Measurement-Driven Estimates of Above-Ground Biomass Change in the Eastern Canadian Boreal Forests from Permanent Sample Plots and Landsat Time Series

Abstract

Monitoring boreal above-ground biomass (AGB) change requires approaches that are both measurement-based and spatially explicit. We integrated permanent sample plots from Quebec and Ontario with Landsat-7 spectral trajectories (1999–2023) to quantify non-fire-related AGB change after excluding wildfire-affected intervals and to evaluate whether annualized AGB change can be predicted from spectral change at the plot-interval scale. Tree height was estimated using a multilayer perceptron model (R² = 0.83) and combined with species-specific allometry to derive plot-level AGB and interval ΔAGB. These estimates were aggregated to ecodistricts using effective sample sizes and confidence intervals. Across well-sampled ecodistricts, mean annualized ΔAGB ranged from −0.82 to +3.54 t ha⁻¹ yr⁻¹, with lower or negative changes mainly occurring in eastern regions. Spectral indices derived from NIR–SWIR bands showed relatively stronger associations with ΔAGB than greenness-based indices, consistent with the sensitivity of moisture- and disturbance-related metrics to canopy stress, including defoliation. An XGBoost ensemble correctly predicted the direction of change in 77% of intervals. These results provide a measurement-constrained and scalable framework for monitoring non-fire-related biomass change and supporting greenhouse-gas reporting across boreal forest landscapes.

1. Introduction

Boreal forests are among the world’s largest terrestrial carbon reservoirs and strongly influence the climate system through coupled exchanges of carbon, water, and energy [1,2,3]. Canada contains roughly one quarter of the world’s boreal forest, making changes in forest biomass particularly consequential for national greenhouse-gas inventories and global carbon budget assessments [4,5,6,7]. Accurate estimation of above-ground biomass (AGB) in forests is therefore crucial for quantifying carbon storage and assessing forest dynamics. Recent studies have shown that integrating ground-plot inventories with multi-source satellite observations and machine or deep learning models substantially improves AGB estimation by linking field-based biomass measurements to complex spectral and structural signals in remote-sensing data [8,9,10,11,12].

At the same time, the magnitude, the drivers, and the spatial patterns of recent AGB change in boreal forests remain uncertain. In Canada’s boreal forests, drought-induced increases in tree mortality and associated reductions in carbon-sink strength, together with biotic disturbances such as insect outbreaks, have already produced measurable losses of live biomass [13,14,15,16,17,18]. Repeated inventory analyses and satellite-based vegetation indices reveal widespread greening and biomass gains in many regions, but also climate-related slowdowns and local declines, with different methods often disagreeing on the relative roles of climate versus disturbance processes [19,20,21,22,23,24]. Near-term (5–30-year) [25] forecasts of ΔAGB are particularly challenging: stand-level growth trajectories are comparatively predictable, whereas partial mortality and other non-stand-replacing processes are multifactorial and stochastic [26,27]. Recent machine learning work combining large permanent sample plot networks, climate covariates, and multi-decadal Landsat time series has begun to map where near-term biomass gains and losses are likely to occur, but mainly at continental or broad ecozone scales rather than at decision-relevant units such as eastern Canadian ecodistricts [25].

Biomass monitoring in Canada rests on two complementary observational pillars. The plot-based pillar consists of the National Forest Inventory (NFI) and aligned provincial remeasurement programmes (e.g., Ontario Ministry of Natural Resources and Forestry; Ministère des Ressources naturelles du Québec), which provide detailed repeated tree and stand measurements from sample plots (including permanent sample plots (PSP), Permanent Inventory Plots (PIP), and Permanent Growth Plots (PGP) networks) that underpins standardized national reporting and model calibration [28]. The remote-sensing pillar encompasses a suite of model-based mapping approaches that integrate plot measurements with optical, radar, and lidar observations to generate wall-to-wall estimates of biomass and related forest attributes. A range of statistical and machine learning approaches has been used to map AGB from optical time series calibrated with inventory plots (e.g., k-nearest neighbour (kNN), regression/ensemble learning, and deep learning) [9,11,29,30,31,32,33,34], yet uncertainties remain in translating plot-level change to spatially explicit, temporally consistent estimates across administrative/ecological units.

Despite major advances, three important gaps remain for the eastern Canadian boreal forest. First, to our knowledge, most carbon budget and biomass studies report trends at national or broad ecozone scales and do not explicitly isolate wildfire-masked, multi-decadal changes in live AGB at the ecodistrict scale, which represents a primary ecological reporting unit used in regional planning and greenhouse-gas inventories; consequently, background growth and non-wildfire biomass losses remain poorly quantified at this scale [4,25,35]. Second, although model-assisted and model-based frameworks that combine forest inventory and remote sensing are well developed, they have rarely been applied to estimate annualized changes in AGB (ΔAGB; t ha⁻¹ yr⁻¹) derived from multi-year remeasurement intervals at the ecodistrict level using long Landsat time series with design-consistent uncertainty, as many operational products prioritize single-epoch biomass or disturbance status [9,12,36]. Third, because near-term biomass change is only intermittently predictable at the plot scale [25], ecodistrict-scale ΔAGB remains difficult to predict, and the sources of disagreement between plot-based and Landsat-derived ΔAGB remain poorly understood, particularly with respect to sampling intensity, disturbance history, and spatial representativeness [37,38,39,40].

The primary objective of this study was to quantify non-wildfire, annualized change in live above-ground biomass (ΔAGB; t ha⁻¹ yr⁻¹) at the ecodistrict scale across boreal Ontario and Québec by integrating sample-plot remeasurements with Landsat-7 time series and quantifying uncertainty consistent with the sampling and remeasurement structure. To isolate background dynamics relevant to management and greenhouse-gas accounting, we exclude stand-replacing wildfire [41] intervals using national burned-area composites and an independent burn-detection map, such that ΔAGB reflects growth, partial harvesting, insects, and other non-wildfire processes. We then quantify how well plot-interval ΔAGB and its components can be predicted from Landsat spectral trajectories, and we diagnose both disagreements and similarities between plot-based and satellite-derived indicators of biomass change as a function of sampling intensity, disturbance history, and spatial representativeness. Finally, we used these relationships to support spatially explicit characterization of biomass gain and loss across ecodistricts, including areas with sparse plot coverage.

2. Materials and Methods

2.1. Study Area

The study domain spans the eastern Canadian boreal forest across Québec and Ontario, covering approximately 193 million ha. The site-level inventory comprises 12,366 unique plots, acquired between 1999 and 2024 from permanent sample plots (PSPs), Permanent Growth Plots (PGPs), and Permanent Inventory Plots (PIPs) (

Table A1. Summary of the permanent plot inventory for the eastern boreal Québec–Ontario study area (1999–2024), reporting total plot-level census observations by plot network (PGP, PSP, PIP), the distribution of sampled plot areas (ha), and the frequency of plot remeasurement intervals for each province and the combined domain.

		Quebec	Ontario			Total
The total number of unique sample plots		7585	PGP	PSP	PIP	12,366
			3234	655	892
Sample plot area (ha)	0.01	0	1			1
	0.02	0	4			4
	0.04	7585	4673			12,258
	0.05	0	14			14
	0.06	0	38			38
	0.09	0	31			31
	0.1	0	20			20
Frequency of sample plot remeasurement intervals	1 time	1248	1975			3223
	2 times	6090	1551			7641
	3 times	216	1108			1324
	4 times	31	147			178

), with sampling primarily concentrated in the southern boreal portion of the region (Figure 1a,b). Following standard PSP protocols, each census represents a complete field measurement for a given plot at a specific time, and successive censuses from the same plot were paired to define remeasurement intervals used to characterize stand dynamics [42]. Plot areas follow provincial design specifications and are overwhelmingly dominated by 0.04 ha plots (n = 12,258), whereas alternative plot sizes (0.01, 0.02, 0.05, 0.06, 0.09, and 0.1 ha) occur only rarely (n ≤ 38 for each class). The dataset also provides repeated measurements over time, with 7641 plots remeasured twice, 1324 three times, and 178 four times, while 3223 plots were measured only once. For each census, we compiled plot identifiers and georeferenced coordinates, measurement year, species-resolved tree records, diameter at breast height (DBH), total height, and tree status from the Ontario Ministry of Natural Resources and Forestry Growth and Yield Program and Québec’s Données Québec repository (Ministère des Ressources naturelles et des Forêts).

