Exploring Interpretability in Deep Learning-Based Forest Classification - Separating Similar Forest Classes using Sentinel-2 Time Series

Maximilian Kirsch; Christine Preisach

cs.LGOct 13, 2026

Exploring Interpretability in Deep Learning-Based Forest Classification - Separating Similar Forest Classes using Sentinel-2 Time Series

Authors: Maximilian Kirsch, Christine Preisach

Organizations: Intelligent Systems Research Group (ISRG), Karlsruhe University of Applied Sciences, Germany

Abstract

The ecological insurance theory posits that diverse forests are more resilient and better able to adapt to changing environmental conditions. Increasing drought and heat stress in Central Europe promotes mixed forests for maintaining ecosystem services. However, a diverse species distribution complicates remote sensing-based species mapping, as distinguishing pure from mixed species compositions at the pixel level remains challenging. We evaluated the potential of recurrent neural networks (RNNs) for tree-species classification from Sentinel-2 time series and investigated the spectral-temporal features underlying class discrimination, with particular emphasis on separating pure from mixed Norway spruce pixels. We compared RNN classification models with Random Forest (RF) and Extreme Gradient Boosting (XGBoost) and used Integrated Gradients (IG) to quantify feature attributions across spectral bands and time. The best-performing RNN achieved a Macro-F1 score of 0.851 on the test data, outperforming XGBoost by 0.012 and Random Forest by 0.103. Beyond classification performance, the attribution analysis revealed distinct spectral-temporal patterns associated with the separation of pure and mixed Norway spruce pixels. Attribution values increased particularly after approximately day of year (DOY) 125, while B11 showed the strongest contribution during periods of increasing spectral divergence between the two classes. These findings show that recurrent models can provide competitive classification performance while also revealing the phenological periods and spectral regions that are most relevant for distinguishing closely related forest compositions. Combining time-series classification with explainability therefore provides additional insight into the spectral-temporal mechanisms underlying tree-species discrimination and may support more detailed mapping of mixed forest structures.

Figures & tables

Figure 1: Data distribution : spatial distribution of sample locations used in this study, grouped by species.

Figure 2: Mixed Sentinel-2 pixels. Illustration of Sentinel-2 pixels containing pure Norway spruce stands and mixed stands. A shows an orthophoto of a forest structure including Norway spruce and varying deciduous trees. B shows a subset of the same area including the outlines of deciduous trees within a 3 $\times$ 3 Sentinel-2 pixel grid. C shows the corresponding Sentinel-2 pixels averaged over one year (2021).

Figure 3: LSTM and GRU architectures. At time step $t$ , $x_{t}$ is the multivariate input. The LSTM unit uses forget ( $f_{t}$ ), input ( $i_{t}$ ), and output ( $o_{t}$ ) gates with candidate cell state $\tilde{C}_{t}$ . The GRU unit uses update ( $z_{t}$ ) and reset ( $r_{t}$ ) gates with candidate activation $g_{t}$ . The LSTM carries forward long- and short-term memory separately as cell state $C$ and hidden state $h$ , while GRU only forwards one combined state. Blue and red frames indicate sigmoid ( $\sigma$ ) and $\tanh$ activations, respectively. The recurrent weights $w,u$ and biases $b$ are learned.

Figure 4: GRU-MHA Classifier Architecture. The RNN output is projected into key and value representations, while a learnable global query provides the query representation. After splitting the projected representations into multiple attention heads, scaled dot-product attention is computed between the global query and the keys of all time steps. The resulting attention scores are normalized using softmax and used to calculate a weighted sum of the value vectors, yielding one pooled representation per head. These are connected and used as inputs for the classification head.

Figure 5: Multi-head attention weighting. The top panel shows the temporal profile of spectral band 11 (SWIR 1), which is included for visualization purposes. The panels below show the attention weights assigned to each time step by the individual attention heads. The blue bars in the top panel represent the attention weights aggregated across all heads for each time step. The attention mechanism operates on the latent representations rather than directly on the spectral band values.

