PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Kal Backman, Jared Wood, Adam Roff

arXiv:2608.26489v1cs.CV

TL;DR

The paper tackles end-user metric misalignment and limited training data for rare, ambiguous woody-change tasks. It trains a bitemporal Sentinel-2 woody-clearing model with α loss scaling and uses generated or augmented imagery for zero-shot transfer. The approach increases woody-clearing precision or recall, improves woody segmentation against prior maps, and reaches an F1 score of 0.845 for zero-shot regrowth.

  • Problem

    Deep-learning change detection may require large datasets unavailable for rare, visually subtle events such as woody regrowth, while standard losses may not align outputs with end-user accuracy priorities.

  • Method

    The paper trains a CNN encoder-decoder on bitemporal Sentinel-2 imagery, introduces α loss scaling to target precision or recall, and generates or augments imagery for zero-shot transfer.

  • Results

    The model achieved 1.85× higher precision or 1.12× higher recall, reduced woody-segmentation error by 18.2%, and attained an F1 score of 0.845 for zero-shot regrowth.

  • Takeaways & Limitations

    The work provides a foundation for zero-shot woody segmentation and future woody-regrowth detection when curated regrowth datasets are difficult to establish.

  • Takeaways & Limitations

    The work is geographically constrained to New South Wales, and zero-shot regrowth and segmentation were evaluated on less than 1% of the woody-clearing samples.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper addresses woody-change monitoring by aligning deep-learning objectives with end-user metrics and transferring a woody-clearing model to data-scarce regrowth and segmentation tasks. It develops this approach using bitemporal Sentinel-2 imagery and a large New South Wales dataset.

  • Motivation: Land clearing is identified as the leading cause of terrestrial biodiversity loss, motivating timely and accurate monitoring data.Such data are needed to support informed biodiversity-protection decisions.
  • Data limitations: Deep-learning change detection requires large training datasets, which are especially difficult to obtain for rare events.Woody regrowth is visually subtle and continuous over several years, and comprehensive consistent mapping remains limited.
  • Zero-shot transfer: Zero-shot learning transfers trained models to unseen tasks without additional training or data, leveraging related datasets such as woody clearing.The paper applies this strategy to woody regrowth detection, where training data are insufficient.
  • Proposed approach: The proposed model uses bitemporal, geospatially aligned Sentinel-2 imagery to detect woody vegetation clearing and introduces α to target end-user Fβ objectives.Woody vegetation is defined as vegetation over 2m, and the objective prioritizes recall for large β values.
  • Proposed approach: The model is a CNN twin encoder-decoder that outputs a mask of woody vegetation absent in the post image relative to the prior image.Its encoder extracts multiscale features, decoder fuses temporal and spatial information, and head produces pixel-level clearing confidence scores.
  • Study scope and contributions: The study trains and validates the woody-clearing network on 7 years of imagery covering 801,137 km2 and 54 billion data points.The contributions also include imagery augmentation and generation for zero-shot transfer to regrowth and woody segmentation.

1) Encoder:

The model uses a Siamese encoder-decoder to extract and fuse multiscale features from aligned prior and post images for woody-change prediction. Its tunable loss coefficient α shifts optimization toward recall or precision to better match target Fβ objectives.

  • Encoder:: A Siamese convolutional encoder independently extracts five multiscale feature maps from the spatially aligned prior and post images.The feature maps use identical backbones and progressively lower resolutions.
  • Decoder and head:: The decoder fuses bitemporal features across spatial scales, progressively upsamples them, and produces predictions at the input-image resolution.A convolutional head applies sigmoid activation to generate per-pixel woody-clearing confidence scores.
  • Loss function:: The loss combines binary cross-entropy and dice terms to address extreme class imbalance in woody-clearing segmentation.Binary cross-entropy is biased toward the majority class, whereas dice loss is more robust to imbalance but can over-focus on erroneous labels in sparse-positive batches.
  • Loss function:: α is introduced as a tunable parameter so values above 1 favor recall and values below 1 favor precision.Recall-favoring terms are multiplied by α, while precision-favoring terms are multiplied by 1/α.

C. Data

The dataset uses seven years of annual, spatially aligned Sentinel-2 imagery across mainland New South Wales with manually validated woody-clearing labels. Training and validation tiles are highly imbalanced, and plantation clearing is encoded as non-clearing.

