Land Cover Classification in Google Earth Engine

Apply random forest to land cover classification in Google Earth Engine by training with labeled Landsat 8 pixels, using many decision trees and majority voting on spectral bands and NDVI.

Overfitting in land cover classification on Landsat data occurs when a random forest memorizes pixels; prevent with bagging, random features, diverse data, balanced classes, limited depth, many trees, and cross-validation.

Create a training dataset for land cover classification by selecting cloud-free imagery from July 1 to August 31, 2022 and labeling points for water, vegetation, urban area, and bare land.

Combine training points and label them 0 for water, 1 for vegetation, 2 for urban, and 3 for bare land for the random forest, then verify names and map points.

Select Landsat bands red, near infrared, and swir1, compute ndVi to highlight vegetation, then combine these features for a land cover classification of fields, small towns, lakes, and bare soils.

Collect red and IR feature values, plus ndvi, at each training point to build a dataset for random forest learning on Landsat imagery.

Split a 113-point land cover dataset into 70% training and 30% testing for a random forest in Google Earth Engine, and verify the sizes in the console to prevent overfitting.

Train a 50-tree random forest in Google Earth Engine from 80 training points to classify pixels as water, vegetation, urban, or bare land using four features.

Apply a trained random forest to label every pixel in Qmra as water, vegetation, urban, or bare land, producing a land cover map in Google Earth Engine.

Map land cover in google earth engine with blue water, green vegetation, grey urban, and yellow bare land, using random forest with training points and red, nir, swir1, ndvi features.

Explore why four features balance speed and accuracy in land cover classification with random forest, why data splitting matters, and how 53 and testing points guide performance without overfitting.

Use a 70/30 train-test split to address overfitting, evaluate a random forest on unseen data, compute accuracy and a confusion matrix, and report overall and per-class performance for map trustworthiness.

Use the trained random forest to predict the class of each testing point (water, vegetation, urban or Maryland) and compare these predictions to the true labels from video four.

Create a confusion matrix to compare true classes with Random Forest predictions in a 4x4 table of water, vegetation, urban, and bare land, revealing per-class and overall accuracy.

Use a confusion matrix to reveal misclassified areas, like water vs vegetation, and adjust training points and features to improve producer's and user's accuracy.