Contains the Approach for False Positive Analysis and Cleaning of Crop Field Prediction Raster of Angola
# Notebook Overview
This Repo contains the approach for false positive analysis and cleaning of my crop field prediction with a UNET model in Angola.
The notebook guides users through a workflow for spatial data analysis and classification using a Random Forest model in Google Earth Engine (GEE) using GEEmap in Python. The goal is to create a mask for the prediction, to exclude mostly false-positive pixels.
The workflow is organized into three main sections:
#### Data Source Acknowledgment
This project uses satellite imagery from the PLANET constellation, provided through the NICFI (Norway’s International Climate and Forest Initiative) program. I gratefully acknowledge the PLANET-NICFI initiative for making this high-resolution data available, which is essential for the spatial analysis and crop field prediction in this notebook.
---
## 1. Data Preparation
- Loads and preprocesses predicted fractional map and land cover raster
- Creates 500 stratified proportional random sampling points in total
- The points were manually labelled with 1 = crop and 0 = non-crop
---
## 2. Sampling of Predictor Variables in Google Earth Engine
- Loads several relevant predictor variables from Google Earth Engine to discriminate crop fields and non-crop fields
- Stacks predictor variables and resamples them to 10 meter spatial resolution
- Extracts values based on sampling points
- Currently used predictor variables used to train the RF model:
- **Elevation** derived from ALOS
- **Precipitation**, **Evapotranspiration (ET)** and **Aridity Index'** derived from CHIRPS and MODIS
- **Sentinel-1 Radar** Imagery Median composite (01-2021 - 12-2024)
- **Enhanced Vegetation Index (EVI)** based on Sentinel-2
- **Class** variable (derived from Section 1 in notebook)
---
## 3. Training of Prediction of RF Model
- Parameters for RF model in GEE:
- `numberOfTrees=500`
- `variablesPerSplit=3`
- `minLeafPopulation=3`
- `bagFraction=0.85`
- `seed=42`
- Trains the model on the prepared dat …