PaperlayerView exact revision on arXiv

Source-linked AI summary

Detection of Christmas tree plantations from high-resolution aerial imagery. A case study in the French Morvan

Francesca Razzano, Emanuele Dalsasso, Adrien Baysse-Lainé, Silvia Liberata Ullo, Gilda Schirinzi, Jocelyn Chanussot

arXiv:2608.27290v1cs.CV

TL;DR

Christmas tree plantation delineation is an underexplored remote-sensing problem because plantations combine distinctive structural and temporal characteristics with visual confusion, sparse labels, and severe class imbalance. The study develops a hard-negative-aware deep-learning segmentation framework and evaluates it through supervised testing, temporal transfer, and large-scale validation. DeepLabV3 with a ResNet-34 encoder achieves IoU 0.733 and F1-score 0.846 on the held-out 2020 test set, while HNM raises area under the precision-recall curve from 0.204 to 0.913.

  • Problem

    Christmas tree plantations lack dedicated remote-sensing mapping methods despite their economic and environmental relevance and their distinct structure, management cycles, and confusing visual contexts.

  • Method

    The study frames plantation mapping as rare-target semantic segmentation and combines hard-negative-aware dataset construction with a hybrid weighted BCE and Tversky objective.

  • Results

    DeepLabV3-R34 achieves IoU 0.733 and F1-score 0.846 on the held-out 2020 test set, while HNM increases area under the precision-recall curve from 0.204 to 0.913.

  • Takeaways & Limitations

    The framework transfers across 2017/2018 and 2023 imagery and supports large-scale detection of sparse plantation patches in predominantly negative landscapes.

Abstract

from arXiv · show

Christmas tree plantations are economically relevant, yet a largely unexplored application domain in Remote Sensing (RS). Their delineation is challenging because of high planting density, short rotation cycles, visual confusion with surrounding vegetation, the availability of dense labels for one reference year only, and severe class imbalance at the landscape scale. Although Deep Learning (DL) methods have shown strong potential for vegetation mapping, existing approaches are typically designed for forests, generic plantation systems, or orchards, and do not explicitly address the structural specificity and hard-negative confusion that characterize Christmas tree plantations. In response to these challenges, this work makes three main contributions: (i) it frames Christmas tree plantation mapping as a distinct rare-target semantic segmentation problem; (ii) it introduces a Hard Negative Mining (HNM) strategy to improve discrimination against confusing background patterns; and (iii) it evaluates the proposed framework across complementary levels, including supervised testing, temporal transfer, and large-scale validation. On the 2020 test set held out, the best model, DeepLabV3 with a ResNet-34 encoder, achieves an IoU of 0.733 and an F1-score of 0.846. HNM substantially improves precision-recall behavior, increasing the area under the precision-recall curve from 0.204 to 0.913. Temporal inference further shows meaningful transferability, reaching IoU/F1 values of 0.751/0.858 on 2017/2018 and 0.691/0.817 on 2023. Large-scale validation further highlights the intrinsic difficulty of the task, as Christmas tree plantations occupied only a very small fraction of the extent of the common evaluation, corresponding to 1,498.4 ha (1.72\%) in 2017/2018 and 1,782.2 ha (2.04\%) in 2023 out of 87,309.4 ha in total.

I. INTRODUCTION

Christmas tree plantations are an underexplored remote-sensing target whose distinctive structure, temporal dynamics, and visual similarity to other vegetation make dedicated mapping methods necessary. This study positions the task as rare-target semantic segmentation and motivates hard-negative-aware deep learning.

