01:15:03] we use different
[01:15:07] one
[01:15:11] represent
[01:15:20] So here uh
[01:15:43] Okay. So here you can uh seen different
[01:15:47] type of classes are present like this is
[01:15:51] the uh
[01:15:53] uh uh euro satellite image h
[01:16:01] captured by the satellites. So here
[01:16:04] different type of image or large amount
[01:16:07] of image are uh represent the annual
[01:16:11] crops and similarly like for forest. So
[01:16:16] different type of image are uh captured
[01:16:20] and pre-processed or labeled for the
[01:16:23] prediction of
[01:16:26] uh uh forest.
[01:16:28] So these type of uh image or these type
[01:16:32] of uh uh data uh are label data sets
[01:16:37] which are used to predict uh different
[01:16:41] classes.
[01:17:08] So here we uh normally uh
[01:17:13] use the Jupyter notebook to visualize to
[01:17:17] uh load to uh these type of large data
[01:17:23] sets uh efficiently.
[01:17:25] So we write the some uh programming code
[01:17:28] for this like import the operating
[01:17:31] system. Uh this is the mainly uh
[01:17:34] different type of AI uh libraries or
[01:17:38] tensorflow libraries. These are used uh
[01:17:42] to mainly in the machine learning or goi
[01:17:47] like pandas for classifying the uh tip
[01:17:51] files or met plot leave for uh plotting
[01:17:55] purposes and skarn models import the uh
[01:18:02] uh
[01:18:04] uh dividing or splitting data set into
[01:18:07] different categories. So many different
[01:18:11] type of uh
[01:18:14] uh libraries here use to the uh perform
[01:18:19] this project. So uh uh
[01:18:24] here we first
[01:18:27] use the data set path to
[01:18:33] load the all images. So you can uh uh
[01:18:38] like
[01:18:46] directly you can
[01:19:14] Yes. So here you can load the data by uh
[01:19:19] uh giving the uh uh path of the data
[01:19:22] like I keep my data in the uh on the
[01:19:27] desktop
[01:19:31] like here. So generally you can just
[01:19:36] copy as a part of data and directly
[01:19:40] paste into here. So here you can
[01:19:42] directly uh load your data sets. So like
[01:19:46] we run this. So here this is the detail
[01:19:50] of uh tensorflow library version like
[01:19:53] 2.161.
[01:19:55] So this is version capable of hand
[01:19:58] handling these type of data sets. So
[01:20:01] here we just run the uh load the data
[01:20:04] sets in second uh cell and in third cell
[01:20:08] we just uh uh uh uh confirm this uh path
[01:20:15] have uh data set or not. So path exist
[01:20:20] yes path is not false because it exist.
[01:20:24] So also after that we verify the classes
[01:20:28] like how many classes are present. So we
[01:20:31] write a code for classes to uh detect
[01:20:36] how much classes it has or what are the
[01:20:39] classes like here you can uh seen.
[01:20:46] Yes. So you can uh seen like animal
[01:20:50] crops, forest is already
[01:20:54] here you can
[01:20:56] seen like
[01:20:58] uh animal crops forest. So these classes
[01:21:02] are already present and we can see here
[01:21:07] h. So and after that we
[01:21:12] uh list the uh different classes and uh
[01:21:19] how much each image contain in each
[01:21:23] classes. So
[01:21:27] generally run and the annual crop has
[01:21:31] 3,000 image and forest has 3,000 image
[01:21:37] and herb vegetation is 3,000 and
[01:21:41] highway. So different
[01:21:44] uh uh classes has different number of
[01:21:49] images satellite images.
[01:21:52] So here uh this is uh the
[01:21:58] totaling the number of classes like how
[01:22:00] much classes you have. So we generally
[01:22:03] we have 10 classes. So load again the
[01:22:08] euro set data sets in two axis like
[01:22:12] xaxis and yaxis. So generally you can
[01:22:16] store the data or uh uh uh call the data
[01:22:21] uh how much data at the uh uh uh axis
[01:22:25] side uh x-axis side and how much data on
[01:22:29] the
[01:22:32] y side and what are the data is x side
[01:22:34] and what are the data is y uh y side. So
[01:22:38] here we generally write the simple code
[01:22:41] to uh uh uh to uh classify or identify
[01:22:47] these classes
[01:22:49] uh uh on the different axis.
