AI_my Slides.pptx

RESEARCH METHODOLOGY: DEEP LEARNING WORKFLOW USE THE MODEL 1 PREPARE SENTINEL-2 IMAGE 4 Classify pixels using deep learning...

RESEARCH METHODOLOGY: DEEP LEARNING WORKFLOW USE THE MODEL 1 PREPARE SENTINEL-2 IMAGE 4 Classify pixels using deep learning Select inputs (Raster, Model) PREPARE TRAINING DATA TRAIN A MODEL 2TRAINING SAMPLE MANAGER Manually digitized training samples 3 Input training data Set the Model Type, Backbone Model EXPORT TRAINING DATA Set the parameters (Epoch, Batch Size, Chip Size, Metadata format, etc.) T E A RAMINAT M RESEARCH METHODOLOGY: DEEP LEARNING WORKFLOW By definition: Epoch - a single pass through the entire training dataset. It is used to measure the number of times the model has seen the entire dataset. Batch Size – the number of training examples used in one iteration on training process. Chip Size - refers to the dimensions of image chips used for training models, typically measured in pixels. Common dimensions include 256x256 pixels or 512x512 pixels depending on the dataset. Backbone model - The backbone in deep learning is like a central processing unit that extracts important features from input data, helping the model understand and interpret information effectively. T E A RAMINAT M RESEARCH METHODOLOGY: DEEP LEARNING WORKFLOW Deep Learning Workflow in ArcGIS Option 1: Use Pre-trained models Option 2: Train your own model Labelling Data Inferencing Train Model Preparation In pre-trained AI models eliminate this process Analysis Image Manageme nt Folder content: 1. ModelCharacteristics Field 2. model_metricts.html mobility, 3. Model2.dlpk monitoring T E A 4. Model2.emd 5. Model2.pth RAMINAT M  FINDINGS, RESULTS AND ANALYSIS 1. Development and consolidation of training libraries of Mangrove Forest INPUT OUTPUT 100 training samples 531 Image chips generated T E A RAMINAT M 2. Develop a deep learning model for the Mangrove Forest in the Province of Marinduque ANALYSIS OF THE MODEL PARAMETERS EVALUATION METRICS Tria l Trainin Model Batch Validatio Chip Backbon Precision Recall F1 Accuracy No. g Epoch Type Size n% Size e Model Sample ResNet- 0.838756 0.738499 0.785441 0.97288 1 100 U-Net 32 10 100 50 34 ResNet- 0.894645 0.922586 0.908407 0.96004 2 100 U-Net 32 10 100 100 34 ResNet- 0.868059 0.766774 0.814279 0.973750 3 100 U-Net 32 20 100 50 34 ResNet- 0.885036 0.856455 0.870511 0.95918 4 100 U-Net 32 20 100 100 34 ResNet- 0.865261 0.889431 0.87718 0.960110 5 100 U-Net 32 20 100 100 50 ResNet architectures (ResNet 34 and 50) Independent variables: Training samples, model type and chip size Dependent variables: Validation percent, backbone model and Epoch T E A RAMINAT M RESEARCH METHODOLOGY: DEEP LEARNING WORKFLOW Learning Curves Three common dynamics that you are likely to observe in learning curves; they are: Underfit Learning Overfit Learning Underfit. Curves Curves Overfit. Good Fit. Good fit Learning Curves T E A RAMINAT M 2. Develop a deep learning model for the Mangrove Forest in the Province of Marinduque Learning Curves X Axis - Progress Y Axis - TRIAL 1 TRIAL 3 Performance TRIAL 5 TRIAL 2 TRIAL 4 T E A RAMINAT M 2. Develop a deep learning model for the Mangrove Forest in the Province of Marinduque EVALUATION METRICS Trial No. Precision Recall F1 Accuracy 1 0.838756 0.738499 0.785441 0.97288 2 0.894645 0.922586 0.908407 0.96004 3 0.868059 0.766774 0.814279 0.973750 4 0.885036 0.856455 0.870511 0.95918 5 0.865261 0.889431 0.87718 0.960110 Trial #4 is the best model for the mangrove forest, it generated a smooth learning curve among the five trials, which is a sign of a good fit. T E A RAMINAT M 3. Comparative Analysis of the Mangrove Forest through Image FITNESS OF FEATURE Validation COMPARATIVE ANALYSIS CLASS PARAMETER 2024 Deep S 2020 Land Cover Learning Study Map Trial No. 4 (ResNet-34) Generally fit in the Mangrove Forest areas, 1. Fitness of the Generally fit in the however there are some generated feature Mangrove Forest areas. “salt-and-pepper noise”, class and misclassifications of non-mangrove areas. 2. Area of the 3,006 hectares 2,490 hectares generated (baseline) 83% area coverage shapefiles Location: Salomagui 3. Stratified Island, Random Sampling 100 sampling points 97 out of 100 sampling Sta. Cruz, Marinduque Points and classified as Mangrove points were classified as misclassifications Forest Mangrove Forest of non-mangrove areas. T E A RAMINAT M 3. Comparative Analysis of the Mangrove Forest through Image Validation MISCLASSIFICATIONS “SALT AND PEPPER” STRATIFIED RANDOM SAMPLING POINTS Sites: Municipalities of Mogpog, Santa Cruz, Boac & Torrijos T E A RAMINAT M  Comparison of conventional and Deep Learning process CONVENTIONAL PROCESS AI (DEEP LEARNING PROCESS) Preliminary Data Preparation for Province of Marinduque Image classification process: OBIA-Object Based Image Analysis (Image segmentation Deep Learning Model process process) Digital Interpretation Digital Interpretation Duration: 5 working days ￭ Downloading and Pre-processing: 3 days 10 working days ￭ Creation of Training Samples: 1 day ￭ Downloading and Pre-processing ￭ Export the data: 30 mins ￭ Preliminary Mapping of Mangrove ￭ Train the data: 1-2 hours ￭ Apply the Model: 1-2 hours Software: eCognition Software ArcGIS Pro with Image Analyst Tool License: Deep Learning ArcMap T E A RAMINAT M 4. Preliminary map of Mangrove Forest in the Province of Marinduque T E A RAMINAT M  CONCLUSIONS The best model developed for the mangrove forest used U- Net (Pixel Classification) with parameters: ResNet-34 as the backbone model, 32 batch size and 100 epochs. Based on the metrics, the model generated 0.96 accuracy (1 equals accurate). A strong correlation was observed in the learning curve using the 20% of the data for testing and the 80% for training. This indicates that the model is performing well enough to generalize the training data. Image validation accuracy of 97% was attained using 2020 Land Cover Map and deep learning model’s derived mangrove feature class. The application of deep learning significantly expedite the mapping of mangrove forest. T E A RAMINAT M T E A RAMINAT M RS