  • Research gap: Christmas tree farming lacks dedicated operational remote-sensing tools despite its economic, environmental, and public relevance.Existing approaches mainly address forests, generic plantation forestry, or fruit orchards.
  • Task specificity: Planting densities of 8,000–10,000 trees per hectare and 5–10-year harvest rotations create a distinctive structural and temporal signature.These characteristics differentiate Christmas tree plantations from orchards and forests.
  • Study focus: The study therefore investigates deep-learning semantic segmentation adapted to Christmas tree plantations as a dedicated Earth-observation task.The framing explicitly addresses the class’s structural specificity and confusion with visually similar backgrounds.
  • Related approaches: Prior research spans agricultural-plot segmentation, tree-crown delineation, multi-temporal vegetation monitoring, and imbalanced-target detection, but not this specific plantation class.Hard-example mining in related vegetation tasks provides a methodological rationale for the present adaptation.
  • Task specificity: Plantation appearance varies with management phase, planting density, row spacing, harvesting, canopy cover, and surrounding context.Examples range from bare-soil establishment to mature dense parcels and sparse canopies near grassland or forest borders.

III. MATERIAL FOR STUDY AREA

The Morvan Regional Natural Park is a relevant study area because it is France’s main Christmas-tree production hub, with concentrated but spatially heterogeneous plantations. Its terrain, orientation, seasonal operations, and land-use turnover create challenging remote-sensing conditions.

  • Study-area relevance: The Morvan supplies approximately one-third of national Christmas tree demand, making it an economically important production region.The area also experiences visible landscape effects from plantation expansion.
  • Physical setting: Plantation parcels concentrate between 560 and 680 m and favor moderate slopes and predominantly south-west, south, and west orientations.Their distribution reflects topographic, mechanization, and climatic constraints.
  • Spatial structure: The production system forms dense plantation hotspots alongside isolated peripheral parcels, complicating remote-sensing modeling.This spatial concentration produces heterogeneous target distributions across the region.
  • Temporal dynamics: Seasonal harvesting and shifting cultivation areas alter parcel appearance over time, while abandoned parcels can evolve into dense monospecific woodland.These transitions complicate separation from active plantations, young reforestation, and surrounding vegetation.

B. Data and layers for GT

The study combines multi-year high-resolution BD ORTHO imagery with a rigorously validated 2020 parcel inventory as supervised ground truth. The resulting data design supports training on 2020 labels and temporal-transfer analysis in 2017/2018 and 2023.

  • Spatial coverage: The dataset covers the Morvan production area across four French administrative units, with most data located in Nièvre.The study’s reference inventory is centered on municipalities where Christmas-tree cultivation is concentrated.
  • Temporal design: Only the 2020 acquisition is initially paired with rigorous vector ground truth for supervised training, validation, and first-stage evaluation.The other years are first used as inference targets for temporal transferability analysis.
  • Imagery: BD ORTHO mosaics provide 20 cm ground-sampling-distance RGB imagery for 2017/2018, 2020, and 2023.The common spatial resolution supports geometric consistency and visual comparison across years.
  • Ground truth: The supervised reference layer contains 1,253 manually annotated active Christmas tree plantation parcels from 2020.The annotations were designed to distinguish plantations from visually similar land covers.
  • Ground truth: Ground-truth construction combined parcel screening, boundary refinement, historical-image checks, visual filtering, field verification, and stakeholder discussions.This multi-stage protocol addressed ambiguity with clear-cuts, early reforestation, fallow parcels, and young forest plantations.

3) Ancillary Data and Temporal Support Layers:

Ancillary land-cover layers and a dedicated patch-construction workflow are used to preserve difficult negative contexts and address the scarcity of plantation pixels. The resulting hard-negative-aware dataset supports supervised and temporal segmentation evaluation.