[01:22:55] So like here you can uh seen
[01:23:00] xaxis data has
[01:23:04] 27,000
[01:23:07] h how much 27,000
[01:23:11] so
[01:23:13] n4
[01:23:16] and 64 it
[01:23:20] pixel size like image has 64x 64 square
[01:23:26] like image uh uh it has uh all images
[01:23:33] must be in same resolution H before
[01:23:36] feeding like these type of uh labeling
[01:23:40] called uh
[01:23:43] supervised uh learning.
[01:23:47] So uh at the uh y-axis it 27,000 data
[01:23:52] sets. So again we print the uh uh what
[01:23:56] are the x shape and what are the y. So
[01:23:59] you can directly
[01:24:02] uh write the code for the print to the
[01:24:05] these type of data. So here uh like you
[01:24:08] can uh see in the unique labels like the
[01:24:11] identified the labels. So we have the 10
[01:24:14] classes or 10 levels like 0 to 9. So
[01:24:18] data set summary.
[01:24:21] So already we discussed total number of
[01:24:23] images 27,000
[01:24:26] image seaps are 24 sorry sorry uh 64
[01:24:32] into 64 and 16 is the input size and
[01:24:35] total number of uh classes is 10.
[01:24:42] So visualization so how the data uh our
[01:24:46] input data looks. So we uh write the
[01:24:50] code for the uh visualize the uh data
[01:24:54] like figure size we can uh give the 16
[01:24:57] and four into length like one time into
[01:25:01] four data h. So what are the uh here the
[01:25:07] images uh classes like what are the
[01:25:10] represent to the RGB, NDVI and uh uh
[01:25:14] NDVI and NDBI.
[01:25:18] So here you can see so annual crop uh
[01:25:22] crops RGB data represents like NDBI,
[01:25:27] NDI
[01:25:29] and NDBI. So different uh uh type of uh
[01:25:37] uh
[01:25:38] uh
[01:25:40] data sets uh represent different type of
[01:25:47] uh bands or
[01:25:55] colors like here
[01:26:17] So uh don't confuse with the uh NDVI,
[01:26:23] NDWI
[01:26:24] and NDVI. So these are uh like spectral
[01:26:28] indices. Uh so like uh uh uh butter
[01:26:35] these indices. So these uh indices
[01:26:38] normally RGB images only show the what
[01:26:42] the human eye can see. So like if the uh
[01:26:46] uh uh different type of uh uh satellite
[01:26:51] image be used or uh for uh spectral
[01:26:55] information that uh help to identify uh
[01:26:59] different classes like vegetation, water
[01:27:02] bodies and human uh sorry urban regions
[01:27:06] uh more accurately. So these indices
[01:27:09] like uh NDVI
[01:27:12] stand for normalized difference
[01:27:14] vegetation index uh purpose of detection
[01:27:18] of the uh vegetation detection crop
[01:27:20] health monitoring and
[01:27:23] and forest monitoring. So I have uh uh
[01:27:27] different uh a uh PPT for uh these uh uh
[01:27:34] indexes.
[01:27:35] So like uh NDWI
[01:27:39] uh so it is represent the uh normalized
[01:27:42] difference uh water index NDBI like
[01:27:47] buildup normalized difference buildup
[01:27:49] index. So
[01:27:52] we represent different type of images or
[01:27:57] by different spectral indices here you
[01:28:01] can see. So generally just write a code
[01:28:04] and it it give everything to you. Huh.
[01:28:08] So this is the uh beauty of the AI and
[01:28:12] uh uh uh ML and DL model. So you can
[01:28:16] write only write a simple code with a I
[01:28:20] think this is a
[01:28:22] uh 20 or 30 lines of the codes but it
[01:28:25] gives like the lot of information uh
[01:28:29] about the spectral indices h of the uh
[01:28:33] different satellites satellite image of
[01:28:36] different classes.
[01:28:41] So here so
[01:28:43] uh like we uh write a code for decoding
[01:28:47] these uh images into number of class. So
[01:28:51] like original uh original number uh
[01:28:54] label shapes like 20,000 and encoded
[01:28:57] also it
[01:28:59] uh uh label shapes in 20,000 images but
[01:29:05] it has 10 class already we discussed.
[01:29:09] So here here we divide the uh uh uh uh
[01:29:15] uh data sets into two categories. So for
[01:29:20] uh two categories like um
[01:29:23] uh training data sets and testing data
[01:29:27] sets.
[01:29:30] So here look at uh X uh represent the X
[01:29:36] training data sets and X test represent
[01:29:39] the X uh uh store the uh uh test data
[01:29:44] sets and such as Y and Y test. So like X
[01:29:50] and X cat uh uh uh divided into two
[01:29:55] categories like test data set size into
[01:29:59] only 20%.