The study area was stratified at the ecodistrict level of the Canadian National Ecological Framework. Ecodistricts are fourth-level units nested within ecozones, ecoprovinces, and ecoregions, delineated as relatively homogeneous areas in terms of climate, landforms, surficial deposits, soils, water and potential vegetation [43]. Within the study area, a total of 85 ecodistricts were considered; 66 ecodistricts contained sample plots whereas 19 ecodistricts lacked plots. Figure 1b shows pronounced spatial variation in sampling intensity, with plot density ranging from 0.02 to 9.83 plots per 10,000 ha and highest densities in the southern and central managed forest belt. Figure 1c summarizes ecodistrict-level sampling effort and temporal coverage, showing that both plot counts and remeasurement intervals vary by more than an order of magnitude across ecodistricts, and that observation periods typically span approximately two to three decades. Collectively, this ecodistrict-based stratification provides a policy-relevant reporting framework for summarizing plot-derived AGB dynamics and for mapping spatial patterns of ΔAGB using wall-to-wall Landsat coverage, while recognizing higher uncertainty in ecodistricts lacking direct plot support [44].

2.2. Ground Sample Plots and AGB Estimation

Field data collection in the provincial plot inventories followed standard boreal forest inventory protocols, with diameter at breast height (DBH; 1.3 m) and total height and species identity recorded for each enumerated stem [11]. However, tree inclusion thresholds and associated biomass estimation conventions differ among inventories. For example, in the United States stems are typically classified as “trees” when DBH ≥ 12.7 cm (USDA Forest Service), whereas the Canadian National Forest Inventory follows a lower threshold of DBH ≥ 9.1 cm [45]. Consistent with Canadian National Forest Inventory conventions and provincial PSP protocols, AGB in this study was quantified for all live trees with DBH ≥ 9.1 cm measured at 1.3 m. Across 1,272,299 DBH records (Ontario: 780,392; Quebec: 491,907), the distribution is strongly dominated by smaller diameter classes. The 9.1–15 cm class accounts for 64.8% of stems, followed by 15–20 cm (22.4%) and 20–30 cm (11.4%). Larger-diameter classes are comparatively rare, with 30–50 cm representing 1.3% of stems and >50 cm only 0.1% (Figure A1). To examine the structural characteristics of the sampled trees, all individuals with both recorded DBH and measured height were compiled across the study area. This subset comprised 341,771 paired DBH–height observations. Within this dataset, DBH averaged 18.08 cm with a range from 9.10 to 107.40 cm. Correspondingly, tree height averaged 14.54 m, ranging from 1.30 to 43.24 m. The distributions of both DBH and tree height are illustrated in (Figure A2), showing that DBH exhibits a right-skewed distribution with most observations concentrated in smaller diameter classes, whereas tree height is primarily distributed within intermediate ranges.

Species identity is a key determinant of AGB because allometric relationships between DBH, height, and biomass vary systematically among taxa owing to differences in wood density, crown architecture, and stem form. Broad allometric compilations for North America have shown that applying generic or mis-specified species groups can introduce stand-level biomass biases on the order of 10%–30%, particularly along hardwood–softwood gradients [46,47]. To reduce such bias and ensure consistency with Canadian forest carbon reporting, we used species-specific or genus-level equations, together with hardwood and softwood group models from the Canadian national biomass equation sets [48,49] with the general functional form described in Equations (1)–(4) to convert tree measurements to individual-tree AGB prior to plot-level aggregation.

Given that about 27% of trees with DBH measurements had observed heights and DBH–height allometry is nonlinear and species-dependent, we developed and validated a multilayer perceptron (MLP) model to predict tree height from DBH and species, thereby obtaining complete height information for AGB estimation. As shown in Figure 2, the height model was trained only on trees with both DBH and height measurements. Height data are available for 71 of the 76 species in the study area; the remaining 5 species have DBH records only. Given this uneven coverage and the strong imbalance in sample sizes among species, we grouped taxa into allometric classes to stabilize model training while retaining all observations. As shown in

Table 1. Classification of tree groups and their associated allometric subgroups.

Group Name	Allometric Subgroup	Number of Tree Records
Softwood Group 1	Spruce/Fir/Hemlock	214,475
Softwood Group 2	Pine	65,540
Hardwood Group 5	Soft Maple/Birch/Cherries	59,131
Hardwood Group 6	Aspen/Poplar/Willow	42,075
Hardwood Group 4	Oak/Hickory/Hard Maples	41,975
Softwood Group 3	Cedar/Larch	10,674
Hardwood Group 7	Mixed Hardwoods	7772

, the grouping follows established biomass–height allometry families, with softwood groups 1–3 (spruce/fir/hemlock, pine, cedar/larch) and hardwood groups 4–7 (oak/hickory/hard maples, soft maple/birch/cherries, aspen/poplar/willow, and mixed hardwoods), adapted from [47,48,50]. Height prediction is therefore conditioned on allometric group rather than individual species, improving model robustness and species coverage. This MLP approach was selected given its capacity to approximate complex non-linear functions [51]. The dataset was partitioned by stratified random sampling across allometric classes into training (70%), validation (15%), and test (15%) subsets, preserving class proportions. To assess potential data leakage associated with tree-level random partitioning, an additional independent validation was conducted using a plot-level split, where all trees from a given plot were assigned to a single subset. This approach ensured that no information was shared between training and evaluation datasets at the plot level. Predictors comprised DBH (cm) as a continuous variable and species identity as a categorical variable (encoded), with total tree height (m) as the response. Hyperparameters were tuned using RandomizedSearchCV over a predefined search space (

Table A2. Summary of the MLP model configuration, training strategy, and selected hyperparameters.

Component	Specification
Model type	Multi-layer Perceptron (MLP)
Objective	predict tree height (m)
Input variables	DBH (cm) and species (allometric group)
Output	Continuous height (m)
Architecture	Fully connected feedforward network
Hidden Layer Sizes	[(256, 128, 64), (128, 64, 32), (64, 32)]
Alpha (Regularization Strength)	[0.0001, 0.001, 0.01]
Activation Functions	[‘ReLU’]
Solvers (Optimization Algorithms)	[‘Adam’, ‘SGD’]
Loss function	Mean Squared Error (MSE)
Train/validation/Test split	70%/15%/15%
Performance	R2; MAE; RMSE

) [52] to balance model flexibility and simplicity. Model performance was assessed using regression metrics for continuous outcomes, including the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), providing complementary measures of model accuracy and error magnitude under species-specific allometry and a skewed DBH distribution [53]. The final tuned MLP was then applied to all trees lacking measured height, yielding a spatially and taxonomically complete height dataset for subsequent AGB computation.

After predicting tree height with the MLP, AGB was estimated at the component, tree, and plot levels and then expressed per unit area. For each biomass component c (wood, bark, branches, foliage), component-level biomass (kg) for tree j was computed using species-specific power-law allometric equations [48,49]:

$$B_{c,j}={\alpha }_{c,s}{DBH}_j^{{\beta }_{\mathrm{c},s}}{H}_j^{{\theta }_{c,s}}$$

(1)

where $B_{c,j}$ is the biomass (kg) of component $c$ for tree $j$; $DB{H}_j$ is diameter at breast height (cm); $H_j$ is total height (m); and ${\alpha }_{c,s},{\beta }_{\mathrm{c},s},{\theta }_{c,s}$ are species- and component-specific coefficients (See

Table A3. Species- and component-specific coefficients (${\alpha }_{c,s}, {\beta }_{c,s}, {\theta }_{c,s}$) of the allometric power-law model used to estimate above-ground biomass (AGB) components, where c denotes biomass components: wood (wo), bark (ba), branches (br), and foliage (fo) [48,49]. Rows with a white background represent softwood species, whereas rows with a grey background represent hardwood species.

