IIT Roorkee — Thesis (May 2025)
Author: Divyang Raj Verma
Supervisor: Emeritus Fellow Dr. P.K. Garg, Geospatial Engineering Group, Dept. of Civil Engineering
Full pipeline for classifying individual urban trees by species family using:
- Segmentation: YOLOv11l-seg trained on Google Maps Static imagery (Bangalore study area)
- Imagery: PlanetScope 8-band SuperDove multispectral imagery (3 m resolution, Bangalore AOI)
- Ground truth: Bengaluru Municipal Tree Census (~50,000 raw points → 14,224 retained after filtering out flowering plants and small canopies; 6,322 usable after polygon matching)
- Classifier: Random Forest with SMOTE oversampling and PCA dimensionality reduction
Best result: 4-class SMOTE + PCA
Train / Val / Test accuracy: 70.1 % / 65.1 % / 62.5 % — Cohen's Kappa: 0.80
Sample counts below are the final labeled set used for classification (6,322 total) after all the filtering and pre-processing steps.
| Family | Common Examples | Samples |
|---|---|---|
| Fabaceae | Indian Rosewood, Acacia, Tamarind, Albizia | 2,604 |
| Meliaceae | Neem, Mahogany | 964 |
| Arecaceae | Coconut Palm, Areca Palm | 933 |
| Bignoniaceae | Trumpet Tree, Jacaranda | 753 |
| Moraceae | Jackfruit, Peepal, Banyan | 385 |
| Combretaceae | Indian Almond, Arjuna | 354 |
| Anacardiaceae | Mango, Indian Hog Plum | 199 |
The 4-class subset uses the top four families (Fabaceae, Meliaceae, Arecaceae, Bignoniaceae; 5,254 samples). The 7-class subset adds Combretaceae, Moraceae, and Anacardiaceae.
01_data_acquisition/ Google Maps Static API → Maxar satellite PNGs
02_preprocessing/ PlanetScope UDM2 masking + vegetation indices
03_segmentation/ YOLO11l-seg training (Optuna) + inference
04_postprocessing/ IoU merge, confidence/area filters, convex-hull smoothing, ground truth merge
05_feature_extraction/ Zonal stats (mean/std/median) + GLCM texture
06_classification/ Dataset preprocessing, RF training, inference
07_analysis/ Confusion matrices, F1 charts, feature importance
| Group | Details | Count |
|---|---|---|
| Spectral zonal stats | 8 bands × 5 months × 3 stats (mean, std, median) | 120 |
| Index zonal stats | 6 indices × 5 months × 3 stats | 90 |
| GLCM texture | 2 bands (Red, NIR) × 3 properties (Dec 2024) | 6 |
| Temporal differences | 4 intervals × 14 features × 3 stats | 168 |
| Total | 384 |
Vegetation indices: NDVI, NDVI2, NDVI-RE (red-edge NDVI), EVI, SAVI, GNDVI
GLCM properties: Contrast, Homogeneity, Correlation
Date suffixes: May2024, Sept2024, Dec2024, Jan2025, Feb2025
git clone https://github.com/Divyang999/individual-tree-classification.git
cd individual-tree-classification
pip install -r requirements.txtGPU note: YOLO training requires CUDA. The classification pipeline runs on CPU.
The script builds the 185 m AOI grid itself (from an AOI vector file) and writes the grid centres CSV before downloading — there's no separate pre-built grid CSV to pass in.
export GOOGLE_API_KEY="your_key_here"
python 01_data_acquisition/download_google_images.py \
--aoi data/aoi_south_zone.shp \
--out data/satellite_images/ \
--grid-csv data/grid_centers.csv \
--zoom 19 --scale 2 --grid-size 185# For each date (May2024, Sept2024, Dec2024, Jan2025, Feb2025):
python 02_preprocessing/preprocess_planetscope.py \
--input data/rasters/May2024_raw.tif \
--udm2 data/rasters/May2024_udm2.tif \
--output data/rasters/May2024_8b.tif
python 02_preprocessing/compute_vegetation_indices.py \
--input data/rasters/May2024_8b.tif \
--outdir data/indices/May2024/Two-stage training (thesis Table 7): an Optuna hyperparameter search on the initial 300-image annotation set, then a transfer-learning fine-tune on the full 350-image set (slower LR, 10 frozen layers) that produces the final best.pt used for inference.
