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

cs.CVAug 27, 2026

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

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

Organizations: Engineering Department, University of Naples Parthenope, Naples, Italy · INRIA, Grenoble, France · CNRS Délégation Alpes: Grenoble, Auvergne-Rhône-Alpes, France · Engineering Department, University of Sannio, Benevento, Italy

Abstract

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.

Figures & tables

Figure 1: Examples of the high visual and structural variability of Christmas tree plantations in high-resolution RGB imagery. Despite belonging to the same target class, parcels exhibit drastic differences depending on their management phase and spatial context: (a) newly established plantation on bare soil; (b) variable density and row spacing across different sub-plots; (c) mature and densely planted parcel with strong linear patterns; (d) irregular harvesting and clearing dynamics; (e) sparse canopy cover with grassy background; (f) early-stage planting adjacent to forest borders.

Figure 2: Study area in the Morvan Regional Natural Park, France. The figure shows the geographical context of the Morvan area within France, the extent of the protected geographical indication “Christmas tree of the Morvan”, the parcel-level inventory of Christmas tree plantations mapped in 2020 (red), and the spatial extent of the orthophoto tiles used in this study (blue dashed rectangle).

Data	Year(s)	Resolution	Source	Reference	Role in workflow
BD ORTHO orthophotos	2017/2018, 2020, 2023	0.2 m	IGN	[ 37 ]	Main RGB imagery for supervised training (2020), temporal inference (2017/2018 and 2023), and visual interpretation.
Manual parcel annotations	2020	Parcel level	Based on expert photo-interpretation and field verification	[ 8 ]	Positive reference layer for Christmas tree plantations; used for training, validation, and test construction.
OSO land-cover map	2018, 2020	10 m	THEIA	[ 38 , 39 ]	Water masking for 2017/2018 and 2020; exclusion of obvious non-target areas.
CLC+ Backbone	2023	10 m	Copernicus Land Monitoring Service	[ 40 ]	Water masking for the 2023 temporal inference dataset.
Natural grassland dataset	2020	10 m	Panhelleux et al.	[ 31 ]	Grassland-related hard negative support for 2020; identification of herbaceous areas likely to be confused with Christmas tree plantations.
HRL Grassland	2017, 2023	10 m	Copernicus Land Monitoring Service	[ 41 , 42 ]	Grassland support layers for temporal inference in 2017/2018 and 2023.

Table I: Summary of the main datasets used in this study, their temporal coverage, and their role in the workflow.

Figure 3: Overview of the proposed workflow for Christmas tree plantation delineation from high-resolution RGB imagery. The pipeline combines BD ORTHO orthophotos, parcel-level reference data, and ancillary land-cover layers to construct a hard-negative-aware training dataset. Positive and hard negative samples are then used to train the semantic segmentation model, which is evaluated both on the 2020 held-out test set and in temporal transfer settings over 2017/2018 and 2023 through visually guided refinement and large-scale analysis.

Figure 4: Land-cover composition inside the verified Christmas tree parcels (2020 dataset). The bar chart illustrates the raw spatial overlap between the manually annotated parcels and the ancillary layers (grassland, clear-cut, and water), motivating the use of explicit priority rules during multi-class mask generation.

Figure 5: Qualitative comparison on the 2020 test set. Columns show (from left to right) RGB patch, GT 2020 multiclass mask, GT binary mask (Christmas tree class), and predictions from DeepLabV3-R18, DeepLabV3-R34, DeepLabV3-R50, FPN-R34, U-Net (b16), and U-Net++-R34 (thresholds as in Table II ).

Model	thr	IoU $\uparrow$	F1 $\uparrow$	Prec. $\uparrow$	Rec. $\uparrow$
DeepLabV3-R18	0.80	0.703	0.826	0.830	0.821
DeepLabV3-R34	0.70	0.733	0.846	0.866	0.827
DeepLabV3-R50	0.80	0.697	0.822	0.807	0.837
FPN-R34	0.75	0.700	0.824	0.833	0.814
U-Net (b16)	0.55	0.512	0.677	0.762	0.609
U-Net++-R34	0.75	0.731	0.845	0.843	0.846

Table II: Performance on the 2020 test set. Thresholds (thr) are selected on the validation set by maximizing the F1-score. Best values are highlighted in bold. Arrows indicate whether higher (

\uparrow

) values correspond to better performance.

