Prediction Mapping Using GIS Data and Advanced ML Algorithms

Discover how the caret package in R uses the train function to select algorithms, perform data splitting, preprocessing, resampling, hyperparameter tuning, and assess variable importance for GIS mapping.

Explore hyperparameters as external configuration and how grid search, random search, and Bayesian optimization tune models, with grid and random searches being common and Bayesian optimization advancing research.

Apply the extreme gradient boosting (xgboost) classifier for classification and regression, explaining how boosted trees grow with misclassified points, include regularisation, cross-validation, and pruning to control overfitting.

Explore the k nearest neighbors classifier, which uses the k closest data points to assign a class. The method is non-parametric and lazy, relying on Euclidean or Manhattan distances.

Explore Naive Bayes classification, its probabilistic approach and independence assumption, with examples of true/false positives and negatives and guidance on model tuning.

Learn how the random forest ensemble builds multiple decision trees to improve predictive accuracy, mitigate overfitting, and combine their votes for the final class prediction.

Explore selecting training and testing data for prediction mapping, including 65/35 splits, confusion matrices, and extending from sample incidents to area wide predictions in geographic information systems and machine learning.

Switch to a 64-bit system to access more RAM for training and testing in prediction mapping with GIS data and advanced ML, and use multiple software versions for compatibility.

Apply PM10 readings pre-processing in Excel by loading station data, handling missing values, and computing seasonal averages from 2017–2019, focusing on summer months.

Allocate the air monitoring stations and record data in QGIS by creating a new station field and entering concentration values to compute an average.

Learn to convert PM10 readings to WHO limits in QGIS by creating a level field, applying a classification equation, and mapping values to good, moderate, and unhealthy categories.

Identify a study area in northern Iraq, download Landsat 8 imagery from 2017–2019, filter by cloud cover, mosaic scenes, and clip to the area of interest for PM10 prediction variables.

Download Landsat 8 imagery via USGS service, choose surface reflectance product, understand corrected vs original data, and access multi-band outputs for GIS-based mapping.

Visualize downloaded Landsat 8 images, mosaic scenes, add land surface temperature and band 7, mask clouds, then clip to study area and average June–August 2017–2019 with code for a scene.

Process Landsat 8 bands and indices in R by setting up a project, loading libraries, and applying surface temperature formulas using metadata and image bands 1 through 11.

Extract and process Landsat 8 data in R to compute the land surface temperature, applying radiometric corrections and haze corrections, while clipping to a study area and stacking bands.

Process Landsat 8 bands in R to compute monthly and annual means for bands 2–6 and savi and built-up index, then derive three-year averages for GIS-based prediction mapping.

Learn to build a road network in QGIS by using the QuickOSM plugin to extract highways, export a road network shapefile, and compute distance to roads with euclidean distance.

Prepare the input dataframe in QGIS by aligning the target with conditioning factors, importing monthly and three-year mean values, and renaming rasters for the ambient air pollution susceptibility map.

Create training data by buffering stations and deriving grid-based conditioning factors in QGIS. Export final input variables to CSV for modeling, integrating join attributes and statistics.

Learn data preparation, feature engineering, and model evaluation for XGBoost in R, including data cleaning, normalization, train-test split, hyperparameter tuning, and generating classification and probability maps.

Learn how to run an XGBoost model with a train control for cross-validation, set hyperparameters, evaluate with accuracy, examine variable importance, and interpret a confusion matrix for three-class prediction.

Explore hyper-parameter tuning of the XGBoost algorithm via grid search and tenfold cross-validation, analyze parameter effects with plots, and identify the best model achieving about 81% accuracy.

Apply the XGBoost algorithm to multiclass air pollution classification, evaluate with ROC AUC plots, and interpret predicted class probabilities and confusion metrics on testing data.

Use the XGBoost algorithm to fit an optimized model on all inventory observations with the recommended tuning grid, then generate predictions from the original data and examine variable importance.

Learn to convert raster stacks to a dataframe for XGBoost, apply normalization and scaling, and generate probability and classification maps from GIS data.

Apply the XGBoost algorithm to generate and plot multi-class probability maps from GIS data, producing class-specific rasters for good, very unhealthy, and unhealthy conditions.

Discover how the XGBoost algorithm enables classification and production of prediction maps, merging three classes in one scene and converting level averages to class IDs for color-coded maps.

Apply the nb algorithm and use ggplot to visualize the linearity between the target and independent variables, evaluating predictors and class discrimination in the data.

Explore running a Naive Bayes model with default training settings, perform three-fold cross-validation on GIS data, assess accuracy across three classes, and interpret variable importance and predictions.

Optimize the NB algorithm through hyper-parameter tuning, comparing three options across three values for nine configurations, and evaluate fit and runtime while considering AUC of ROC plots and normalized rasters.

The nb algorithm produces probability maps for three classes and corresponding classification maps, while addressing prediction delays and adjusting grid resolution to improve accuracy and visualization.