  • Ancillary data: Ancillary layers describe water bodies, grasslands, and clear-cuts to identify difficult negative contexts resembling Christmas tree plantations.Temporal grassland products and clear-cut layers support inference-year validation and reduce confusion with harvested forest parcels.
  • Dataset construction: The dataset-construction pipeline increases verified plantation representation, preserves difficult non-target contexts, and avoids spatial leakage.These design goals respond to the small landscape fraction occupied by plantations and the need for reliable evaluation.
  • Workflow: The workflow constructs a hard-negative-aware dataset from RGB orthophotos, parcel annotations, and ancillary layers before training and evaluation.It supports both the 2020 supervised test and temporal-transfer analysis.
  • Patch preparation: Original 5 × 5 km orthophoto tiles are subdivided into non-overlapping 512 × 512-pixel patches covering about 102.4 × 102.4 m.The subdivision makes the imagery compatible with segmentation-model processing.
  • Hard negatives: Semantically meaningful hard negatives are encoded in masks during dataset construction rather than treated as random background.This preserves difficult non-target examples for subsequent model training.

2) Mask Integration and Label Priority:

Mask generation resolves overlaps between verified Christmas tree parcels and ancillary land-cover layers through explicit label-priority rules. Sampling then emphasizes representative plantations and visually confusing non-target contexts.

  • Label priority: Christmas tree parcels receive priority over grassland, clear-cut, and water labels when overlapping layers are fused into multi-class masks.Remaining ancillary labels are added sequentially only where no higher-priority label has been assigned.
  • Overlap assessment: 68.7% of verified parcel area matched plantations before filtering, while overlaps reached 28.2% with grassland and 3.9% with clear-cuts.These raw intersections indicate that direct layer fusion would create nonnegligible label ambiguity.
  • Sampling: Positive patches contained at least 30% plantation pixels, with positive samples capped at Pcap = 150 patches per tile.The threshold targeted representative plantation patterns and limited over-representation of highly productive areas.
  • Sampling: Negative patches contained no Christmas tree pixels and at least 25% grassland or clear-cut area, concentrating sampling on confusing backgrounds.Water remained a separate class, with a limited number of water-containing patches also retained.

A. Model Training Setup

Training addresses limited data, class imbalance, and visual confusion through augmentation and a HardNegativeLoss combining weighted pixel discrimination with overlap-focused optimization.

  • Data augmentation: RandomAugment applies flips, 90° rotations, brightness and contrast jitter, and Gaussian noise during training to reduce overfitting and improve invariance.Geometric transformations are applied jointly to images and masks, while radiometric perturbations affect only images.
  • Loss function: HardNegativeLoss combines weighted Binary Cross-Entropy with Tversky loss to address class imbalance and confusion with similar land covers.The objective emphasizes pixel-wise discrimination, overlap consistency, and minority-class recall.
  • Loss function: The loss uses λbce = 0.4 and λtversky = 0.6, balancing the two objective terms empirically.The weighted BCE component incorporates a spatially explicit weight map to guide discrimination against hard-negative classes.
  • Loss function: Pixel weights are 1.0 for background, 4.5 for Christmas Tree, 2.0 for Grassland, 2.5 for Clear-cut, and 6.0 for Water.Higher weights penalize errors on the target and difficult non-target regions more strongly than generic-background errors.
  • Loss function: Tversky parameters α = 0.2 and β = 0.8 penalize false negatives more strongly than false positives.This configuration favors high recall and limits missed detections of small or isolated plantation parcels.
  • Loss function: Combining weighted BCE and Tversky loss targets both local hard-negative discrimination and global structural coherence in predicted masks.

3) Training Protocol:

Training uses mixed-precision optimization and validation-IoU model selection to reduce computational demands while retaining the best-performing model weights.

  • Training protocol: 20 epochs maximum and torch.amp mixed-precision computation reduce memory usage and accelerate processing during training.Each iteration computes HardNegativeLoss from logits, masks, and pixel-wise weight maps before backpropagation updates the parameters.
  • Training protocol: The checkpoint with the highest validation IoU is selected after monitoring IoU at each epoch.

B. Model Architecture and Hyperparameters

The study compares several semantic-segmentation architectures for RGB aerial imagery and tunes their operating thresholds and evaluation metrics for severe class imbalance.