  • Data: Seven years of annual Sentinel-2 imagery from 2018 to 2024 cover mainland New South Wales, Australia.The imagery uses 10m data across a landmass of 801,137 km2.
  • Data: The imagery is organized into 109 roughly 100km × 100km scenes with radiometric correction, cloud masking, and annually selected prior-post image pairs.Each era’s prior image is the previous era’s post image.
  • Label generation: Woody-clearing regions are proposed using a pixel-based clearing regression index and then manually checked and independently validated by operators.The proposal model emphasizes recall to avoid excluding clearing events.
  • Label limitation: Pine-plantation clearing is encoded as non-clearing because it was not required to be reported.This creates erroneous zero labels for plantation-clearing pixels.
  • Dataset split: 284,215 tiles are used for training and 56,851 for validation, with woody clearing occupying 0.05% of all pixels.Labels from 2018-2019 through 2022-2023 are used for training, while 2023-2024 is reserved for validation.

III. EXPERIMENT DESIGN

The experiments evaluate woody-clearing segmentation, loss-scaling behavior, and zero-shot transfer using augmented bitemporal Sentinel-2 tiles. Evaluation includes pixel, segment, and spatial-distance metrics on held-out scenes.

  • III. EXPERIMENT DESIGN: The experiments cover woody-clearing segmentation, loss-scaling effects, and zero-shot woody-regrowth and woody-segmentation tasks.The latter tasks are assessed without further training on their target labels.
  • Training procedure: Training samples pair prior and post images with labels, while half of each batch is drawn from tiles known to contain woody clearing.Random appearance, noise, and geometric augmentations are applied to sampled tiles.
  • Inference: Full-scene predictions use 1024w × 1024h patches with 50% overlap and weighted fusion that reduces the influence of patch borders.The model processes these patches with batch size 32.
  • Evaluation metrics: Performance is evaluated with precision, recall, F1-score, F2-score, and intersection over union.True positives, false positives, and false negatives are defined from thresholded predictions and ground-truth labels.
  • Spatial evaluation: Segment analysis evaluates overlap between predicted and ground-truth clearing patches, while distance analysis evaluates spatial proximity between predicted and reference pixels.Both analyses redefine TP, FP, and FN using segment overlap or distance thresholds.

B. Woody clearing loss scaling

The loss-scaling experiment fine-tunes models across multiple α values and evaluates their precision-recall behavior using generalized Fβ scores. Curve fitting provides estimates of precision and recall as functions of α.

  • B. Woody clearing loss scaling: Fifteen models are fine-tuned with α values ranging from 0.1 to 10 using an additional 100,000 optimization iterations.The remaining training methodology is kept identical to the base procedure.
  • Evaluation: Each α-conditioned model is evaluated on the 2023-2024 era at a 50% confidence threshold.True positives, false positives, and false negatives are computed using the same evaluation procedure.
  • Fβ analysis: Fβ generalizes F1 and F2 by controlling the relative weighting of precision and recall.β > 1 favors recall, whereas β < 1 favors precision.
  • α selection: A fitted curve estimates precision and recall as functions of α to guide selection of α for a target metric.The precision and recall models use a linearized α transformation and offsets based on the default α = 1 model.

C. Woody clearing comparison

The study evaluates zero-shot transfer from woody clearing detection to woody segmentation by generating post images through handcrafted and activation-maximization techniques. It benchmarks nine post-image generation methods and uses masked weighted averaging with image overlap for prediction generation.

  • Nine post-image generation techniques were benchmarked for zero-shot woody segmentation.The methods included zero and empty-scene inputs, clearing mosaics, contextual augmentations, woody-vegetation patch insertion, and activation maximization.
  • Handcrafted generation methods require task- and dataset-specific understanding to construct suitable post images.Activation maximization was introduced as a more generalizable alternative that optimizes the input image while keeping model weights fixed.
  • Predictions used a 50% confidence threshold, masked weighted averaging, and 50% image overlap.The masking covered contextual borders and pasted woody vegetation within generated post images.
  • Activation maximization searches for a post image that maximizes the model’s output prediction.The generated image is optimized as an input parameter array rather than updating model weights.

1) Zero-shot woody segmentation evaluation:

The zero-shot woody segmentation evaluation uses labeled vegetation datasets and compares generated-image strategies with SamGeo and a prior labeled-point benchmark. Performance is measured with overall accuracy because the segmentation data are more class-balanced than the clearing task.