Species	${\mathit{\alpha }}_{\mathbf{w}\mathbf{o},\mathbf{s}}$	${\mathit{\beta }}_{\mathit{w}\mathit{o},\mathbf{s}}$	${\mathit{\theta }}_{\mathbf{w}\mathbf{o},\mathbf{s}}$	${\mathit{\alpha }}_{\mathbf{b}\mathbf{a},\mathbf{s}}$	${\mathit{\beta }}_{\mathbf{b}\mathbf{a},\mathbf{s}}$	${\mathit{\theta }}_{\mathbf{b}\mathbf{a},\mathbf{s}}$	${\mathit{\alpha }}_{\mathbf{b}\mathbf{r},\mathbf{s}}$	${\mathit{\beta }}_{\mathbf{b}\mathbf{r},\mathbf{s}}$	${\mathit{\theta }}_{\mathbf{b}\mathbf{r},\mathbf{s}}$	${\mathit{\alpha }}_{\mathbf{f}\mathbf{o},\mathbf{s}}$	${\mathit{\beta }}_{\mathbf{f}\mathbf{o},\mathbf{s}}$	${\mathit{\theta }}_{\mathbf{f}\mathbf{o},\mathbf{s}}$
Black Spruce	0.0335	1.7389	0.9835	0.0132	1.7657	0.5775	0.0405	3.1917	−1.3674	0.2078	2.5517	−1.3453
Balsam Fir	0.0294	1.8357	0.8640	0.0053	2.0876	0.5842	0.0117	3.5097	−1.3006	0.1245	2.5230	−1.1230
Eastern Hemlock	0.0257	1.9277	0.8576	0.0118	1.9893	0.4700	0.0215	2.6553	−0.4682	0.1471	2.0108	−0.6080
Eastern Red Cedar	0.0520	1.7731	0.7054	0.0283	1.7079	0.0000	0.0219	2.3585	0.0000	0.2575	2.5136	−1.5565
Eastern White Cedar	0.0295	1.7026	0.9428	0.0076	1.7861	0.6132	0.0501	2.5165	−0.8774	0.0813	2.2180	−0.7907
Eastern White Pine	0.0170	1.7779	1.1370	0.0069	1.6589	0.9582	0.0184	3.1968	−1.0876	0.0584	2.2389	−0.5968
Jack Pine	0.0199	1.6883	1.2456	0.0141	1.5994	0.5957	0.0185	3.0584	−0.9816	0.0325	1.7879	0.0000
Red Pine	0.0106	1.7725	1.3285	0.0277	1.5192	0.4645	0.0125	3.3865	−1.1939	0.0731	2.3439	−0.7378
Red Spruce	0.0143	1.6441	1.4065	0.0274	2.0188	0.0000	0.0005	3.3136	0.0000	0.0106	2.2709	0.0000
White Spruce	0.0265	1.7952	0.9733	0.0124	1.6962	0.6489	0.0325	2.8573	−0.9127	0.2020	2.3802	−1.1103
Other softwood species	0.0276	1.6868	1.0953	0.0101	1.8486	0.5525	0.0313	2.9974	−1.0383	0.1379	2.3981	−1.0418
Eastern Cottonwood	0.0051	1.0697	2.2748	0.0009	1.3061	2.0109	0.0131	2.5760	0.0000	0.0224	1.8368	0.0000
Balsam Poplar	0.0117	1.7757	1.2555	0.0180	1.8131	0.5144	0.0112	3.0861	−0.7164	0.0617	1.8615	−0.5375
American Basswood	0.0168	1.9844	0.8989	0.0057	1.5881	1.1472	0.0039	2.0084	0.8588	0.0147	1.8300	0.0000
American Beech	0.0432	2.0378	0.7000	0.0049	1.9057	0.6770	0.0355	2.3749	0.0000	0.0452	1.5567	0.0000
Black Ash	0.0306	2.1836	0.5740	0.0897	2.2634	−0.567	0.0994	2.1630	−0.4809	0.0124	1.0325	0.8747
Black Cherry	0.0181	1.7013	1.3057	0.0101	1.5956	0.9190	0.0005	2.8004	0.8603	0.1976	1.4421	−0.5264
Gray Birch	0.0295	1.9064	0.9139	0.0148	1.8433	0.5021	0.0150	3.0347	−0.7629	0.0455	2.6447	−1.4955
Bitternut Hickory	0.0139	1.5913	1.5080	0.0081	1.4943	1.1324	0.0050	3.0463	0.0000	0.0121	2.0865	0.0000
Large-tooth Aspen	0.0128	2.0633	0.9516	0.0240	2.3055	0.0000	0.0131	3.1274	−0.8379	0.0382	2.1673	−0.6842
Red Maple	0.0315	2.0342	0.7485	0.0283	2.0907	0.0000	0.0225	2.4106	0.0000	0.0571	1.4898	0.0000
Silver Maple	0.0274	1.7126	1.1086	0.0123	1.8250	0.5010	0.0543	3.7343	−1.6497	6.6808	2.1092	−2.1697
Sugar Maple	0.0301	2.0313	0.8171	0.0103	1.7111	0.8509	0.0661	2.5940	−0.4933	2.5019	2.4527	−2.3008
Trembling Aspen	0.0142	1.9389	1.0572	0.0063	2.0819	0.6617	0.0137	2.9270	−0.6221	0.0270	1.6183	0.0000
White Ash	0.0224	1.7438	1.1899	0.0126	1.6456	0.7893	0.0354	2.3046	0.0000	0.0195	1.0509	0.7836
American Elm	0.0207	2.2276	0.6488	0.0078	2.4540	0.0000	0.0393	2.1880	0.0000	0.0516	1.4511	0.0000
White Oak	0.0442	1.6818	1.0310	0.0308	1.7479	0.3504	0.0022	2.0165	1.3953	0.0053	1.2822	1.1323
Yellow Birch	0.0259	1.9044	0.9715	0.0069	2.0834	0.5371	0.0325	2.3851	0.0000	0.1683	1.2764	0.0000
Black Maple	0.0315	2.0342	0.7485	0.0283	2.0907	0.0000	0.0225	2.4106	0.0000	0.0571	1.4898	0.0000
Bur Oak	0.0285	1.8501	1.0204	0.0326	1.8100	0.4153	0.0013	3.0637	0.3153	0.0582	1.5438	0.0000
Shagbark Hickory	0.0139	1.5913	1.5080	0.0081	1.4943	1.1324	0.0050	3.0463	0.0000	0.0121	2.0865	0.0000
Swamp White Oak	0.0442	1.6818	1.0310	0.0308	1.7479	0.3504	0.0022	2.0165	1.3953	0.0053	1.2822	1.1323
Manitoba Maple	0.0315	2.0342	0.7485	0.0283	2.0907	0.0000	0.0225	2.4106	0.0000	0.0571	1.4898	0.0000
Paper Birch	0.0338	2.0702	0.6876	0.0080	1.9754	0.6659	0.0257	3.1754	−0.9417	0.1415	2.3074	−1.1189
Other hardwood species	0.0353	2.0249	0.7048	0.0090	1.8677	0.7144	0.0448	2.6855	−0.5911	0.0869	1.8541	−0.5491

). Total tree biomass (kg) was then obtained as the sum of component-level biomass across all components:

$${AGB}_j={\sum }_{c\in \left\{wood, bark, branches, foliage\right\}}{B}_{c,j}.$$

(2)

Plot-level biomass (kg) for plot $p$ was obtained by summing over all trees $j$ within that plot:

$${AGB}_p^{\left(\mathrm{k}\mathrm{g}\right)}= {\sum }_{j\in P}{AGB}_j,$$

(3)

finally, biomass per unit area (${AGB}_p^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})}$) for plot $p$ is computed by converting kilograms to tonnes and normalizing by plot area (${Area}_p^{\left(\mathrm{h}\mathrm{a}\right)}$):

$${AGB}_p^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})}={\displaystyle \frac{{AGB}_p^{\left(\mathrm{k}\mathrm{g}\right)}/1000}{{Area}_p^{\left(\mathrm{h}\mathrm{a}\right)}}}.$$

(4)

Figure 3 summarizes the workflow used to convert ground-plot measurements to plot-level AGB.

2.3. Plot-Level AGB Change and Ecodistrict Aggregation

Annualized change in live AGB for each remeasurement interval $i$ ${(∆AGB}_{yr,i}(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1}{\mathrm{y}\mathrm{r}}^{-1}))$ was computed using plot-level biomass expressed per unit area $(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})$, as defined in Equation (4), as:

$$\mathsf{\Delta }AG{B}_{\mathrm{y}\mathrm{r},i}=({AGB}_{p, end, i}^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})}-{AGB}_{p, start, i}^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})})/\mathsf{\Delta }{t}_i$$

(5)

where ${AGB}_{p,start, i}^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})}$ and ${AGB}_{p, end, i}^{(\mathrm{t} {\mathrm{h}\mathrm{a}}^{-1})}$ denote the plot-level AGB per unit area at the beginning and end of remeasurement interval $i$, respectively and $\mathsf{\Delta }{t}_i$ is the length (years) of interval $i$. After initial screening, interval-level quality control (QC) was applied to a set of 10,147 remeasurement intervals prior to aggregation [25,54]. Intervals with missing required variables (e.g., plot-level attributes or interval metrics) were excluded, although no records were removed at this stage. Intervals with short remeasurement periods ($\mathsf{\Delta }{t}_i<4 \mathrm{y}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{s}$) were excluded to avoid unstable annualized estimates, resulting in the removal of 28 intervals. In addition, implausible values of $\mathsf{\Delta }AG{B}_{\mathrm{y}\mathrm{r},i}$ were filtered using a conservative, data-informed threshold of |$\mathsf{\Delta }AG{B}_{\mathrm{y}\mathrm{r},i}$|> 10 t ha⁻¹ yr⁻¹, removing a further 83 intervals. This threshold was selected to exclude extreme outliers inconsistent with observed boreal biomass dynamics while retaining realistic variability. After QC, 10,036 intervals were retained for analysis. To assess sensitivity to the plausibility threshold, alternative cutoffs of ±8 and ±12 t ha⁻¹ yr⁻¹ were also evaluated. Resulting ecodistrict-level estimates and classifications were largely unchanged, supporting the robustness of the selected cutoff.

