raflisbk/FETA-fasterrcnn-6band-canopy

🌳 Faster R-CNN 6-Band Canopy Detection

Faster R-CNN ResNet50-FPN v2 trained on 6-band multispectral satellite imagery for tree canopy detection.

Model Details

Property	Value
Backbone	ResNet50-FPN v2 (COCO pretrained)
Input	6 channels (B, G, R, RE, NIR1, NIR2)
Classes	1 (canopy)
Image size	640×640
Best mAP@50	0.9735
Parameters	~43M

Auto-Detect Band Support

Input Format	Band Mapping
3-band (RGB)	[R, G, B, 0, 0, 0]
5-band	[B, G, R, RE, NIR1, 0]
6-band	Direct input
7-band	Drop Band 7 (anomalous NIR3)

Usage

import rasterio, numpy as np, torch
from torchvision.models.detection import fasterrcnn_resnet50_fpn_v2

# Load model (modify first conv for 6 channels)
model = ...  # See training notebook for full setup
model.load_state_dict(torch.load('best.pt'))
model.eval()

# Load & preprocess
with rasterio.open('tile.tif') as src:
    img = src.read()  # (C, H, W)
# Map to 6 channels, normalize, run inference