Species	Description	Samples
Species	Description	Train	Test
Fagus sylvatica	pure stand	13131	6828
Quercus spp.	pure stand	7738	2244
Pinus sylvestris	pure stand	16702	32484
Picea abies	pure stand	15989	19688
Picea abies	mixed stand	16965	40511

Table 1: Final Sequences Sample Size

Model			Performance metrics
RNN	Attention	BN Dim.	Acc	BA	Macro-F1
LSTM	MHA	32	0.832	0.909	0.825
		64	0.827	0.920	0.814
		128	0.848	0.930	0.843
	None	32	0.821	0.912	0.813
		64	0.830	0.923	0.830

Table 2: Comparison of Model Configurations

Panel of Figure 7 — Figure 7: Confusion matrices for the two best-performing RNN models, GRU with and without attention, and the XGBoost model

Tree species	Metric	RF	XGBoost	LSTM-128		GRU-128
Tree species	Metric	RF	XGBoost	MHA	No Att.	MHA	No Att.
Norway spruce	F1	0.614	0.677	0.677	0.684	0.691	0.689
	Sensitivity	0.704	0.759	0.731	0.742	0.763	0.747
	Specificity	0.895	0.914	0.917	0.918	0.914	0.919
Norway spruce mixed	F1	0.697	0.774	0.775	0.782	0.784	0.784
	Sensitivity	0.610	0.707	0.717	0.729	0.723	0.728

Table 3: Per-class comparison of model configurations

Panel of Figure 8 — Figure 8: Map with distribution statistics of predicted forest classes.

Band Identifier	Band Name	Central Wavelength (nm)
B2	Blue	492.7
B3	Green	559.8
B4	Red	664.6
B5	Red Edge 1	704.1
B6	Red Edge 1	740.5
B7	Red Edge 1	782.8

Table 4: Sentinel-2 Level-2A bands used in this study.

Panel of Figure 9 — Figure 9: Integrated Gradients attribution results for the six investigated classes against a zero-baseline.

Figure 10: Attribution map for the separation of pure and mixed Norway spruce.

Figure 11: Attribution Map for the separation of beech and oak

Figure 12: Separation between pure and mixed Norway spruce. The top row shows the median test-set time series with confidence bands. The bottom row displays Integrated Gradients attributions along the Norway spruce median time series, highlighting time steps relevant for distinguishing the two species.

Panel of Figure 14 — Figure 14: False Predictions over years: Here the mean percentage $\pm$ SD of false predictions for each class is visualized for every observed year across all RNN models. Norway spruce is abbreviated to NWS . The disturbed Norway spruce class has no observations for the years 2021 and 2022, as there was no data where the disturbance year was after 2020.

Panel of Figure 15 — Figure 15: Mean percentage of false predictions ( $\pm$ SD) for pure and mixed Norway spruce combined, grouped by the number of time steps per time series.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Model	RF	XGBoost	LSTM		GRU
Model	RF	XGBoost	MHA	No Att.	MHA	No Att.
Accuracy	0.792	0.851	0.848	0.850	0.854	0.853
Balanced accuracy	0.890	0.925	0.930	0.931	0.933	0.933
Macro F1	0.748	0.839	0.843	0.843	0.851	0.847

Appendix

Table 5: Comparison of model configurations

Model	Hyperparameter	Tested values
Random Forest	n_estimators	100, 250, 500, 750
	max_depth	3, 6, 10, 15
	min_samples_split	2, 5, 10
	min_samples_leaf	1, 2, 5
	max_features	sqrt , log2 , 0.5, 0.7
	bootstrap	True

Appendix

Table 6: Hyperparameter search spaces for Random Forest and XGBoost models

Interpretable Time-Series Classification

See all 16 papers

Sep 3, 2026

Tree species mapping in Denmark: A comparison of spectral-temporal features with geospatial foundation model embeddings

We map tree species across Denmark using National Forest Inventory plots and EO data, while evaluating the potential of foundation models for large-scale forest characterization. We compare two alternative input representations for tree species classification: (i) manually engineered spectral-temporal features (STF) derived from multi-temporal Sentinel-1 and Sentinel-2 observations, and (ii) embeddings generated by the EO FMs TESSERA and AlphaEarth. Both representations are complemented with canopy height information. Random forest, XGBoost, and Multi-Layer Perceptron (MLP) classifiers are evaluated for all input representations, with separate assessments for pure and mixed forest stands. The STF-based MLP achieves the highest classification performance, yielding macro F1 scores of 0.843 and 0.653 for pure and mixed stands, respectively. The MLP trained on TESSERA embeddings delivers competitive performance for pure stands, achieving results within 1.1 percentage points of the best-performing model. TESSERA consistently outperforms STF-based models when fewer than approximately 25% of training plots are available, demonstrating a substantial advantage under limited training data. Multi-year observations systematically improve classification accuracy relative to single-year inputs, while ablation experiments reveal the complementary contributions of Sentinel-1 backscatter, spectral indices, and canopy height data. The best-performing model is subsequently applied at the national scale to generate a 10 m tree species map of Denmark. Area-adjusted validation indicates an overall map accuracy of 79.9%. The resulting map, released as an open-access product, is the first high-resolution national tree species map of Denmark and provides a valuable resource for forest monitoring, ecological research, and land management applications.

