Learning Woody Clearing With Loss Alignment for Zero-Shot Regrowth and Woody Segmentation

Kal Backman; Jared Wood; Adam Roff

cs.CVAug 27, 2026

Learning Woody Clearing With Loss Alignment for Zero-Shot Regrowth and Woody Segmentation

Authors: Kal Backman, Jared Wood, Adam Roff

Organizations: New South Wales Department of Climate Change, Energy, the Environment and Water, Parramatta, 2150, NSW, Australia

Abstract

Detecting woody clearing is vital for managing biodiversity. Deep learning models can detect change in woody vegetation from bitemporal remote sensing imagery, however generated products may not meet end-user specifications due to unaligned loss definitions. Further limitations of deep learning models are the reliance on large datasets which can be difficult to attain for spatially rare and ambiguous events such as regrowth detection. In this work we train a model to detect woody change using bitemporal Sentinel-2 imagery consisting of 7 years' worth of annual imagery across the state of New South Wales, Australia. To align the objective of the model with end-user metrics, we introduce the loss scaling coefficient $α$ which transforms the objective to optimize for specific $F_β$ scores. Introducing $α$ was found to increase precision by 1.85x or recall by 1.12x. We propose input imagery augmentation and generation techniques that allow the woody change detection model to zero-shot transfer to regrowth and woody segmentation tasks. For woody segmentation, image generation techniques using activation maximization with low $α$ values for stability and image generation techniques derived from handcrafted features utilizing a mosaic of clearing patches and artificial trees for contextual grounding were found to outperform prior woody segmentation works of the study area, reducing the overall error by up to 18.2%. For zero-shot woody regrowth, creating pseudo-post and prior images resulted in the model achieving an F1 score of 0.845, creating a foundation for future regrowth detection work.

Figures & tables

Fig. 1: Overview of the woody change detection model architecture. The encoder individually extracts features from the prior and post images at different resolution scales using convolution-based bottleneck blocks. The extracted features are subsequently fused to encode woody change at each resolution scale. The lowest resolution feature is progressively upscaled and fused with the subsequent higher resolution features to produce woody change detection predictions at an identical resolution as the input imagery.

Fig. 2: Example predictions of the woody clearing model.

Fig. 3: Examples of the different post image generation techniques used to generate zero-shot woody predictions. Predictions were generated using a confidence threshold of 50% using a masked weighted average and 50% image overlap.

Fig. 4: Example images used to evaluate woody regrowth, where the pseudo-post image derives from 2018 Sentinel-2 imagery whilst the pseudo-prior image derives from 2024 Sentinel-2 imagery.

Confidence	Precision	Recall	F1	F2	IoU
Pixel-wise metrics
10%	0.289	0.830	0.429	0.604	0.273
25%	0.312	0.819	0.452	0.618	0.292
50%	0.352	0.797	0.488	0.636	0.323
75%	0.390	0.775	0.519	0.647	0.351
90%	0.409	0.763	0.533	0.651	0.363

TABLE I: Performance metrics for woody clearing detection on a pixel-wise and segment-wise basis

Distance	Precision	Recall	F1	F2	IoU
0m	0.352	0.797	0.488	0.636	0.323
10m	0.451	0.871	0.594	0.734	0.423
20m	0.478	0.891	0.622	0.760	0.452
30m	0.492	0.901	0.636	0.772	0.467
40m	0.503	0.909	0.648	0.783	0.479
50m	0.510	0.914	0.655	0.789	0.487

TABLE II: Performance metrics for woody clearing detection using distance thresholds

Fig. 5: Results from the woody clearing loss scaling experiment. Figures A) and B) plot the attained precision and recall metrics across confidence thresholds ranging from 5% to 95% against various $\alpha$ conditioned models where $\alpha$ ranges from $\frac{1}{10}$ to 10. Figure C) plots the $F_{\beta}$ scores attained for each $\alpha$ conditioned model where $\beta$ ranges from $\frac{1}{15}$ to 15. The green line indicates the maximum $F_{\beta}$ score for a specific $\beta$ value, highlighting the optimal $\alpha$ value which maximizes the particular $F_{\beta}$ score.