The plot to ecodistrict aggregation and uncertainty-estimation workflow is summarized in Figure 4. Aggregating interval-level changes rather than static stocks provides a more direct measure of contemporary carbon fluxes and disturbance–recovery dynamics in boreal forests [25,35,40,55,56].

To obtain spatially explicit statistics suitable for linking with remote sensing predictors, we grouped intervals by ecodistrict and summarized them using domain-level summary estimators. For each ecodistrict $e$, we computed the arithmetic mean annual AGB change (${∆AGB}_{yr,e}$ (t ha⁻¹ yr⁻¹)), as:

$${∆AGB}_{yr,e}={\displaystyle \frac{1}{n_e}}{\sum }_{i\in e}{∆AGB}_{yr,i}$$

(6)

where $n_e$ is the number of intervals in ecodistrict $e$. We also computed the sample standard deviation $s_e$ of interval-level ${∆AGB}_{yr,i}$ and the number of contributing plot intervals. Because multiple intervals within an ecodistrict are not fully independent, we adjusted the nominal sample size using an effective sample size $N_{eff, e}$ derived from the intraclass correlation coefficient (ICC) and calculated an effective standard error as

$${SE}_{eff,e}=S_e/\sqrt{N_{eff,e}}.$$

(7)

A 95% confidence interval (${\mathrm{CI}}_{\mathbf{95},\mathrm{e}}$) for the ${∆AGB}_{yr,e}$ was obtained using a $t$-critical value with $N_{\mathrm{eff},\mathrm{e}}-1$ degrees of freedom. We do not assume that interval-level ${∆AGB}_{yr,i}$ itself follows a symmetric Student’s t distribution; rather, the t-based CI is used as an approximation for the ecodistrict-level mean ${∆AGB}_{yr,e}$ and aggregation across intervals reduces the effect of skewness. Ecodistricts were classified as “Gain” if the lower CI bound of ${\mathrm{CI}}_{\mathbf{95},\mathrm{e}}$ was >0, “Loss” if the upper CI bound was <0, and “No change” otherwise, and a separate “Insufficient” class was assigned to domains with very few intervals or undefined variance. Full formulas for $N_{\mathrm{eff},\mathrm{e}}$, $S{E}_{\mathrm{eff},\mathrm{e}}$, the t-critical values, and the 95% confidence bounds, as well as the significance (SIG) classification rule, are provided in Appendix B. This combination of interval-level QC, domain-level effective sample sizes, and CI-based significance labelling is consistent with commonly used approaches for large-area biomass-change estimation from NFI-type data [25,40,55,57].

Monte Carlo simulation is commonly used to propagate uncertainty in environmental models by repeatedly perturbing uncertain input variables and propagating these uncertainties through the full modelling chain [58,59]. In this study, uncertainty in tree height estimates was propagated through biomass calculations and ΔAGB estimation using repeated stochastic simulations (n = 100 runs). For trees with imputed height, perturbations were applied assuming approximately Gaussian prediction errors:

$$H_j^{\left(r\right)}=H_j+{\epsilon }_j^{\left(r\right)},{\epsilon }_j^{\left(r\right)}~\mathrm{N}(0,{\sigma }_H^2)$$

(8)

where $H_j^{\left(r\right)}$ is the perturbed height of tree $j$ in simulation $r$, $H_j$ is the original (predicted) height of tree $j$, and ${\epsilon }_j^{\left(r\right)}$ is a random error term drawn independently for each tree and simulation. The parameter ${\sigma }_H$ represents the standard deviation of the height prediction error, approximated here by the RMSE of the height prediction model. The perturbed heights were then propagated through allometric equations, plot-level aggregation, interval-level ΔAGB estimation, and ecodistrict-level confidence interval and classification (SIG) calculations. In addition, a sensitivity analysis was performed to evaluate the influence of allometric uncertainty by applying a uniform ±5% scaling factor to AGB estimates. Together, these analyses provide an approximate quantitative assessment of the sensitivity of ΔAGB estimates and ecodistrict classifications to uncertainties in tree height prediction and allometric scaling.

2.4. Remote Sensing Predictors and ΔAGB Modelling

In Google Earth Engine (GEE), we generated Landsat-7 Collection 2, Level-2 surface reflectance composites for July–August over 1999–2023 to represent peak growing-season conditions in the eastern Canadian boreal forest [25]. For each calendar year $Y$, we constructed a two-year moving window ($Y-1$ to $Y$), applied the standard USGS radiometric scaling, and removed cloud, cloud shadow, snow, and cirrus using the pixel-quality band, following national Landsat compositing protocols [8,30]. To mitigate the Landsat-7 SLC-off (scan line corrector failure) striping effects after 2003, we relied on multi-temporal compositing (two-year window) combined with median aggregation to reduce data gaps, together with a 3 × 3 focal median filter to suppress residual striping artefacts. In addition, the use of a single-sensor (Landsat-7) time series ensured spectral consistency across the study period, avoiding inter-sensor radiometric differences that could affect spectral indices and ΔAGB modelling. Following these steps, we computed the median surface reflectance in the BLUE, GREEN, RED, NIR, SWIR1, and SWIR2 bands, as well as the per-pixel number of clear observations. Plot-level predictors were obtained by sampling the start- and end-year composites (30 m resolution) at plot centroids. At the ecodistrict scale, the same July–August composites were spatially aggregated within each ecodistrict polygon after masking permanent water bodies and stand-replacing burned areas [41]. Band medians and mean clear-observation counts were then derived for an “early” period (2000–2012) and a “late” period (2012–2023), consistent with recent Landsat-based biomass-change studies [25].

From these July–August Landsat-7 composites, we derived a suite of spectral vegetation indices characterizing canopy greenness, moisture status, structure, and senescence at both plot and ecodistrict scales. For each plot interval, we calculated Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR), Normalized Difference Infrared Index (NDII), Normalized Difference Water Index (NDWI), Moisture Stress Index (MSI), Plant Senescence Reflectance Index (PSRI), Soil-Adjusted Total Vegetation Index (SATVI), Enhanced Vegetation Index (EVI), and Near-Infrared Reflectance of Vegetation (NIRv) at interval start and end (

Table A4. Landsat vegetation index (VI) metrics.

Vegetation Index	Formulas	Biophysical Meaning	Reference
Normalized Difference Vegetation Index (NDVI)	(NIR − R)/(NIR + R)	photosynthetic activity and canopy density	[86,87,88]
Enhanced Vegetation Index (EVI)	2.5 × (NIR − R)/(NIR + 6R − 7.5B + 1)
Normalized Difference Infrared Index (NDII)	(NIR − SWIR1)/(NIR + SWIR1)	canopy and soil moisture and fire-related structural change	[89,90]
Normalized Burn Ratio (NBR)	(NIR − SWIR2)/(NIR + SWIR2)
Moisture Stress Index (MSI)	SWIR1/NIR	emphasize water stress	[88,91]
Normalized Difference Water Index (NDWI)	(G − NIR)/(G + NIR)
Soil-Adjusted Total Vegetation Index (SATVI)	((SWIR1 − R)/(SWIR1 + R + 0.5)) × 1.5 − (SWIR2/2)	track senescence and soil/background effects	[92,93]
Plant Senescence Reflectance Index (PSRI)	(R − G)/NIR
Near-Infrared Reflectance of Vegetation (NIRV)	NDVI × NIR	provides a proxy for canopy structure and primary productivity	[94]

). Annualized change metrics ($\mathsf{\Delta }{\mathrm{VI}}_{\mathrm{yr}}$, index units ${\mathrm{y}\mathrm{r}}^{-1}$) were obtained by differencing end and start interval values and dividing by the remeasurement interval length. At the ecodistrict level, the same indices and annualized changes were computed for the early (2000–2012) and late (2012–2023) periods, yielding spatially aggregated spectral trajectories analogous to the plot-level $\mathsf{\Delta }{\mathrm{AGB}}_{\mathrm{yr}}$ estimates and used as predictors in subsequent XGBoost-based $\mathsf{\Delta }\mathrm{AGB}$ modelling [25].

To model annualized plot-level AGB change, we used an Extreme Gradient Boosting (XGBoost) regression ensemble [60]. XGBoost was selected because it can efficiently model nonlinear relationships and interactions in large remote-sensing datasets and has been widely used in biomass modelling studies [25]. To benchmark the added value of XGBoost, two simpler models were also evaluated using the same predictors and grouped cross-validation framework: Ridge regression [61] and Random Forest [62]. Additional configuration details for the primary and benchmark models are summarized in

Table A7. Summary of model configurations, validation strategy, and tuning procedures for XGBoost and benchmark models.