Apply the knn model with 10-fold cross-validation and a k grid (1–31). Compare default versus tuned settings, analyze accuracy and confusion matrix, and note that default often matches tuning.

Explore k-nearest neighbors, ROC AUC evaluation, and probability and classification prediction maps using GIS data and advanced machine learning algorithms.

Conduct grid-search tuned random forest on air pollution data under limited ram; prepare data, train with 65/35 split, optimize mtry and max nodes, and assess feature importance and accuracy.

Compare grid search and random search for tuning a random forest in GIS data, optimizing roc auc, and interpreting variable importance for predictions.

Learn how to convert raster images to a dataframe for RF analysis, preparing the study area, handling rosters, normalizing data, and managing memory in large GIS workflows.

Load and prepare data for a random forest, rename features to match training input, and run probability predictions to generate a GIS-based probability map for the study area.

Apply the random forest algorithm to generate a classification prediction map from GIS data, after normalizing variables, aligning names with training data, and computing the final study-area projection.

Compare four models and visualize their nine-class probability and classification maps in QGIS using GIS data, random forest, and nearest neighbors.

Add developed tools to the QGIS processing library by standardizing coordinate systems across layers and using grid unit analysis to generate training and testing data.

Convert land cover string classes to numeric codes in a shapefile within QGIS, then convert the vector to a 30 meter raster for seamless integration into prediction mapping workflows.

Run the tools session guides you to prepare polygon or point data, convert polygons to centroids, compute landslide metrics within slopes, and identify safe areas with a slope threshold.

Map landslides and non landslides points using a 100 by 100 grid to train and test a GIS based ML model, with 30 by 30 prediction.

Run step 3 tools to build a landslide hazard model with GIS data and ML, integrating elevation, slope, aspect, land cover, curvature, roughness, and soil type for training and testing.

Copy pixel values into Excel, convert land cover attributes to numbers, and prepare training and testing data for GIS-based prediction mapping with advanced ML algorithms.

Clean GIS data in Excel for ML: remove columns, rename headers, extract mean and mode, check missing values, save training and testing CSVs, then import to an artificial neural network.

In R, this lecture shows preparing landslide GIS data in Iran for XGBoost by converting aspect and land cover to features, handling missing values, and normalizing for training and testing.

Apply the XGBoost train function using default settings, perform tenfold cross-validation with five times for small data, and report about 88 percent accuracy and 81 percent on testing.

Explore hyper-parameter tuning of the XGBoost model through grid search, evaluating rounds, depth, gamma, and subsample with 10-fold cross-validation to maximize accuracy and interpret feature importance.

Apply the XGBoost algorithm to binary predictions, evaluate with AUC of the ROC plot, and interpret prediction outputs for yes/no classifications.

Apply the XGBoost algorithm to run an optimized model on GIS data, generate probability prediction maps, compare with an artificial neural network, and identify key predictors like slope and elevation.

Use XGBoost for binary classification to create GIS-based prediction maps, join prediction keys to data frames, and visualize 30x30 cell grids in green and red.

Explore landslide data analysis with the arc nearest neighbor model: clean data, prepare training and testing sets, select features, scale values, and visualize elevation and slope to improve prediction.

Compare k-nearest neighbors models with and without hyper-parameter tuning using grid search and 10-fold cross-validation, evaluating accuracy and feature importance for land cover and elevation.

Learn how to apply the k-nearest neighbors algorithm, evaluate with roc auc, and troubleshoot data entry issues while training, predicting, and plotting results on gis data.

Develop probability prediction maps with the knn approach on gis data by stacking rasters, applying max–min scaling, and generating a yes/no probability map of outcomes.

Explore KNN classification for prediction maps using GIS data, interpreting probability and yes/no outcomes, and learn how to plot results on an x and y grid.

Train a Naive Bayes model for landslide classification using GIS data, entering training data, scaling, and visualizing variable relationships across categorical and continuous features.

Train the NB algorithm using a controlled cross-validated setup, tune hyper-parameters, and evaluate performance with the AUC of the ROC plot to optimize binary classification.

Use the naive bayes (nb) algorithm to produce probability and classification prediction maps from GIS data, detailing data prep, frame conversion, feature handling, and model outputs.

Prepare GIS training data for a random forest model by cleaning data, converting continuous to categorical variables, normalizing features, and assembling landslide incident data for training and visualization.

Explore the random forest training workflow with the default train function, tuning hyper-parameters via grid and random search, and evaluating accuracy and AUC of ROC plot.

Explore a random forest algorithm to generate probability and classification prediction maps from a geographic information system, using random search optimization of hyperparameters and careful data preparation for landslide prediction.

Visualize landslide susceptibility with four models in QGIS, comparing classification results and probability maps. Learn to interpret model differences, identify areas and create professional plots for decision makers.

Learn to map predictions with GIS data and advanced ML algorithms for air pollution and landslide susceptibility using four models including extreme gradient boosting, K nearest neighbors, and random forest.