# Stage 1 — hyperparameter search (GPU server, ~100 trials)
python 03_segmentation/optuna_yolo.py
# Stage 2 — transfer-learning fine-tune on the +50-image set, from the
# Optuna checkpoint
python 03_segmentation/transfer_learning.py \
--checkpoint runs/segment/custom/best.pt \
--data data/roboflow_350img/data.yaml \
--epochs 16 --freeze 10
# Run inference to get tree crown polygons
python 03_segmentation/inference_yolo.py \
--model runs/segment/transfer/best.pt \
--images data/satellite_images/ \
--output data/raw_segments.csvpython 04_postprocessing/polygon_postprocess.py \
--input data/raw_segments.csv \
--output data/processed_segments.csv
python 04_postprocessing/merge_ground_segments.py \
--polygons data/processed_segments.csv \
--trees data/ground_truth_trees.csv \
--output data/merged_output.gpkg# Edit 05_feature_extraction/feature_config.yaml with your raster paths
python 05_feature_extraction/zonal_statistics_glcm.py \
--config 05_feature_extraction/feature_config.yamlpreprocess_dataset.py writes a whole directory of per-combination parquet
splits (not a single file) — that directory is what train_random_forest.py
takes as --input.
python 06_classification/preprocess_dataset.py \
--input data/zonal_stats.gpkg \
--outdir data/preprocessed/
python 06_classification/train_random_forest.py \
--input data/preprocessed/ \
--outdir model_outputs/python 06_classification/predict.py \
--input-gpkg data/zonal_stats.gpkg \
--prep-dir data/preprocessed/ \
--model-4 model_outputs/4_class_smote_pca \
--model-7 model_outputs/7_class_smote_pca \
--outdir inference/# Classification charts (confusion matrix, feature importance, F1 — Fig 27-32)
python 07_analysis/classification_results.py \
--input-dir model_outputs/ \
--outdir results/plots/classification/
# Segmentation training-curve charts (loss/precision/recall/mAP — Fig 24-25)
python 07_analysis/segmentation_results.py \
--results-csv runs/segment/transfer/results.csv \
--outdir results/plots/segmentation/| Dataset | Combination | Test Acc | Test F1 | Kappa |
|---|---|---|---|---|
| 4-class SMOTE | PCA | 0.625 | 0.687 | 0.801 |
| 4-class SMOTE | Spectral+Temporal+GLCM | 0.609 | 0.611 | 0.790 |
| 4-class SMOTE | Spectral+Temporal | 0.618 | 0.619 | 0.783 |
| 7-class SMOTE | PCA | 0.515 | 0.526 | 0.751 |
individual-tree-classification/
├── README.md
├── requirements.txt
├── 01_data_acquisition/
│ └── download_google_images.py
├── 02_preprocessing/
│ ├── preprocess_planetscope.py
│ └── compute_vegetation_indices.py
├── 03_segmentation/
│ ├── optuna_yolo.py
│ ├── transfer_learning.py
│ └── inference_yolo.py
├── 04_postprocessing/
│ ├── polygon_postprocess.py
│ └── merge_ground_segments.py
├── 05_feature_extraction/
│ ├── zonal_statistics_glcm.py
│ └── feature_config.yaml
├── 06_classification/
│ ├── preprocess_dataset.py
│ ├── train_random_forest.py
│ └── predict.py
└── 07_analysis/
├── classification_results.py
└── segmentation_results.py
- Satellite imagery: PlanetScope SuperDove PSB.SD (8-band, 3 m)
- Google basemap: Google Maps Static API (zoom 19, scale 2 → 0.30 m/pixel nominal, ~0.15 m/pixel effective)
- Ground truth: Bangalore Municipal Tree Census (OpenCity Portal), KGIS shapefiles
- Segmentation training data: Manually annotated via Roboflow — 10,287 trees across 350 images (300 initial + 50 added for transfer learning)
If you use this code, please cite:
Divyang Raj Verma, "Individual Tree Classification using ML and
High-Resolution Satellite Imagery," M.Tech Thesis,
IIT Roorkee, May 2025.