Component	Specification
Primary model	XGBoost (gradient boosting)
Benchmark models	Random Forest regression; Ridge regression
Predictors	Spectral vegetation indices at the start of each interval; annualized changes in spectral indices; interval duration; interval mid-year
Response variable	Annualized AGB change, ΔAGB (t ha−1 yr−1)
Validation design	Five-fold grouped cross-validation
Preprocessing	Outlier control, imputation, and Ridge standardization
XGBoost tuning	Tree depth (4–6), learning rate (0.03–0.10), minimum child weight (1–5), row and predictor subsampling (0.70–0.90), L1/L2 regularization, and early stopping
RF tuning	Number of trees (300–500), maximum tree depth (10–unlimited), minimum terminal node size (1–2), and predictors considered per split (square-root rule)
Ridge tuning	Regularization strength, α (0.01–100)
Hyperparameter tuning	Grid search over a compact parameter set
Loss function	Mean Squared Error (MSE)
Performance metrics	R2; MAE; RMSE; Directional Accuracy, precision, recall, F1-score, false negative rate

. Predictors included start-of-interval values and annualized changes in spectral indices, along with interval duration, interval mid-year, a binary indicator identifying intervals spanning both the early (2000–2012) and late (2012–2023) periods, and a clear-sky observation metric. We used five-fold cross-validation grouped by plot identifier [63] to avoid leakage among intervals from the same plot, tuning a compact grid of XGBoost hyperparameters with early stopping and winsorizing ΔVI features to limit the influence of outliers. The XGBoost model was fit for absolute ΔAGB (t ha⁻¹ yr⁻¹) and model performance was summarized using MAE, RMSE, $R^2$, and directional accuracy, defined as the proportion of intervals for which the model correctly predicts the sign of ΔAGB. Additional binary classification metrics (precision, recall, F1-score, balanced accuracy, and false negative rate (FNR)) were computed from out-of-fold predictions to assess model performance for loss detection [25]. A final seed ensemble was then refit using all QC intervals, after which permutation-based feature importance was applied to rank predictors.

In the final step, we used the trained XGBoost ensemble of plot-level ΔAGB predictions to generate ecodistrict-level estimates of ${∆AGB}_{yr,e}$ (t ha⁻¹ yr⁻¹). For each ecodistrict, we assembled a feature table of Landsat-7 spectral indices for the early (2000–2012) and late (2012–2023) periods, together with their annualized changes, and passed these predictors through each member of the absolute ΔAGB ensemble, averaging across seeds to obtain a single model-based ΔAGB estimate per ecodistrict. We then combined these predictions with ecodistrict-level sampling statistics from the plot network (number of intervals, within ecodistrict variance, effective sample size) to compute standard errors and 95% CI, following model-assisted small area estimation practice in large area forest inventories [36,64,65]. Ecodistricts whose CIs were entirely above zero were classified as biomass “Gain”, those entirely below zero as “Loss”, and all others as “No change”, yielding a directionally explicit map of ΔAGB consistent with the gain/loss summaries in [25]. The overall remote-sensing workflow is summarized in Figure 5.

3. Results

3.1. Plot-Level AGB Estimates from MLP-Based Height Prediction and Allometric Equations

The study area contains 76 tree species in total, including 18 softwoods and 58 hardwoods. Softwoods account for 81% of all records, while hardwoods comprise 19% of records. Species composition was dominated by three boreal softwoods, including black spruce (35.42% of stems), jack pine (19.25%), and balsam fir (18.59%), while the leading hardwoods are paper birch (8.27%) and trembling aspen (7.37%). Together, these five species represent about 89% of all records. The sampled dataset was dominated by spruce, fir, and pine, whereas birch and aspen represented the most common broadleaf associates (see

Table A5. Tree species composition of the plot inventory, reporting relative abundance (%) by species and functional group, and summarizing the overall contributions of softwoods (81%) and hardwoods (19%) across the study area.

Softwood: 81%			Hardwood: 19%
Species	Percentage	Type	Species	Percentage < 0.01	Type
Black spruce	35.42	Softwood	Black maple		Hardwood
Balsam fir	19.25		Shagbark hickory
Jack pine	18.59		Scots pine		Softwood
Paper birch	8.27	Hardwood	European larch
Trembling aspen	7.37		American hornbeam		Hardwood
White spruce Eastern white cedar	4.67 1.22	Softwood	Speckled alder
Red maple	0.89	Hardwood	Bur oak
American larch	0.72	Softwood	Rock elm
Sugar maple	0.56	Hardwood	Swamp white oak
Red pine	0.46	Softwood	Hybrid poplar
Yellow birch	0.35	Hardwood	Slippery elm
Balsam poplar	0.28		Cucumber tree
Eastern white pine	0.20	Softwood	Big-leaf linden
Red spruce	0.13		Hybrid larch		Softwood
American beech	0.12	Hardwood	Sorbus species		Hardwood
Eastern hemlock	0.11	Softwood	Chokecherry
Pin cherry	0.09	Hardwood	Sassafras
Black ash	0.06		Manitoba maple
Salix tree species	0.04		Carolina poplar
American mountain-ash	0.03		Sweet pignut hickory
Large-tooth aspen	0.03		Sweet cherry
Northern red oak	0.03		Eastern cottonwood
American basswood	0.03		Bay-leaved willow
White ash	0.02		Populus species
Ironwood	0.02		Black walnut
Striped maple	0.02		Amelanchier species
Northern mountain-ash	0.01	Hardwood	Tulip tree		Hardwood
Black cherry			Buckthorn
Gray birch			Eastern red cedar		Softwood
American elm			Unknown tree species
Silver maple			Sweet chestnut		Hardwood
Green ash			Chinquapin oak
Mountain maple			Black oak
Bitternut hickory			Unknown hardwood
White oak			Black oak
Pitch pine	0.01	Softwood	White willow
Norway spruce			Common hackberry

).

The final MLP configuration selected by grid search comprised three hidden layers with 128, 64, and 32 neurons, ReLU activation, stochastic gradient descent (SGD), and modest L2 regularization (α = 0.0001). This architecture showed consistent performance across the validation and independent test datasets, with R² ≈ 0.83 in both cases. Corresponding error metrics were also nearly identical, with RMSE values of 2.32 and 2.31 m and MAE values of 1.76 and 1.75 m for the validation and test sets, respectively. The similarity between validation and test metrics suggests good generalization with limited evidence of overfitting.

Scatterplots of observed versus MLP-predicted tree height for both the validation and test datasets (Figure 6a,b) show a dense, approximately linear distribution of points aligned with the 1:1 line. The model captures the primary gradient of tree height across the full range of observations, with no strong evidence of systematic curvature or severe heteroscedasticity. These patterns suggest that a single global non-linear model was adequate across species groups and DBH classes. A slight tendency toward underestimation is observable for taller trees (>30–35 m), likely reflecting the limited representation of large stems in the training data; however, these deviations are confined to the upper tail of the distribution and do not materially affect overall model performance. An additional independent validation using a plot-level split yielded nearly identical results to those obtained from the original partitioning (R² ≈ 0.83; RMSE ≈ 2.30 m; Figure A3), indicating that model performance is robust and not materially influenced by sample dependence.

Model errors were relatively consistent across the seven allometric groups, with MAE on the test set ranging from 1.52 to 2.20 m (

Table 2. Test-set performance metrics (MAE, RMSE, and R²) of the MLP tree height model across seven allometric groups.

Group Name	Allometric Subgroup	Count	MAE	RMSE	R2
Softwood Group 1	Spruce/Fir/Hemlock	32,172	1.52	2.02	0.82
Softwood Group 2	Pine	9831	2.20	2.80	0.79
Hardwood Group 5	Soft Maple/Birch/Cherries	8870	1.78	2.32	0.76
Hardwood Group 6	Aspen/Poplar/Willow	6311	1.87	2.43	0.85
Hardwood Group 4	Oak/Hickory/Hard Maples	6296	1.93	2.54	0.81
Softwood Group 3	Cedar/Larch	1601	2.08	2.78	0.64
Hardwood Group 7	Mixed Hardwoods	1166	1.89	2.48	0.86

). RMSE values showed a similar pattern (2.02–2.80 m), while R² ranged from 0.64 to 0.86, indicating broadly consistent predictive performance across groups. The lowest errors occurred for the dominant softwood group comprising spruce, fir, and hemlock (MAE = 1.52 m; R² = 0.82), whereas higher errors were observed for the pine group (MAE = 2.20 m) and the cedar–larch group (MAE = 2.08 m), possibly reflecting greater variability in height–diameter relationships and the relatively small sample size of the cedar–larch group (n = 1601). Hardwood groups showed intermediate performance. Overall, these results indicate that the MLP provides sufficiently consistent height predictions across the main species groupings required for subsequent biomass estimation.

