notesum.ai

Published at November 21

Contrasting local and global modeling with machine learning and satellite data: A case study estimating tree canopy height in African savannas

cs.AI

cs.CV

Released Date: November 21, 2024

Authors: Esther Rolf¹, Lucia Gordon², Milind Tambe², Andrew Davies²

Aff.: ¹University of Colorado, Boulder; ²Harvard University

Arxiv: http://arxiv.org/abs/2411.14354v1

Refer to caption

XceptionS2
	data used		RMSE (m)		MAE (m)		$R^{2}$
model description	local	global	average	std dev.	average	std dev.	average	std dev.
global init.	✓	✓	1.98	0.03	1.48	0.02	0.37	0.02
random init. (no latlon)	✓	✗	1.94	0.03	1.47	0.02	0.40	0.02
random init.	✓	✗	2.05	0.04	1.55	0.03	0.33	0.03
U-Net
SSL init., decoder fine-tuned	✓	✓	2.05	0.04	1.54	0.03	0.32	0.03
SSL init., everything fine-tuned	✓	✓	2.08	0.13	1.55	0.10	0.30	0.09
random init., everything trained	✓	✗	2.27	0.69	1.59	0.20	0.10	0.69
Fully Convolutional Network
5-layers, 128 filters	✓	✗	1.64	0.01	1.20	0.01	0.57	$<$ 0.01
Pixelwise random forest
parameters vary by split	✓	✗	2.28	$<$ 0.01	1.80	$<$ 0.01	0.16	$<$ 0.01
Globally available TCH maps
GLAD (Potapov et al., 2021)	✗	✓	3.09	-	2.44	-	-0.53	-
ETH (Lang et al., 2023)	✗	✓	4.51	-	3.83	-	-2.26	-
Meta (Tolan et al., 2024)	✗	✓	3.51	-	2.87	-	-0.98	-
Pauls et al. (2024)	✗	✓	2.43	-	2.01	-	0.05	-