  • Model architecture: DeepLabV3 with a ResNet34 encoder is used as the primary model and compared with U-Net, DeepLabV3-R18, DeepLabV3-R50, FPN-R34, and U-Net++-R34.
  • Model architecture: All architectures process three-channel RGB imagery and output a single-channel binary plantation mask.
  • Model architecture: The custom U-Net uses an encoder-decoder with skip connections, 16 initial filters doubling during downsampling, and dropout p = 0.3.Dropout is applied within convolutional blocks to reduce overfitting while preserving spatial detail.
  • Hyperparameters: Adam optimization starts at 1×10^-4, while ReduceLROnPlateau halves the learning rate after three consecutive epochs without validation-IoU improvement.
  • Threshold selection: Validation thresholds from [0.05, 0.95] are searched separately for each architecture, selecting the value that maximizes validation F1-score.This avoids deriving binary predictions from an arbitrary cutoff.
  • Evaluation: Primary evaluation relies on global pixel-level IoU and F1-score because they reflect segmentation quality under strong plantation-background imbalance.
  • Evaluation: Precision and recall characterize false-positive and false-negative trade-offs, while F1-score summarizes them through their harmonic mean.
  • Evaluation: Large-scale inference additionally reports specificity and balanced accuracy, with balanced accuracy averaging recall and specificity.Global confusion matrices are accumulated across patches and millions of pixels.

V. RESULTS

The study benchmarks semantic segmentation architectures on the held-out 2020 test set and selects DeepLabV3-R34 as the reference model. Qualitative comparisons show differences in boundary coherence and robustness to confusing backgrounds.

  • 1) Selection of the semantic segmentation model:: 258 samples provide the held-out 2020 test bed for quantitative and qualitative comparison because dense, reliable GT masks are available.The test set supports model comparison and analysis of plantation delineations.
  • Temporal transfer and validation:: Temporal analysis extends inference to 2017/2018 and 2023, using manual visual annotation to enable quantitative and qualitative assessment without initial pixel-level GT.The analysis first uses spatial support corresponding to the 2020 test set.
  • 1) Selection of the semantic segmentation model:: DeepLabV3-R34 achieves the highest IoU (0.733) and F1-score (0.846) among the evaluated architectures.It also reaches precision 0.866 and recall 0.827.
  • 1) Selection of the semantic segmentation model:: U-Net++-R34 follows closely with IoU 0.731 and F1 0.845, while DeepLabV3-R18 remains competitive but slightly below.The ViT-based U-Net baseline has substantially lower overlap and recall in this setting.
  • 1) Selection of the semantic segmentation model:: DeepLabV3-R34 is selected as the reference model for subsequent experiments after producing the most stable delineations with fewer spurious detections.The qualitative comparison covers RGB patches, ground-truth masks, and predictions from six architectures.

2) Ablation on the training objective:

The training-objective ablation compares loss formulations on DeepLabV3-R34, while HNM is examined through controlled visual and threshold-free comparisons. The experiments emphasize precision–recall control under severe class imbalance and confusing non-target backgrounds.

  • 2) Ablation on the training objective:: The hybrid weighted BCE + Tversky formulation shows a consistent advantage over Tversky and weighted BCE alone.The comparison uses the same architecture, multiple seeds, and a fixed test threshold of 0.70.
  • 2) Ablation on the training objective:: Tversky loss alone is strongly recall-oriented but yields the weakest overall overlap and precision in the ambiguous segmentation setting.The result indicates that overlap-driven optimization alone does not control false positives sufficiently.
  • 3) Assessment of the proposed HNM strategy:: HNM suppresses diffuse false activations and produces more spatially coherent plantation responses in 2020 test patches.The visual comparison contrasts DeepLabV3-R34 trained with and without HNM using shared probability scales.
  • 3) Assessment of the proposed HNM strategy:: HNM reduces false positives to 1,648,335 and yields precision = 0.866, recall = 0.827, F1 = 0.846, and IoU = 0.733.The strategy combines selective hard-negative sampling with BCE weighting based on similarity to challenging negative classes.
  • Temporal transfer and annotation refinement:: Targeted refinement corrects proxy-label errors, including missing plantations and boundary or semantic mismatches, while preserving unchanged areas.The workflow is applied to 2023 and corrected 2017/2018 annotations for evaluation.