After tree height prediction with the MLP, measured and predicted heights were used in the species-specific allometric equations to compute tree-level AGB. The resulting biomass distribution is summarized below. Most stems fell within low to intermediate biomass classes (10–50 kg), with progressively fewer trees in higher AGB classes. Detailed counts by biomass class and tree type, together with summary statistics, are provided in Figure A4. Plot-level AGB ranged from approximately 0.1 to 380 t ha⁻¹, with a mean of 71.2 t ha⁻¹ and a standard deviation of 46.6 t ha⁻¹, indicating considerable variability among plots. The boxplot (Figure A5) shows a right-skewed distribution with several high-AGB outliers above approximately 200 t ha⁻¹, likely representing a relatively small number of very dense stands.

Mean plot-derived AGB at the ecodistrict scale ranged from about 23 to 148 t ha⁻¹ (Figure 7). The lowest mean values (23–43 t ha⁻¹) occurred mainly in northern and interior ecodistricts with colder climates and a higher prevalence of open or low-productivity forest types, whereas intermediate values (44–77 t ha⁻¹) dominated much of central Québec and Ontario. The highest mean AGB classes (78–148 t ha⁻¹) were concentrated in southern and southwestern ecodistricts, where longer growing seasons and the predominance of productive mixedwood and hardwood stands likely contributed to higher biomass, consistent with previous findings for Canadian boreal forests [60,61,62,63]. Ecodistricts lacking plots are shown as “No plots” in this plot-based summary and were not included in the calculation of mean plot-derived AGB.

3.2. Ecodistrict-Level Mean Annualized AGB Change

Having characterized static AGB patterns, we next examined annualized biomass change and associated uncertainty. Effective sample sizes ($N_{\mathrm{eff},e}$) varied from 3 to more than 1000 remeasurement intervals (median ≈ 130), producing 95% confidence-interval widths (CI widths) between 0.27 and 5.44 t ha⁻¹ yr⁻¹ (median ≈ 0.72). Figure 8 illustrates the inverse relationship between data density and precision: CI width declines steeply as the number of remeasurement intervals increases and gradually levels off at higher sampling frequencies. This pattern indicates that ecodistricts with relatively few remeasurement intervals yield more uncertain estimates of ΔAGB, whereas ecodistricts with larger numbers of intervals provide substantially narrower confidence intervals for mean biomass change. The observed trend is broadly consistent with the expected reduction in sampling uncertainty with increasing sample size. Variance decomposition further showed that most variability occurred within ecodistricts rather than between them (${\sigma }_{within}^2=8.60$ vs. ${\sigma }_{between}^2=1.18$), corresponding to an ICC of ${\rho }_d=0.12$. This moderate within-ecodistrict correlation supports the use of an effective sample size adjustment.

Across the 59 ecodistricts with sufficient data, ${∆AGB}_{yr,e}$ ranged from −0.82 to +3.54 t ha⁻¹ yr⁻¹, with a median of 0.78 and an overall mean of 0.97 t ha⁻¹ yr⁻¹, indicating that modest net biomass gains predominated overall, although two ecodistricts experienced substantial net losses. The strongest gains occurred mainly in southern ecodistricts, including 403, 404, 406, and 389, whereas pronounced losses were concentrated in eastern ecodistricts including 447 and 445. Based on the CIs, 37 ecodistricts were classified as significant biomass gains, 2 as losses, and 20 as no change; the remaining 7 ecodistricts containing plots were classified as “Insufficient data” (Figure 9a,b). Sensitivity analyses confirmed that these patterns were robust to the choice of plausibility threshold. Only two ecodistricts (393 and 419) changed classification under the stricter ±8 t ha⁻¹ yr⁻¹ threshold, shifting from no change to gain due to minor shifts in their confidence intervals around zero. No classification changes were observed between the ±10 and ±12 thresholds, indicating that the main conclusions were insensitive to the selected cutoff.

Uncertainty propagation analysis indicated that ecodistrict-level ΔAGB estimates were robust to both tree height prediction error and allometric scaling uncertainty. Monte Carlo simulations (n = 100) showed that, on average, fewer than one ecodistrict changed classification in a typical run (mean = 0.16), and only four out of 59 ecodistricts exhibited any classification change across all simulations, with only one ecodistrict showing a classification-change probability greater than 10%. The mean absolute change in interval-level annualized ΔAGB was 0.17 t ha⁻¹ yr⁻¹, while changes at the ecodistrict level were minimal, with a mean absolute difference of approximately 0.025 t ha⁻¹ yr⁻¹ and a mean change in confidence interval width of approximately 0.04 t ha⁻¹ yr⁻¹ (

Table A6. Sensitivity of ecodistrict-level ΔAGB estimates to tree height uncertainty, assessed using Monte Carlo simulations (n = 100). Reported metrics include the number of ecodistricts with classification changes (SIG), as well as changes in ΔAGB magnitude and confidence interval (CI) width.

Metric	Value
Number of ecodistricts	59
Mean number of SIG changes per run	0.16
Maximum number of SIG changes per run	1
Number of ecodistricts with any SIG change	4
Number with probability of SIG change ≥5%	1
Number with probability of SIG change ≥10%	1
Mean absolute interval-level ΔAGB change (t ha−1 yr−1)	0.17
Mean absolute change in ecodistrict mean ΔAGB (t ha−1 yr−1)	0.025
Mean absolute change in CI width (t ha−1 yr−1)	0.04

). These variations are small relative to the overall range and distribution of ΔAGB values (Figure 9a). Similarly, a ±5% sensitivity analysis of allometric scaling resulted in no changes in ecodistrict classification (SIG), with only minor proportional effects on ecodistrict mean ΔAGB (≈0.025 t ha⁻¹ yr⁻¹) and confidence interval width (≈0.04 t ha⁻¹ yr⁻¹). Together, these results demonstrate that the spatial patterns of biomass gain, loss, and no change are insensitive to uncertainties in tree height estimation and moderate variations in allometric equations.

3.3. Remote-Sensing Predictors and ΔAGB Model Performance

Before fitting the XGBoost models, we assessed whether July–August Landsat spectral indices contained a detectable signal of annualized ΔAGB change. Across 8123 quality-controlled (QC) plot intervals, annual changes in all indices were generally small in magnitude; however, several moisture and disturbance-sensitive indices exhibited moderate but consistent associations with ΔAGB (Figure 10). In particular, NDII and NBR showed relatively higher absolute Pearson correlation coefficients with ΔAGB. These relationships are consistent with the physical interpretation of these indices. NBR, which contrasts near-infrared and shortwave infrared reflectance, is widely used to detect canopy disturbance and vegetation recovery and can therefore respond to structural changes in vegetation [66]. Similarly, NDII is sensitive to variations in canopy water content and vegetation moisture stress, conditions that are frequently associated with changes in forest productivity and biomass [67]. Consequently, intervals characterized by increasing canopy moisture and structural recovery (higher NDII and NBR and lower MSI) tended to correspond to biomass gains, whereas intervals exhibiting the opposite spectral responses were more likely to show biomass losses. In contrast, greenness indices such as NDVI and EVI exhibited comparatively lower correlations with ΔAGB, suggesting that simple changes in photosynthetic activity alone are a relatively poor proxy for structural biomass change over 5–10-year periods. Hexagonally binned scatterplots of ΔNBR and ΔNDVI (Figure A6) further illustrate these patterns. The highest densities of points occur near zero change in both ΔAGB and ΔVI, reflecting the predominance of low-magnitude gains and losses across intervals, while larger biomass decreases tend to be associated with modest declines in indices such as NBR and NDVI. However, the clouds of points are broad and approximately elliptical, with substantial overlap between gain and loss intervals, emphasizing that individual spectral indices explain only a modest fraction of the variance in ΔAGB and that the underlying relationships are non-linear and noisy. Together, these diagnostics motivate the use of a multi-index, non-linear ensemble model rather than reliance on any single spectral index or threshold to infer biomass change.

At the plot level, the XGBoost ensemble for predicting ${\mathrm{\Delta }AGB}_{yr}$ achieved an MAE of 1.54 t ${\mathrm{h}\mathrm{a}}^{-1}{\mathrm{y}\mathrm{r}}^{-1}$, an RMSE of 2.42 t ${\mathrm{h}\mathrm{a}}^{-1}{\mathrm{y}\mathrm{r}}^{-1}$, and $R^2=0.40$ across QC remeasurement intervals. The model correctly predicted the direction (sign) of biomass change for 77% of intervals, with directional accuracy of 0.88 for gains and 0.48 for losses. Additional classification metrics further quantified this asymmetry. Loss detection showed a precision of 0.57, a recall of 0.48, and an F1-score of 0.52, with a balanced accuracy of 0.68 and a false negative rate of 0.52. The confusion matrix indicated that 1363 observed loss intervals were misclassified as gains, whereas 1265 loss intervals were correctly identified, highlighting that omission of loss events remains the primary limitation. This asymmetry suggests that the ensemble more reliably identifies biomass gains than losses. Loss events are captured less consistently, which may reflect their lower frequency, and greater variability among disturbance processes. In addition, this limitation likely reflects the reduced sensitivity of optical predictors to abrupt disturbance-driven biomass losses compared to more gradual gain processes [40]. Residual diagnostics (Figure 11) indicate that errors are centred near zero for moderate ΔAGB values, with a tendency to underestimate the magnitude of large losses and slightly overestimate strong gains (Figure 11a), consistent with regression toward the mean. Residuals show no systematic bias with census interval length over 5–20 years (Figure 11b), indicating that the weighting scheme and explicit inclusion of interval duration effectively control duration-related biases and support subsequent aggregation of plot-level predictions to the ecodistrict scale.

