For research use only. Not for clinical decisions. Copyright (c) 2026, Payton Jachec

S2_Feature_finder

An ML process for locating classes of features using RF and XGBoost across an extent using common geospatial analysis indices as inputs.

S2_Feature_Finder Contents

• ml_pipe.py
• environment.yml

Required Packages and Environment

Known Working Configuration

Package	Version
arrow	1.4.0
bokeh	3.9.0
dask	2026.3.0
distributed	2026.3.0
geopandas	1.1.3
mgrs	1.5.4
numpy	2.4.6
pandas	3.0.3
planetary-computer	1.0.0
pystac-client	0.9.0
pystac	1.14.3
python	3.11.15
rasterio	1.4.4
scikit-learn	1.9.0
scipy	1.17.1
shapely	2.1.2
stackstac	0.5.0
tqdm	4.68.0
xarray	2026.4.0
xgboost	3.2.0

Required Channels

• conda-forge

Environment Setup

It is recommended to use Anaconda or Mamba to create a venv for this program suite. If running on an HPC, default configurations will almost certainly break the code.

Suggested Environment

~] $ micromamba create -n ENV_NAME python=3.11.15 -c conda-forge bokeh dask distributed geopandas mgrs numpy pandas>=2.2.3 planetary-computer pystac-client rasterio scikit-learn scipy shapely stackstac tqdm xarray xgboost

or using the included environment.yml

~] $ micromamba create -n ENV_NAME environment.yml

Job Submission

ml_pipe.py

This script takes geospatial analysis indices, positive locations, and negative locations as inputs and then outputs all positive locations over the extent.

Field	Line	Use
INDEX_PATHS	13-19	Sets paths to analysis indices
DEM_PATH	22	Sets path to DEM file
POSITIVE_GEOJSON	25	Sets path to geojson of known positives
NEGATIVE_GEOJSON	26	Sets path to geojson of known negatives
OUT_DIR	29	Sets path to output directory
BUFFER_RADII_M	34	Sets buffer radii around known locations
CONFIDENCE_THRESHOLD	40	Sets threshold (0-1) for a location to be considered a positive
USE_RAW_INDICES	43	Toggles direct use of analysis indices
USE_TERRAIN	44	Toggles use of DEM-derived terrain variables
USE_INDEX_RATIOS	45	Toggles use of analysis index ratios
USE_TEXTURE	46	Toggles use of local variability in other predictors
USE_CONTEXT_ZSCORE	47	Toggles use of z-score comparred to surrounding area
TEXTURE_WINDOW_PX	49	Sets window for texture pixel groups (must be odd)
CONTEXT_WINDOW_PX	50	Sets neighborhood window for z-score context (must be odd)
TPI_WINDOW_PX	54	Sets context window for TRI (topographic position index) (must be odd)
RF_N_ESTIMATORS	57	Sets number of decision trees for random forest
RF_MAX_DEPTH	58	Sets max depth per tree in random forest
RF_MIN_SAMPLES_LEAF	59	Sets minimum number of pixels required for each tree to continue growing
RF_N_JOBS	60	Sets number of CPU cores to use for parallel tasks
XGB_N_ESTIMATORS	62	Sets number of boosting rounds for XGBoost
XGB_MAX_DEPTH	63	Sets depth of boosting rounds for XGBoost
XGB_LEARNING_RATE	64	Sets learning (error correction) rate for each subsequent boost round
XGB_SUBSAMPLE	65	Sets fraction of training pixels randomly sampled to grow each boost tree in XGBoost
XGB_N_JOBS	66	Sets number of parallel jobs for training/prediction
ENSEMBLE_WEIGHT_RF	69	Sets weight of RF in ensemble
ENSEMBLE_WEIGHT_XGB	70	Sets weight of XGBoost in ensemble
TEST_SPLIT	73	Sets fraction of samples held for evaluation
RANDOM_SEED	74	Sets random seed for entire ensemble
CLASS_WEIGHT	75	Toggles between methods of handling class imbalance
PREDICT_CHUNK_ROWS	78	Sets number of rows to predict at once
MIN_DETECTION_AREA_M2	83	Sets plausible minimum area on square meters of a positive location
POLYGON_SIMPLIFY_TOL	86	Sets polygon simplification in degrees
SAVE_CONFIDENCE_RASTER	89	Toggles whether the confidence raster is saved to the output directory
SAVE_POINT_GEOJSON	90	Toggles whether the geojson of calculated positive location centroids is saved to the output directory
SAVE_POLYGON_GEOJSON	91	Toggles whether the geojson of calculated positive location polygons is saved to the output directory
SAVE_MODEL	92	Toggles whether the model parameters and weights are saved to reuse without training
OVERWRITE	93	Toggles whether existing outputs are overwritten

Then,

~] $ python3 ml_pipe.py

For research use only. Not for clinical decisions. Copyright (c) 2026, Payton Jachec