B. Temporal Inference on 2017/2018 and 2023

The framework was evaluated across fixed test locations and a common large-scale extent for 2017/2018 and 2023, using refined ground truth where available. Temporal results remained similar to 2020, while large-scale validation exposed severe imbalance and residual false positives.

  • Temporal evaluation: 2017/2018 and 2023 imagery were evaluated on the same 2020 test patch locations to isolate temporal effects while keeping the footprint fixed.The analysis considered radiometry, phenology, acquisition conditions, and landscape evolution.
  • Ground-truth refinement: Corrected masks were stored separately, with metrics computed against refined ground truth when available and proxy labels otherwise.The refinement added missing parcels and removed falsely labeled parcels without overwriting original annotations.
  • Evaluation scope: 2017/2018 evaluation was constrained by unavailable or invalid imagery, making its effective test area smaller than the 2023 evaluation.Some 2020 test patches could not be assessed for 2017/2018 because source imagery did not cover all locations.
  • Temporal results: The refined-GT metrics remained in a similar range to 2020, supporting temporal generalization of the HNM pipeline.Table V reports temporal inference results on fixed 2020 test patch locations.
  • Large-scale validation: 87,309.4 ha formed the common large-scale extent, with 2023 plantations covering 1,782.2 ha (2.04%) and background comprising 97.96%.The resulting background-to-positive area ratio was approximately 48:1.
  • Large-scale validation: 2023 achieved precision = 0.653, recall = 0.736, F1-score = 0.692, and IoU = 0.529 after post-processing.False-positive and false-negative areas were 695.81 ha and 471.39 ha, respectively.
  • Large-scale validation: 2017/2018 achieved precision = 0.597, recall = 0.879, F1-score = 0.711, and IoU = 0.551 under the same post-processing configuration.The false-positive area was 889.80 ha and the false-negative area was 181.72 ha.
  • Qualitative assessment: The wall-to-wall maps located isolated plantations beyond main production clusters, while residual false positives occurred mainly over disturbed forest patches and herbaceous clearings.The detailed overlays showed predicted masks closely adhering to active parcel boundaries.

VI. CONCLUSION AND DISCUSSION

The study frames Christmas tree plantation delineation as a distinct rare-target segmentation problem and designs the framework around class imbalance and hard-negative confusion. The resulting pipeline transfers across acquisition periods, while broader geographic testing and object-aware formulations remain open directions.

  • Conclusion: Christmas tree plantation delineation requires a dedicated remote-sensing task formulation because targets are sparse and visually similar to managed or transitional land covers.The conclusion distinguishes the task from straightforward forest or orchard mapping.
  • Conclusion: Hard Negative Mining substantially improves rejection of visually plausible non-target patterns and is a central component of the operational pipeline.The conclusion identifies HNM as more than a secondary refinement.
  • Conclusion: DeepLabV3 with a ResNet-34 backbone provided the strongest compromise on the 2020 benchmark and robust behavior on 2023 and 2017/2018 imagery.The conclusion links performance to contextual representation, discrimination capacity, and post-processing design rather than architectural complexity alone.
  • Future work: Object-aware or instance-level formulations could reduce residual fragmentation and improve parcel-level delineation in visually ambiguous areas.This is identified as an open direction rather than an evaluated result.
  • Future work: Broader transfer experiments across production regions are needed to assess the framework’s geographical generalizability.The current conclusion identifies geographic coverage as an essential future evaluation boundary.
Loading 2608.27290v1…