Loss	IoU $\uparrow$	F1 $\uparrow$	Prec. $\uparrow$	Rec. $\uparrow$
Tversky	0.577 $\pm$ 0.067	0.730 $\pm$ 0.054	0.655 $\pm$ 0.053	0.826 $\pm$ 0.058
Weighted BCE	0.667 $\pm$ 0.042	0.800 $\pm$ 0.031	0.769 $\pm$ 0.102	0.849 $\pm$ 0.081
Weighted BCE + Tversky	0.695 $\pm$ 0.009	0.820 $\pm$ 0.006	0.756 $\pm$ 0.017	0.895 $\pm$ 0.011

Table III: Multi-seed loss ablation on DeepLabV3-R34 over the 2020 test set with fixed threshold 0.70. Values are reported as mean

\pm

standard deviation across seeds. Best values are highlighted in bold.

Figure 6: Examples of visual comparison for DeepLabV3–R34 trained with and without Hard Negative Mining (HNM) on 2020 test patches. Columns: RGB; GT (multiclass) over RGB; no-HNM mask over RGB; no-HNM $P(\mathrm{CT})$ ; HNM mask over RGB; HNM $P(\mathrm{CT})$ (shared $[0,1]$ scale). HNM suppresses diffuse false activations and produces more spatially coherent plantation responses.

Training regime	AP	AUPRC
no-HNM	0.2042	0.2044
HNM	0.9125	0.9127

Table IV: Threshold-free performance on the hard-negative 2020 test split (DeepLabV3–R34). AP corresponds to the area under the Precision–Recall curve (AUPRC).

Year	IoU $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	Recall $\uparrow$
2023	0.691	0.817	0.876	0.766
2017/2018	0.751	0.858	0.893	0.826

Table V: Temporal inference results on the refined GT on the fixed 2020 test patch locations. Metrics are computed for the HNM pipeline.

Panel of Figure 9 — Figure 9: Qualitative evaluation of the large-scale detection framework over the Morvan Regional Natural Park for the 2017/2018 and 2023 evaluation years. The main panels display the wall-to-wall spatial distribution of the predicted Christmas tree plantations against their respective ground-truth inventories across the entire 87,000 ha extent. The side panels provide representative zoomed-in areas detailing the local matching behavior: solid yellow masks indicate the model predictions, while red boundaries denote the reference ground-truth polygons. The selected insets highlight the model’s accuracy and temporal consistency across different spatial contexts, including dense central clusters and isolated peripheral regions.

Metric	2023	2017/2018
GT area (ha)	1,782.21	1,498.44
Pred. area (ha)	2,006.63	2,206.52
FP (ha)	695.81	889.80
FN (ha)	471.39	181.72
Precision	0.653	0.597
Recall	0.736	0.879

Table VI: Final large-scale area-based validation over the common study extent after the selected post-processing step.

Remote Sensing Image Segmentation

Remote sensing image segmentation labels land cover and objects in satellite, aerial, radar, and multispectral imagery.

See all 220 papers

Jun 23, 2026

Counting Trees from Satellite Imagery with Noisy Supervision

Counting individual trees is a fundamental task for environmental monitoring, yet remains largely unexplored with satellite imagery. At these resolutions, isolated trees may still be identifiable, but crown boundaries become ambiguous in dense forests, making the notion of an individual tree inherently ill-defined. Moreover, large-scale manual annotations of individual trees are prohibitively expensive. While scalable supervision can be derived from airborne LiDAR, the resulting annotations are noisy and difficult to exploit effectively. We address these challenges by formulating tree counting as a spatial density matching problem supervised through Unbalanced Optimal Transport. This formulation naturally accommodates both precise localization of isolate trees and robust density estimation in dense forests. We further introduce a self-correction mechanism that leverages transport residuals to progressively refine noisy supervision during training. We evaluate our approach on TinyTrees, a new benchmark spanning three continents and three satellite sensors, comprising over 216 million tree annotations (including 639k manually verified instances) across

25\,890

^2

. Our method consistently outperforms detection-based, regression-based, and transport-based distribution-matching baselines, demonstrating the effectiveness of unbalanced transport and reliability-aware supervision for large-scale tree counting from satellite imagery. Code, data and models are available at https://github.com/dgominski/treematch.