[01:30:02] So like to
[01:30:09] so here you can uh see in here training
[01:30:13] data set divided already 80% of data
[01:30:17] sets
[01:30:20] for training and for traing purpose
[01:30:26] only 20% data of used for testing
[01:30:30] purpose.
[01:30:31] So build the CNN model. So here the
[01:30:35] building of the CNN model. So uh uh uh
[01:30:39] different type of uh libraries are used
[01:30:44] to uh
[01:30:47] uh build or uh uh these
[01:30:58] model
[01:31:07] So, so here the uh uh in the uh uh CNN
[01:31:14] convolution neural network model uh it
[01:31:18] uh has the input. So we give the input
[01:31:20] like input what are the your input?
[01:31:22] input is a 16x 16
[01:31:27] sorry 64x 64
[01:31:30] uh into 16 input images h how much
[01:31:35] images so for training purposes we used
[01:31:39] uh
[01:31:41] 20,600
[01:31:44] data sets and for testing we used the uh
[01:31:49] uh 5,400 100 data points but input set
[01:31:55] uh uh uh shape is uh 64 by 64 and by 16.
[01:32:02] So convolution uh 2D uh dimensional. So
[01:32:07] here we use uh this is the filter like
[01:32:11] for 2D we use 32 layer here max pooling
[01:32:16] batch normalization flatten dense and
[01:32:19] draw. So here model sequential. So for
[01:32:22] the u uh convolution we used the 32
[01:32:27] filters for max pooling we used uh but
[01:32:31] uh 3 into 3 + 3 max pooling and batch
[01:32:35] normalization
[01:32:37] uh
[01:32:39] and uh so in the batch normalization
[01:32:42] here in the here we use like 32
[01:32:48] uh
[01:33:02] Okay.
[01:33:06] Ah yes yes. So we use the 16 into 16 by
[01:33:10] 18.
[01:33:12] So these these these uh different type
[01:33:15] of uh uh uh uh model architecture uh
[01:33:23] uh uh component uh and it it defined
[01:33:27] value
[01:33:28] like we used max pooling to max pooling
[01:33:32] unit uh again convolution like stack
[01:33:36] convolution again 64 and 3x3 uh uh uh
[01:33:41] Max pooling h activation function right
[01:33:45] uh relu activation function are used and
[01:33:48] padding like same h so this is the uh uh
[01:33:53] uh model uh parameters like how much
[01:33:57] output set uh output shape of these uh
[01:34:01] uh different type of uh convolution
[01:34:05] neural network architectures.
[01:34:08] So here uh you can uh see the dimension
[01:34:15] like here 64 is a height and this one is
[01:34:19] a 64 is a height but you know 32 is a
[01:34:23] layer of this conolution to the network.
[01:34:26] So here the uh total uh total parameter
[01:34:31] like uh
[01:34:34] like uh here you can uh see the two lakh
[01:34:38] uh uh uh
[01:34:41] like 20 uh 21 lakh 97,000
[01:34:45] uh uh 86 so it this data uh contain 8 MB
[01:34:52] okay so 8.3 MB H. So
[01:34:58] this is the uh model volume like uh uh
[01:35:05] uh model uh uh uh train or process up to
[01:35:12] 8 MB data sets or data have a volume 8
[01:35:17] MB
[01:35:18] like compile. After that they compile
[01:35:21] the data uh uh model by using the
[01:35:25] different type of optimizer like we use
[01:35:27] here uh uh Adam optimizer learning rate
[01:35:31] like 10 to power uh minus 3 and compile
[01:35:35] with different type of category and
[01:35:38] matrices. So here you can uh see the
[01:35:42] history of the fit models.
[01:35:48] So
[01:35:50] also you can uh you can uh see the
[01:35:52] different we use the uh learning H. So X
[01:35:57] train and Y train. So validation speed
[01:36:00] here uh and epoic epochs are used like
[01:36:04] only three and huh bad side batch size
[01:36:07] is here 32 32 and verb is one. So I used
[01:36:12] only three uh epochs for uh if you uh
[01:36:16] for uh uh uh uh uh less time consuming
[01:36:22] because uh
[01:36:25] number of epoxa the model uh prediction
[01:36:29] must be accurate
[01:36:33] or will be accurate. So CNN evaluation
[01:36:37] how do we evaluate the uh performance of
[01:36:40] the uh uh accuracy of the models? So
[01:36:44] total um uh performance accuracy is only
[01:36:48] 56%.