  • The woody vegetation evaluation dataset covered 108 km2 across six spatially diverse New South Wales scenes.More than 1M pixels were labeled, with 28% classified as woody vegetation.
  • Predictions from all nine generation methods were produced with 256w × 256h inputs and masked weighted averaging.A 50% image overlap was used, including for methods that inserted prior-image borders.
  • Overall accuracy replaced F2 for woody segmentation evaluation.The segmentation dataset had a more even class balance, and the task did not focus on maximizing recall.
  • SamGeo was evaluated as a prior-work comparison using the text prompt “Woody vegetation.”Its pipeline combines GroundingDino for zero-shot visual grounding with SAM for object segmentation.
  • A labeled-point benchmark contained 6,648 stratified-random points across New South Wales.The points were manually classified using 0.5m-resolution Leica ADS40 imagery captured in 2011.

E. Zero-shot woody regrowth detection

The model was transferred zero-shot to woody regrowth by swapping bitemporal inputs to form pseudo-prior and pseudo-post images. Evaluation used multi-year Sentinel-2 pairs and compared the approach with Segment Any Change and an NDVI-difference baseline.

  • Woody regrowth was defined as present vegetation over 2m resulting from past clearing events.The model extracted regrowth information by swapping the original prior and post images.
  • Pseudo-prior images came from 2024 Sentinel-2 imagery, while pseudo-post images came from 2018 imagery.The longer interval was used because regrowth is visually subtle and develops continuously over several years.
  • The regrowth dataset contained over 490,000 labeled pixels across 49 km2, with 16% classified as woody regrowth.Five scenes across New South Wales were proposed, and labeling focused on regions flagged by the model.
  • Segment Any Change was evaluated as a prior-work comparison for zero-shot woody regrowth segmentation.It uses SAM latent vectors from bitemporal image pairs to detect semantic changes.
  • An NDVI-difference threshold baseline was used to estimate reliance on individual-pixel spectral values.Thresholds from 0.0 to 2.0 in 0.1 increments were selected using held-out image pairs.

B. Woody clearing loss scaling

The loss scaling experiment tests whether α can shift woody clearing models toward precision, recall, or selected Fβ objectives. The resulting curves are compared with equations that estimate precision and recall from an α = 1 validation point.

  • Increasing α raises recall at the cost of precision, while decreasing α boosts precision at the expense of recall.Optimizing α can attain greater overall precision or recall than traditional confidence thresholding alone.
  • For a target Fβ score, an optimal α value exists, and larger β values favor models trained with larger α values.The experiment evaluated β values from 1/15 to 15 and identified optimal α values with the green curve.
  • The α-to-metric relationship is calibrated from initial precision and recall at α = 1.Equations 13 and 14 estimate α′ values for achieving specific precision or recall targets.
  • Estimated precision and recall closely followed observed values, with average errors of 0.025 and 0.016 respectively.Most error came from lower α′ estimates.

D. Zero-shot woody segmentation

Zero-shot woody segmentation transferred a woody-clearing model to segmentation by generating post images that expose contextual woody features. Handcrafted clearing-scene and tree generation performed strongly, including against supervised and alternative baselines.

  • Post-image generation: F1 0.899 was achieved by the clearing scene + trees post-image generation method.Handcrafted methods using real reference scenes outperformed those using zeroed inputs.
  • Post-image generation: A clearing scene plus randomly placed trees doubled the base clearing-scene F1 performance.
  • Post-image generation: 1.6 percentage points higher overall accuracy was achieved by the best handcrafted method than by activation maximization.
  • Comparisons: 1.3× higher F1 was achieved by the best zero-shot configuration compared with SamGeo’s 0.685 F1 score.
  • Fisher et al. dataset: 89.2% overall accuracy with clearing scene + trees reduced error by 18.2% relative to the Fisher et al. model’s 86.8%.Activation maximization reached 87.3% overall accuracy and reduced error by 3.8%.
  • Zero-shot woody regrowth: At a 50% confidence threshold, zero-shot woody regrowth reached an F1 score of 0.845 and overall accuracy of 94.7%.This exceeded Segment Any Change and NDVI difference thresholding, which achieved F1 scores of 0.383 and 0.603, respectively.

B. α loss scaling

The α loss scaling coefficient conditions binary cross-entropy and dice loss to target end-user precision–recall objectives. It improved metric control beyond confidence thresholding, but α must be calibrated for each task and dataset.