Alkiviadis Koukos, Spyros Kondylatos, Thomas Nord-Larsen +3

EO Centre of Excellence, DHI, Agern Alle 5, Hørsholm, 2970, Denmark · Department of Geoscience and Natural Resource Management, University of Copenhagen, Øster Voldgade 10, Copenhagen, 1350, Denmark · ceometrics, Strandvejen 273C, Charlottenlund, 2920, Denmark

Geospatial Foundation Models Geospatial Representation Learning Tree Species Classification Remote Sensing Multi-Angular Remote Sensing

Ask PDF

May 7, 2026

A Novel Graph-Regulated Disentangling Mamba Model with Sparse Tokens for Enhanced Tree Species Classification from MODIS Time Series

Although tree species classification from Moderate Resolution Imaging Spectroradiometer (MODIS) time series data is critical for supporting various environmental applications, it is a challenging task due to several key difficulties: the subtle signature differences among tree species, strong spatial-spectral-temporal information coupling, and the difficulty of modeling large-scale topological context information. To better address these challenges, this paper presents a novel Graph-regulated Disentangled Sparse Mamba model (GDS-Mamba) for enhanced tree species classification, with the following contributions. (1) First, to improve large-scale context modeling, we design a mini-batch graph-regulated approach that explicitly explores topological correlation effects among input images. (2) Second, to disentangle the high-dimensional spatial-spectral-temporal information coupling for improved feature extraction, we propose a novel disentangling Mamba architecture tailored for capturing independent spatial patterns, spectral signatures, and temporal phenology behaviors in MODIS time series. (3) Third, to improve efficiency and subtle feature learning, we design novel sparse token approaches that adaptively learn the optimum subset of tokens to better address the correlation decay problem that bottlenecks standard Mamba models. Extensive experiments using large-scale annual MOD13Q1 data across two Canadian provinces (i.e., Alberta and Saskatchewan) achieved an overall accuracy of 93.94% in Alberta and 80.19% in cross-provincial evaluations, outperforming twelve state-of-the-art classification models.

Motasem Alkayid, Zhengsen Xu, Saeid Taleghanidoozdoozan +7

Department of Geomatics Engineering, University of Calgary, Canada

Remote Sensing Image Understanding Mamba Disentangled Representation Learning Time Series Classification Tree Species Classification

Ask PDF

Apr 17, 2026

From Articles to Canopies: Knowledge-Driven Pseudo-Labelling for Tree Species Classification using LLM Experts

Hyperspectral tree species classification is challenging due to limited and imbalanced class labels, spectral mixing (overlapping light signatures from multiple species), and ecological heterogeneity (variability among ecological systems). Addressing these challenges requires methods that integrate biological and structural characteristics of vegetation, such as canopy architecture and interspecific interactions, rather than relying solely on spectral signatures. This paper presents a biologically informed, semi-supervised deep learning method that integrates multi-sensor Earth observation data, specifically hyperspectral imaging (HSI) and airborne laser scanning (ALS), with expert, ecological knowledge. The approach relies on biologically inspired pseudo-labelling over a precomputed canopy graph, yielding accurate classification at low training cost. In addition, ecological priors on species cohabitation are automatically derived from reliable sources using large language models (LLMs) and encoded as a cohabitation matrix with likelihoods of species occurring together. These priors are incorporated into the pseudo-labelling strategy, effectively introducing expert knowledge into the model. Experiments on a real-world forest dataset demonstrate 5.6% improvement over the best reference method. Expert evaluation of cohabitation priors reveals high accuracy with differences no larger than 15%.

Michał Romaszewski, Dominik Kopeć, Michał Cholewa +6

Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Bałtycka 5, Gliwice, 44-100, Poland · University of Lodz, Narutowicza 68, 90-136, Lodz, Poland · MGGP Aero, Kaczkowskiego 6, 33-100, Tarnów, Poland +1

Tree Species Classification Graph-Based Semi-Supervised Learning Pseudo-Labeling Hyperspectral Image Classification Semi-Supervised Learning

Ask PDF

Abstract

Interpretable Time-Series Classification

Explore similar work

Tree species mapping in Denmark: A comparison of spectral-temporal features with geospatial foundation model embeddings

A Novel Graph-Regulated Disentangling Mamba Model with Sparse Tokens for Enhanced Tree Species Classification from MODIS Time Series

From Articles to Canopies: Knowledge-Driven Pseudo-Labelling for Tree Species Classification using LLM Experts