Fig. 6: Plots comparing the observed precision and recall metrics for the woody clearing detection task to that of the estimated precision and recall using equations 13 and 14

Model	Precision	Recall	F1	F2	IoU
Proposed approach	0.166	0.735	0.270	0.436	0.156
Global Forest Change	0.084	0.494	0.144	0.251	0.078

TABLE III: Woody clearing performance on MapBiomas dataset

Model	Precision	Recall	F1	IoU	OA
Zero image	0.829	0.000	0.000	0.000	0.723
Zero feature	0.666	0.058	0.106	0.056	0.731
Empty-scene	0.591	0.136	0.221	0.124	0.734
Empty scene + border	0.587	0.189	0.286	0.167	0.738
Empty scene + trees	0.684	0.375	0.485	0.320	0.779
Clearing scene	0.921	0.298	0.451	0.291	0.798

TABLE IV: Performance metrics on zero-shot woody segmentation

Model	Woody	Non-woody	Overall
Model	accuracy	accuracy	accuracy
Zero image	0.000	1.000	0.536
Zero feature	0.132	0.978	0.585
Empty-scene	0.220	0.959	0.616
Empty scene + border	0.356	0.910	0.653
Empty scene + trees	0.468	0.891	0.695

TABLE V: Zero-shot woody segmentation performance metrics on Fisher et al. point dataset

Model	Precision	Recall	F1	IoU	OA
Proposed approach	0.815	0.878	0.845	0.731	0.947
Segment Any Change	0.474	0.322	0.383	0.237	0.828
NDVI threshold	0.659	0.556	0.603	0.432	0.878

TABLE VI: Performance metrics on zero-shot woody regrowth

Fig. 7: Example woody clearing predictions of the proposed approach, Global Forest Change [ 57 ] and MapBiomas [ 49 ] overlayed onto the Sentinel-2 post image.

Fig. 8: Example woody segmentation predictions of the proposed approach and SamGeo [ 58 ] overlayed on Sentinel-2 imagery.

Fig. 9: Example woody regrowth predictions of the proposed approach, Segment Any Change [ 62 ] and NDVI difference thresholding overlayed onto the Sentinel-2 Pseudo-prior image.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

		$\alpha$
		$\frac{1}{10}$	$\frac{1}{8}$	$\frac{1}{6}$	$\frac{1}{4}$	$\frac{1}{3}$	$\frac{1}{2}$	$\frac{3}{4}$	1.00	$\frac{4}{3}$	2.0	3.0	4.0	6.0	8.0	10.0
Confidence	5%	0.706	0.673	0.640	0.562	0.503	0.423	0.331	0.286	0.240	0.182	0.135	0.107	0.077	0.063	0.052
	10%	0.719	0.687	0.652	0.570	0.510	0.429	0.337	0.291	0.245	0.186	0.139	0.110	0.080	0.066	0.055
	15%	0.728	0.698	0.662	0.580	0.520	0.439	0.345	0.299	0.252	0.192	0.144	0.114	0.084	0.069	0.058
	20%	0.737	0.708	0.671	0.588	0.528	0.447	0.353	0.306	0.258	0.198	0.148	0.118	0.087	0.072	0.061
	25%	0.745	0.718	0.682	0.599	0.540	0.458	0.364	0.316	0.268	0.206	0.155	0.124	0.092	0.076	0.065

Appendix

TABLE VII: Woody clearing precision

		$\alpha$
		$\frac{1}{10}$	$\frac{1}{8}$	$\frac{1}{6}$	$\frac{1}{4}$	$\frac{1}{3}$	$\frac{1}{2}$	$\frac{3}{4}$	1.00	$\frac{4}{3}$	2.0	3.0	4.0	6.0	8.0	10.0
Confidence	5%	0.421	0.450	0.502	0.596	0.653	0.720	0.785	0.816	0.842	0.876	0.904	0.920	0.942	0.949	0.955
	10%	0.409	0.438	0.491	0.590	0.648	0.716	0.782	0.813	0.839	0.874	0.902	0.918	0.941	0.947	0.953
	15%	0.402	0.430	0.483	0.584	0.643	0.712	0.778	0.809	0.836	0.871	0.900	0.916	0.939	0.945	0.952
	20%	0.396	0.423	0.476	0.578	0.638	0.707	0.774	0.806	0.833	0.868	0.897	0.914	0.937	0.944	0.950
	25%	0.390	0.416	0.469	0.572	0.631	0.701	0.769	0.801	0.829	0.864	0.894	0.911	0.935	0.941	0.948

Appendix

TABLE VIII: Woody clearing recall

Fig. 10: Plots comparing the observed precision and recall metrics for the zero-shot woody segmentation task to that of the estimated precision and recall. Observed values were taken using the clearing scene + trees post image generation technique

Fig. 11: Plots comparing the observed precision and recall metrics for the zero-shot woody regrowth detection task to that of the estimated precision and recall.