Comparative model performance is summarized in

Table 3. Comparison of models for plot-level ΔAGB prediction; Overall Direction Accuracy (ODA).

Model	R2	RMSE	MAE	ODA	Accuracy
					Gain	Loss
Ridge regression	0.29	2.63	1.70	0.75	0.86	0.44
RF	0.38	2.45	1.54	0.77	0.88	0.44
XGBoost	0.40	2.42	1.54	0.77	0.88	0.48

. The comparison shows that the linear Ridge model performs substantially worse than the tree-based models, whereas both Random Forest and XGBoost outperformed Ridge regression, highlighting the importance of nonlinear relationships. XGBoost provided the best overall performance, with modest improvements over Random Forest in R², RMSE, and loss-detection accuracy.

Across the 85 ecodistricts with valid remote-sensing-derived predictors, the XGBoost model predicts that biomass dynamics between the early (2000–2012) and late (2012–2023) periods are dominated by net gains (Figure 12). Using a conservative threshold of |${\mathrm{\Delta }AGB}_{yr}$| > 0.05 t ha⁻¹ yr⁻¹ to avoid over-interpreting near-zero model outputs, 61 ecodistricts are classified as experiencing biomass gains, with a mean ${\mathrm{\Delta }AGB}_{yr}$ of 0.44 t ha⁻¹ yr⁻¹, and a maximum of 1.10 t ha⁻¹ yr⁻¹. Eleven ecodistricts exhibit net losses, with mean ${\mathrm{\Delta }AGB}_{yr}$= −0.17 t ha⁻¹ yr⁻¹ and a range from −0.29 to −0.06 t ha⁻¹ yr⁻¹, while a further 13 ecodistricts fall within the no change band (|${\mathrm{\Delta }AGB}_{yr}$| ≤ 0.05 t ha⁻¹ yr⁻¹) and have near-zero mean change (~0.01 t ha⁻¹ yr⁻¹). Spatially, positive ${\mathrm{\Delta }AGB}_{yr}$ is widespread across most of Ontario, except for some western ecodistricts, and is also prevalent in southern Quebec. In contrast, biomass losses are more spatially localized and occur in a limited number of ecodistricts, particularly in parts of western Ontario (e.g., 372, 384, 395) and eastern Québec (e.g., 436, 444), as well as in northern Québec (e.g., 288, 289, 291, 292, 308). These localized losses may be associated with non-fire disturbances such as harvesting, insect outbreaks, windthrow, or stand-level mortality. Predictions for ecodistricts with sparse or no plot coverage should be interpreted with greater caution, as they rely more heavily on model extrapolation.

4. Discussion

4.1. Disturbance Regimes and Interpretation of Biomass Gain and Loss

Disturbance-composition analysis, based on ecodistrict-level aggregation of plot-derived ΔAGB intervals and Landsat-based disturbance classifications [68], indicates that non-fire biomass dynamics are strongly dominated by defoliation, with comparatively smaller contributions from harvest and windthrow (Figure 13a,b and Figure A7). Defoliation accounts for 76.8%, 74.4%, and 92.3% of total non-fire disturbance proportion in the gain, no change, and loss classes, respectively, in the plot-based framework (Figure 13a), increasing to 87.4%, 93.2%, and 94.4% in the Landsat-based framework (Figure 13b). The consistently high contribution of defoliation in the loss class across both frameworks suggests that biomass decline in our study area is primarily associated with defoliation-driven disturbance regimes. This interpretation is consistent with previous Canadian forest-health studies showing that insect defoliation can cause cumulative canopy stress, growth reduction, mortality, and changes in forest carbon dynamics [69]. In contrast, harvest shows its highest contribution in the gain class and declines toward the loss class, indicating that harvesting was not consistently associated with long-term biomass decline at the ecodistrict scale, likely due to regeneration and spatial averaging effects. Windthrow remains a minor component across all classes. These results further suggest that Landsat time series may preferentially capture spatially coherent, multi-year canopy stress signals, whereas plot-based observations remain more sensitive to local heterogeneity and stand-level management effects.

In contrast, harvest shows its highest contribution in the gain class and declines toward the loss class, indicating that harvesting does not necessarily lead to long-term biomass decline at the ecodistrict scale, likely due to regeneration and spatial averaging effects. Windthrow remains a minor component across all classes. These results further suggest that Landsat time series preferentially capture spatially coherent, multi-year canopy stress, whereas plot-based observations retain sensitivity to local heterogeneity and management effects [68,69]. This is ecologically plausible for Quebec and Ontario, where spruce budworm and other defoliators drive widespread canopy stress.

This disturbance-driven interpretation also helps explain the asymmetric model performance observed between biomass gains and losses. While gains are detected more reliably, losses are more difficult to predict, likely due to their spatial heterogeneity, episodic nature, and weaker or less consistent expression in optical spectral signals [25,40]. This asymmetry is consistent with previous studies showing that AGB loss is inherently more difficult to predict than gain because biomass losses are often driven by stochastic and multifactorial processes, including mortality, partial disturbance, insect outbreaks, and climate stress [25]. In contrast, biomass gains tend to follow more gradual and predictable growth trajectories, resulting in higher model skill.

It is also important to note that not all ecodistricts classified as losses in the Landsat-based results are equally supported by field observations. A comparison among the plot-based map (Figure 9), the Landsat-based prediction map (Figure 12), and the disturbance composition results (Figure A7) indicates that several northern ecodistricts (e.g., 288, 289, 291, 292, and 308), where plot coverage is limited or absent, were classified as losses primarily through model extrapolation rather than direct plot support. Consistent with this, Figure A7 shows that some of these ecodistricts exhibit little or no detectable non-fire disturbance signal, suggesting a potential mismatch between predicted biomass decline and observed disturbance patterns. In contrast, ecodistricts such as 445–447, which were also classified as losses in the plot-based results, are better supported by field data and therefore provide more robust evidence of biomass decline. These discrepancies highlight the need for caution when interpreting loss patterns in sparsely sampled regions and emphasize the importance of improved field coverage and validation in northern ecodistricts.

These patterns indicate that disagreement between plot-based and Landsat-based ΔAGB estimates arises from the interaction of sampling intensity, disturbance history, and spatial representativeness.

4.2. Temporal, Spatial, and Methodological Factors Affecting AGB Trend Interpretation

Discrepancies between plot-based and satellite-derived AGB trends are often driven by temporal mismatches in their observation windows. Ground plots typically record biomass change over discrete remeasurement intervals of varying length, whereas Landsat-based analyses summarize dynamics across fixed and temporally consistent epochs. For example, ecodistricts 445 and 447 exhibited statistically significant AGB losses based on plot-based results (Figure 9) during approximately 2000–2014. However, Landsat-based estimates (Figure 12) aggregated over the early (2000–2012) and late (2012–2023) periods instead indicated a small gain and no change, respectively. Such differences can arise because plot networks may capture disturbance phases (e.g., harvest or insect outbreaks) without fully representing subsequent recovery, whereas satellite composites integrate both decline and regrowth processes over broader time windows. Previous studies have similarly shown that asynchronous comparison periods can introduce systematic bias and generate apparent inconsistencies in biomass trend assessments [34,70].

Spatial support differences between field plots and satellite observations provide a second source of discrepancy. Permanent sample plots typically represent areas of approximately 400 m², whereas a single Landsat pixel covers about 900 m². Consequently, localized disturbances or recovery processes detected within individual plots may be diluted when averaged to the pixel scale and further smoothed during aggregation to the ecodistrict level. This mismatch is particularly relevant in heterogeneous forest landscapes, where fine-scale mortality, partial disturbance, or patchy regeneration may strongly influence plot measurements but remain muted in coarser satellite summaries. As a result, plot-based and remotely sensed estimates may differ even when both accurately characterize biomass dynamics at their respective spatial supports [38,39].