[01:36:50] So how we predict the uh uh uh uh
[01:36:55] prediction at the uh uh y levels. So and
[01:37:00] next term is a visualiz visualization of
[01:37:03] the prediction. So here you can see the
[01:37:07] annual cough image look like in the uh
[01:37:11] uh
[01:37:14] actual image like here but so it
[01:37:18] represents so in the different type of
[01:37:24] confidence huh like so at annual crops
[01:37:28] and predict predict permanent crops h
[01:37:32] but the permanent crops
[01:37:34] are here.
[01:37:37] So what are the confidence level of this
[01:37:40] model? So this model like it has
[01:37:44] confidence up to 73%
[01:37:49] like it is the permanent crops in this
[01:37:54] area
[01:37:57] also like industrial you can see
[01:38:00] different type of bands which represent
[01:38:02] the uh uh industrial and other things.
[01:38:07] Uh similarly annual crops you can see
[01:38:10] 90%
[01:38:12] forest forest up to like uh uh 93%
[01:38:19] uh and uh industrial again here
[01:38:23] permanent crop. So different type of
[01:38:26] these MSS uh
[01:38:30] predicted.
[01:38:33] So here simply you can just uh without
[01:38:37] the uh uh uh confidence level or
[01:38:42] accuracy level. So you can like you can
[01:38:45] see here highway how look like but here
[01:38:48] you can see how the highway predict by
[01:38:50] the AI models.
[01:38:55] Huh.
[01:38:58] Similarly
[01:39:02] no budget.
[01:39:05] So here what what are the precision? How
[01:39:09] the model are uh precise like for the
[01:39:12] annual crop it it precision accuracy uh
[01:39:15] up to one uh it's highest uh 0 to one be
[01:39:20] like 100% forest for 94% and uh uh it
[01:39:25] recall ability like like 11% and uh
[01:39:29] [clears throat]
[01:39:30] it uh F1 score uh for this uh
[01:39:37] accuracy like the uh uh uh
[01:39:42] uh 20%. So uh this this these are
[01:39:46] balance the the FN score balance the
[01:39:49] precision and recall accuracy. H so like
[01:39:53] forest the it precision uh
[01:39:59] [clears throat] 94%.
[01:40:00] Huh. So lowest precision in highway.
[01:40:07] Huh. So but it recall accuracy up to
[01:40:11] 79%.
[01:40:15] So also we verify by this the confusion
[01:40:18] matrix like you can see in here. So this
[01:40:22] is the image represent the uh data sets
[01:40:28] are false. Hm like you can uh see in
[01:40:33] here like these data sets like annual to
[01:40:38] annual and 4 is to this this represent
[01:40:41] the uh how much uh the prediction you're
[01:40:44] accurate
[01:40:55] like the uh diagonal value of This uh uh
[01:41:00] uh uh uh data set uh represent the uh
[01:41:06] prediction h or off diagonal data sets
[01:41:11] it's represent the uh
[01:41:16] mismatch or misclassifications.
[01:41:22] So here
[01:41:23] we run.
[01:41:28] Yes, it work like you can uh see in here
[01:41:30] the annual crop accuracy is 11%, 6%, h
[01:41:35] so highly up to lake. So lake is a uh up
[01:41:39] to 97% accuracy.
[01:41:45] So here you can also plot the uh uh uh
[01:41:52] training uh accuracy versus epoch. So
[01:41:56] how we increase the epochs? So how the
[01:41:58] accuracy
[01:42:00] are increased right and how the training
[01:42:03] losses are
[01:42:07] decreases.
[01:42:10] Here also we uh uh uh get the uh most
[01:42:17] importance of the output like the actual
[01:42:21] crops take a predicted permanent crops.
[01:42:24] H but it confidence level 73%. H so
[01:42:30] industrial and industrial 98%. So like
[01:42:33] this like it predict the sea lake to sea
[01:42:36] lake 100%. So it has no uh uh
[01:42:40] mclassification in sele. So here the
[01:42:43] confidence level you can see the
[01:42:45] different type of images like up to uh
[01:42:50] we uh test the uh up to 539 data sets.
[01:43:12] Okay.
[01:43:14] Hello.
[01:43:19] Yes.
[01:43:22] So
[01:43:25] I placed my from my side. So any
[01:43:28] question
[01:43:30] from the audience?