  • Loss alignment: α loss scaling conditions binary cross-entropy and dice loss to target specific Fβ scores and end-user metrics.
  • Metric control: Confidence thresholding alone was ineffective for fine-tuning precision and recall because model predictions were overconfident.Extreme thresholds produced only 2.9% and 0.9% increases over less extreme thresholds for precision and recall, respectively.
  • Metric control: 1.85× higher precision or 1.12× higher recall was achieved with α loss scaling than with confidence thresholding alone.Precision increased from 0.359 to 0.785, while recall increased from 0.778 to 0.937 at a 50% confidence threshold.
  • Zero-shot transfer: A 1.32× F1 increase was obtained for zero-shot woody segmentation by using α = 1/10 with activation maximization.The lower-α model counteracted activation maximization’s high-recall tendency by prioritizing precision.
  • Calibration: The appropriate α depends on the deployed task and dataset and requires validation performance at α = 1 for calibration.The estimated value can then be used to fine-tune the α = 1 model toward the intended metric.
  • Zero-shot transfer: Handcrafted post-image construction enabled strong zero-shot segmentation but depends on users creating artificial features that expose latent model information.Activation maximization avoids this dependence but reduces zero-shot performance.

E. Model comparison

The proposed woody-change model transfers zero-shot to segmentation and regrowth, while comparisons reveal strengths on small features and limitations in broader forested regions, geography, evaluation size, and supervised performance.

  • Woody clearing: The proposed approach detected small, thin clearing areas from roads and paths that prior methods often missed, but performance degraded around cloud cover.It flagged pixels associated with clearing operations rather than broadly proposing large continuous regions.
  • Woody segmentation: The proposed zero-shot segmentation avoided SamGeo’s tendency to favor dense forest regions while ignoring individual tree crowns.The comparison used predictions over Sentinel-2 imagery.
  • Woody regrowth: The proposed zero-shot regrowth predictions were compared with Segment Any Change and NDVI difference thresholding, whose outputs favored macro-regions or contained salt-and-pepper noise.Segment Any Change over-segmented large continuous regions, while NDVI thresholding lacked spatial information pooling.
  • Limitations: The work is geographically constrained to New South Wales, and zero-shot evaluation datasets contained less than 1% of the woody-clearing evaluation samples.Further evaluation across a greater global context and larger zero-shot datasets is required.
  • Limitations: The model struggled with large continuous forested regions because of limited contextual information, and zero-shot models are expected to underperform comparable supervised models.The paper recommends zero-shot transfer mainly as a supplementary option when sufficient labels are infeasible.
  • Zero-shot transfer: Input image generation enabled zero-shot woody segmentation without additional training; handcrafted mosaics with artificial trees and activation maximization outperformed prior work by reducing error by 18.2% and 3.8%, respectively.The handcrafted technique provided contextual grounding, while activation maximization performed comparably to the best handcrafted technique.
  • Zero-shot transfer: The zero-shot regrowth transfer attained an F1 score of 0.845, providing a foundation for future regrowth detection work where labeled data are difficult to acquire.The difficulty reflects regrowth’s rarity and ambiguity.

APPENDIX A WOODY CLEARING α-CONFIDENCE TABLES

The appendix compares precision and recall behavior across confidence thresholds and loss functions, showing accurate metric estimation and stronger precision from the proposed loss at lower α′ values.

  • α-confidence tables: Precision and recall scores were reported for models conditioned on α across confidence thresholds from 5% to 95%.The corresponding results appear in Tables VII and VIII.
  • Metric estimation: Across all three tasks, average estimation error was 0.028 for precision and 0.040 for recall.For segmentation, errors were 0.017 and 0.079; for regrowth, they were 0.043 and 0.025.
  • Metric estimation: Woody segmentation recall sharply decreased at α′ values below -2 because negative values favor precision over recall and reduce susceptibility to the input-generation techniques.Most observed recall-estimation errors originated from lower α′ values.
  • Loss comparison: The Tversky loss uses φ and γ to scale false positives and false negatives, whereas the proposed loss uses a single α parameter to control precision–recall preference.Higher φ favors precision, while higher γ favors recall; varying either parameter affects both metrics.
  • Loss comparison: The proposed loss achieved marginally greater recall and significantly greater precision than Tversky loss at lower α′ values.The comparison was performed using additional models trained with Tversky loss and constrained parameters.
Loading 2608.26489v1…