Dimitri Gominski, Maurice Mugabowindekwe, Qiue Xu +6

University of Copenhagen · University of Rwanda · University of Chinese Academy of Sciences +3

Remote Sensing Learning with Noisy Labels Remote Sensing Benchmarks Optimal Transport

Ask PDF

Oct 14, 2026

Deep Learning for Forest Plantation Mapping in Godavari Districts of Andhra Pradesh, India

Small-area forest plantations play a vital role in the socioeconomic well-being of farmers in Southeast Asia. Most plantations are established on former agricultural land, often on land less suitable for agriculture. Plantations that are converted from natural forest have adverse impacts on biodiversity. Mapping small-area plantations is thus important to understand the dynamics of forest cover in Southeast Asia and to study the social, economic, and ecological effects of this important land cover and land use change. While the small size of forest plantations makes it difficult to detect them using moderate resolution satellite sensors, the problem is exacerbated by the high degree of mixing between plantations, surrounding vegetation, and other land covers, which often show variegated responses in satellite signals across space and time. In this paper, we study the problem of mapping small-area forest plantations in East and West Godavari districts of Andhra Pradesh, India using deep learning methods. Remotely sensed cloud-free data from the Harmonized Landsat Sentinel-2 S10 product were classified using a pixel-level neural network and training data labeled using a field-based survey in concert with expert aerial photo interpretation. We compare the performance of deep learning methods with a baseline random forest classifier in our study region of 21543 sq. km over a period of 3 years and analyze the differences in the results across land cover classes and seasons.

Snehal More, Anuj Karpatne, Valerie A. Thomas +1

Department of Forest Resources and Environmental Conservation, Virginia Tech, Virginia, USA. · Department of Computer Science, Virginia Tech, Virginia, USA

Land Cover Classification Remote Sensing

Ask PDF

Oct 29, 2025

Mapping and Classification of Trees Outside Forests using Deep Learning

Trees Outside Forests (TOF) play an important role in agricultural landscapes by supporting biodiversity, sequestering carbon, and regulating microclimates. Yet, most studies have treated TOF as a single class or relied on rigid rule-based thresholds, limiting ecological interpretation and adaptability across regions. To address this, we evaluate deep learning for TOF classification using a newly generated dataset and high-resolution aerial imagery from four agricultural landscapes in Germany. Specifically, we compare convolutional neural networks (CNNs), vision transformers, and hybrid CNN-transformer models across six semantic segmentation architectures (ABCNet, LSKNet, FT-UNetFormer, DC-Swin, BANet, and U-Net) to map four categories of woody vegetation: Forest, Patch, Linear, and Tree, derived from previous studies and governmental products. Overall, the models achieved good classification accuracy across the four landscapes, with the FT-UNetFormer performing best (mean Intersection-over-Union 0.74; mean F1 score 0.84), underscoring the importance of spatial context understanding in TOF mapping and classification. Our results show good results for Forest and Linear class and reveal challenges particularly in classifying complex structures with high edge density, notably the Patch and Tree class. Our generalization experiments highlight the need for regionally diverse training data to ensure reliable large-scale mapping. The dataset and code are openly available at https://github.com/Moerizzy/TOFMapper

Moritz Lucas, Hamid Ebrahimy, Viacheslav Barkov +3

Osnabrück University, Institute for Computer Science, Remote Sensing Osnabrück, Germany · Leibniz Institute for Agricultural Engineering and Bioeconomy, System Process Engineering, Germany · Osnabrück University, Institute of Cognitive Science, Artificial Intelligence, Germany +1

Remote Sensing Image Segmentation Agricultural Remote Sensing Land Cover Classification Remote Sensing

Ask PDF

Abstract

Remote Sensing Image Segmentation

Explore similar work

Counting Trees from Satellite Imagery with Noisy Supervision

Deep Learning for Forest Plantation Mapping in Godavari Districts of Andhra Pradesh, India

Mapping and Classification of Trees Outside Forests using Deep Learning