Fig. 12: Plots comparing the proposed loss and the Tversky loss for the observed precision and recall metrics in the woody clearing detection task.

Remote Sensing Change Detection

Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.

See all 89 papers

Mapping Woody Vegetation from Multi-Source Imagery and Prediction Fusion for Enhanced Data Efficiency and Accuracy
Aug 26, 2026Kal Backman, Jared Wood, Adam RoffRemote Sensing Image Segmentation Land Cover Classification
New South Wales Department of Climate Change, Energy, the Environment and Water, Parramatta, 2150, NSW, Australia · Monash University, Clayton, 3800, Vic, Australia · Earth Observation Lab, University of New England, Armidale, 2351, NSW, Australia +1
Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation
Jul 8, 2026Heeju Mun, Tackang Yang, Yunsoo Nam +1Self-Supervised Pre-Training Domain Generalization
Department of Landscape Architecture and Rural Systems Engineering, Seoul National University, Seoul, South Korea · Interdisciplinary Program in Landscape Architecture, Seoul National University, South Korea · Integrated Major in Smart City Global Convergence, Seoul National University, Seoul, Republic of Korea +1
Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping
May 7, 2026Gabriel Jeanson, David-Alexandre Duclos, William Larrivée-Hardy +5Synthetic Data Augmentation Remote Sensing Image Segmentation
1Northern Robotics Laboratory, Université Laval, Québec, QC, G1V 0A6, Canada · 2Département d’informatique et de génie logiciel, Université Laval, Québec, QC, G1V 0A6, Canada

Aug 26, 2026J/K move · Enter open · S save

Mapping Woody Vegetation from Multi-Source Imagery and Prediction Fusion for Enhanced Data Efficiency and Accuracy

Kal Backman, Jared Wood, Adam Roff

New South Wales Department of Climate Change, Energy, the Environment and Water, Parramatta, 2150, NSW, Australia · Monash University, Clayton, 3800, Vic, Australia · Earth Observation Lab, University of New England, Armidale, 2351, NSW, Australia+1

Remote Sensing Image Segmentation Land Cover Classification Multi-Source Image Fusion

Open paper AskPDF arXiv

Tree cover maps are a fundamental remote sensing product, used to derive ecological insights about the landscape and are essential to change detection, vegetation mapping and fire monitoring programs. However, comprehensive tree cover mapping requires reliable and high-quality imagery, free of cloud and weather defects to ensure accurate model outputs. Deep learning approaches can generate high quality maps with minimal human intervention but require large amounts of human annotated data to be successful. In this work we propose a framework consisting of methods that aim to improve the data efficiency and robustness of deep learning models using data fusion techniques to segment woody vegetation defined as vegetation over the height of 2m across the state of New South Wales, Australia. To improve robustness against varying image quality, we propose an image composition method that normalizes the imagery and removes defects, whilst also minimizing the reliance on individual image quality by proposing a prediction fusion method. The two methods resulted in an error reduction of 38.2% and 53.6% respectively compared to single-source imagery. To address deep learning approaches' limitation of requiring large amounts of data, we apply label transfer to multiple sources of imagery as a form of data augmentation to improve data efficiency. Learning from multiple image sources was shown to be the biggest improvement in performance, resulting in an error reduction between 28.1% to 76.2% across the different validation experiments, whilst reducing the standard deviation of performance across image dates by a factor of 13.

Abstract

Remote Sensing Change Detection

Explore similar work

Mapping Woody Vegetation from Multi-Source Imagery and Prediction Fusion for Enhanced Data Efficiency and Accuracy

Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation

Leveraging Image Generators to Address Training Data Scarcity: The Gen4Regen Dataset for Forest Regeneration Mapping