[01:43:31] So all the participants who have a
[01:43:33] question kindly raise your hands.
[01:43:36] Okay.
[01:43:37] Sir, we have a question from Samradish.
[01:43:40] Okay.
[01:43:42] Hello, sir. Am I audible?
[01:43:44] Yes.
[01:43:45] Sir, uh can you please uh share your
[01:43:48] email id so that I can ask doubts there
[01:43:50] as well?
[01:43:53] Okay. So, uh
[01:43:56] can you write it in chat box?
[01:43:58] Yes, please write in chat box.
[01:44:02] So, kindly share your email id in the
[01:44:03] chat box. I have enabled the chat box
[01:44:05] for everybody.
[01:44:18] Yes, I posted on chat box.
[01:44:28] Any other student?
[01:44:35] Sir, I'm unable to see it.
[01:44:39] Uh ma'am, please share.
[01:44:44] Yes sir, I have
[01:44:46] in the chat box.
[01:44:49] I hope SR you able to see his email.
[01:44:53] Yes, ma'am.
[01:44:54] I have a next question.
[01:44:57] Uhhuh.
[01:44:58] Um yes. Uh
[01:45:02] I didn't understand the code you taught.
[01:45:06] Uh can you please uh how can I uh get
[01:45:09] it? No line by line.
[01:45:14] So uh first you have to learn about the
[01:45:18] uh uh Anaconda platform. So uh in which
[01:45:23] is a platform Jupiter notebook. So you
[01:45:28] can learn these uh code from uh
[01:45:33] different type of the uh uh websites or
[01:45:38] uh YouTube channels or different books.
[01:45:42] So I have uh uh two books I will share
[01:45:46] with uh via email or uh two uh uh ma'am
[01:45:53] and she's uh share with you directly. So
[01:45:57] in uh uh books uh the code
[01:46:02] identification how you can write it and
[01:46:04] how uh you implement implement these uh
[01:46:09] quotes or data set at a small level. So
[01:46:13] you can learn anything uh by using also
[01:46:18] uh you can use the AI boards for uh
[01:46:21] understand the uh uh uh codes in your
[01:46:25] own way. So uh there is a mainly
[01:46:29] simplest way you can use the AI and
[01:46:32] verify by these uh
[01:46:52] Can I speak in Hindi? Yes sir. Okay. So
[01:47:01] S information.
[01:47:04] Okay sir. And one more thing can you
[01:47:06] share this P also? Uhhuh. Yes. Yes I
[01:47:10] share.
[01:47:12] Okay. Thank you. Okay.
[01:47:18] Sir next we have a question from Juhi.
[01:47:20] Juhi please ask a question.
[01:47:25] Good evening sir.
[01:47:27] Yes good evening
[01:47:28] sir. Actually I also have a difficulty
[01:47:31] to understand the code the same said by
[01:47:34] the
[01:47:37] sir please can you share the ppt and the
[01:47:40] jupyter notebook with us? Yes sir. I uh
[01:47:44] I will share uh my PPT as well as code
[01:47:48] with us. Okay.
[01:47:54] Sir, next we have a question from
[01:47:55] Darini. Darini, please ask a question.
[01:48:01] Dashi, do you have a question?
[01:48:08] Okay.
[01:48:09] So I guess there are no more doubts. Do
[01:48:12] you have anything to say or shall I
[01:48:14] conclude the session now?
[01:48:19] So am I if you Yes. Yes, you are
[01:48:22] audible. Uh
[01:48:24] so
[01:48:26] so do you have anything to say? Shall I
[01:48:28] conclude the session now?
[01:48:38] Am I audible?
[01:48:47] Participants. Am I audible?
[01:49:00] Sir, am I audible to you?
[01:49:13] Yes ma'am.
[01:49:15] Now I'm on a
[01:49:17] So So do you have anything to say? Shall
[01:49:19] I
[01:49:25] I request to participants uh if you
[01:49:28] I I request to participants uh if you
[01:49:28] have any question you can directly use
[01:49:30] my email to ask question.
[01:49:37] So sir shall I conclude?
[01:49:39] Yes please. Okay.
[01:49:41] Uh I would like to inform all the
[01:49:43] participants that I have once again
[01:49:44] dropped sir's email id in the chat box.
[01:49:46] You can check it. I'm dropping it once
[01:49:48] again. Thank you sir for such an
[01:49:50] insightful session. Thank you all the
[01:49:52] participants. Now you all can leave the
[01:49:54] meeting.