A further source of discrepancy may arise from uneven plot representation within large ecodistricts. Where permanent sample plots are clustered in accessible, productive, or less disturbed portions of an ecodistrict, plot-based summaries may not fully capture the broader spatial mosaic of disturbance and recovery conditions. This issue may help explain pronounced inconsistencies in ecodistrict 372, where plot-derived estimates indicated biomass gain, whereas Landsat-based results suggested biomass decline. In such cases, field plots may overrepresent intact or recovering stands while underrepresenting disturbed areas occurring elsewhere within the ecodistrict. Consequently, discrepancies between plot-based and remotely sensed estimates may reflect sampling representativeness as much as true ecological disagreement.

A further source of mismatch may arise from preprocessing choices applied to the Landsat predictors. In our workflow, plot-level spectral values were derived from two-year July–August composites and a 3 × 3 focal-median window to reduce noise, cloud contamination, and data gaps. While this preprocessing improves radiometric stability and reduces the influence of anomalous individual pixels, it may also attenuate highly localized disturbance or recovery signals at individual plot locations by incorporating information from neighbouring pixels and adjacent years. As a result, abrupt stand-level changes detected in field measurements may appear weaker in the remotely sensed predictors. This scale-dependent smoothing effect is consistent with known challenges in linking plot observations to landscape-level AGB estimates [38,39].

In addition to temporal and spatial mismatches, spectral limitations inherent to optical remote sensing further constrain the detection of ΔAGB. In high-biomass forests, vegetation indices often exhibit saturation effects, showing limited sensitivity to continued biomass accumulation despite measurable increases at the plot level [71,72]. Similarly, disturbance- and moisture-sensitive indices may capture initial biomass loss but may fail to fully represent gradual post-disturbance recovery [73,74]. These limitations highlight the need for multi-index and non-linear modelling approaches when interpreting biomass dynamics from satellite data.

Taken together, these sources of discrepancy do not imply that either framework is inherently flawed; rather, they reflect expected differences in temporal coverage, spatial support, sampling representativeness, preprocessing choices, and spectral sensitivity between field and satellite observations.

4.3. Spectral Vegetation Indices (VIs) as Imperfect but Complementary Proxies of AGB Change

Landsat-derived vegetation indices provide indirect proxies of ΔAGB, but their capacity to capture structural dynamics is limited. As shown in Section 3.3, greenness indices such as NDVI and EVI exhibited relatively lower correlations with ΔAGB, consistent with previous findings that photosynthetic proxies often fail to reflect changes in woody biomass over decadal periods [71,75]. In contrast, indices derived from the NIR and SWIR regions, including NDII, NBR, MSI, and SATVI, showed more pronounced relationships with ΔAGB than greenness-based indices [76], which likely makes them more informative for capturing multi-year biomass dynamics at the ecodistrict scale. However, index performance varied among ecodistricts, echoing broader findings that no single spectral index provides universally robust predictions across forest types [77,78].

Hexagonal binned scatterplots (Figure A6) reveal dense elliptical point clouds centred near zero change, with substantial overlap among gain, loss, and no change classes, confirming that spectral–biomass relationships are nonlinear and confounded by multiple drivers [79]. To illustrate these relationships more directly at the plot level, Figure 14a,b presents two representative examples, NDVI (Figure 14a) and NDII (Figure 14b), based on 100 randomly selected non-fire plot intervals for visual illustration only. For each sampled interval, Landsat July–August annual medians were used to calculate Theil–Sen slopes of spectral indices over the interval period, which were then compared against observed annual ΔAGB. These illustrative plots are not intended to replace the full correlation analysis, but rather to clarify how spectral trajectories align with plot-level biomass change classes. NDVI shows substantial overlap among gain, loss, and no change classes, indicating that greenness trends alone have limited ability to distinguish the direction of biomass change. By contrast, NDII provides clearer directional separation, with positive slopes more frequently associated with biomass gains and near-zero or negative slopes more often associated with losses.

However, these patterns should not be interpreted deterministically. This does not imply that NDII uniquely identifies gain or loss conditions; rather, it indicates that indices sensitive to canopy moisture and NIR–SWIR contrast are generally more informative for representing the direction of ΔAGB change than greenness-only metrics. More broadly, the additional charts for NBR, MSI, and SATVI show the same general tendency, although with varying degrees of overlap and dispersion. These results indicate that spectral–biomass relationships are nonlinear and influenced by multiple interacting factors, including stand structure, recovery stage, disturbance legacy, and compositional change [79]. Consequently, spectral indices should be interpreted as complementary indicators of different ecological processes rather than as direct one-to-one measures of biomass change. Their greatest value emerges when they are integrated within a multi-index, non-linear modelling framework. This interpretation is consistent with [20], who showed that spectral greenness trends in boreal forests are often influenced by disturbance and post-disturbance recovery dynamics. Together, these results suggest that integrating plot-based measurements with remote sensing is essential to disentangle structural biomass change from spectral signals, which may otherwise be confounded by disturbance–recovery processes.

4.4. Study Limitations and Priorities for Future Work

Our analysis is constrained by 7 ecodistricts (e.g., 364, 369, 384, 394, 395, 396, 1028) that currently have only a single remeasurement period, corresponding to regions classified as “insufficient data” (grey) in Figure 9b, which provides a temporal snapshot rather than robust multi-year ΔAGB estimates. Continued remeasurement of existing plots in these ecodistricts is therefore required before they can fully support long-term trend analyses. In addition, some ecodistricts that met our minimum data filter remain weakly sampled, resulting in wider CI and reduced precision. Increasing plot numbers and remeasurement frequency in these sparsely sampled units should be a clear priority for future work.

An additional limitation is that our remote-sensing predictors were derived from a single-sensor Landsat-7 time series. As described in Section 2.4, this choice was made to ensure radiometric consistency across the study period and to mitigate Landsat-7 SLC-off artefacts using multi-temporal compositing and spatial filtering. While reliance on a single sensor may reduce temporal data density compared to harmonized multi-sensor approaches, it avoids potential biases introduced by cross-sensor differences in spectral response and calibration [25]. Importantly, because our analysis focuses on relative spectral changes and their relationship with ΔAGB, rather than absolute reflectance values, this choice is unlikely to affect the main conclusions. Nevertheless, future research could extend this framework by incorporating harmonized Landsat 5/7/8/9 time series [20,80,81,82] and additional data sources such as LiDAR (e.g., GEDI) and SAR-based biomass products to further improve temporal coverage and cross-validation at broader spatial scales [83,84,85].

5. Conclusions

This study demonstrates that integrating forest permanent sample plots with Landsat time series can provide a measurement-constrained framework for assessing non-fire-related AGB change at the ecodistrict scale in the eastern Canadian boreal forest. Across 59 plot-supported ecodistricts (63% of the study area), biomass dynamics are dominated by gains: 37 ecodistricts, accounting for 55% of the plot-supported area, exhibit significant increases, while 20 ecodistricts (39% of this area show no significant change and only 2 ecodistricts exhibit losses. These plot-based results indicate that, in the absence of wildfire, a substantial portion of the boreal forest functions as a net biomass sink, with additional potential for carbon uptake in ecodistricts classified as no change, where targeted management of non-fire disturbance processes (e.g., harvesting, insect outbreaks, and windthrow), could shift these systems toward net biomass gains. At the full spatial extent, the Landsat-based wall-to-wall assessment suggests a predominance of biomass gains. A total of 61 ecodistricts (71.8%) are classified as gains, while 11 ecodistricts (12.9%) show losses, and 13 ecodistricts (15.3%) exhibit no detectable change. Disturbance composition analyses further indicated that biomass declines were more commonly associated with defoliation-dominated disturbance regimes, whereas harvested regions often exhibited neutral or recovering longer-term biomass trajectories. However, predicted losses in sparsely sampled northern ecodistricts should be interpreted cautiously because they rely more heavily on model extrapolation than direct plot support.

Strong data-density effects on confidence intervals and systematic differences between plot-based and satellite-derived indicators underscore the importance of sampling design, spatial representativeness, and scale when extrapolating from plots to landscapes. Moisture- and disturbance-sensitive indices carried the clearest spectral signal of ΔAGB, and the XGBoost ensemble captured biomass gains more consistently than losses, but severe losses remained under-predicted, reflecting persistent challenges in modelling low-frequency, high-impact disturbance processes.

Overall, the convergence between plot-based and Landsat-derived evidence indicates that much of the eastern Canadian boreal forest exhibits either no significant change or increasing biomass trajectories. The resulting gain–loss–no change framework provides a transparent, spatially explicit basis for greenhouse-gas reporting and for prioritizing monitoring and management in ecodistricts with persistent departures from stability. Future work should extend this framework to harmonized multi-sensor time series, improve representation in data-sparse regions, and explicitly attribute ΔAGB to specific non-wildfire disturbance agents, while complementing these analyses with wildfire-driven dynamics.