EVALUATION OF TREE SPECIES CLASSIFICATION METHODS USING MULTI-TEMPORAL SATELLITE IMAGES

Arnav Saha, Srikumar Sastry, Viral Dave, Rajib Ghosh2020 · DOI

TL;DR

Researchers have developed a highly accurate method for identifying tree species from space by tracking their "seasonal fingerprints." By combining Gaussian Mixture Models (GMM) for clustering with Dynamic Time Warping (DTW) to analyze multi-temporal NDVI profiles, the team achieved over 94% accuracy in classifying tropical deciduous forests, significantly outperforming traditional pixel-based methods.

What Problem Does This Paper Address?

Monitoring forest health and biodiversity is critical for conservation, but doing it on the ground is a monumental task. While satellite imagery offers a "bird's-eye view," identifying specific tree species from a single snapshot is notoriously difficult. Many species look nearly identical in a single multispectral image.

The core challenge is phenology—the study of cyclic biological events. Different trees have different schedules for shedding leaves, flowering, and fruiting. A single image misses these temporal signatures. This paper addresses the gap in high-accuracy species classification by leveraging multi-temporal data to capture how different trees change over several months.

Methods

The researchers proposed a workflow that moves beyond simple pixel-by-pixel classification:

  1. Data Stacking: They used multispectral images (R, G, B, and NIR) collected over four consecutive months (December to March). These were stacked into a 16-band composite image, where each pixel contained a time-series of spectral data.
  2. Clustering (The Grouping Phase): They compared K-Means and Gaussian Mixture Models (GMM) to group similar pixels. GMM proved superior because it is "soft" and flexible, handling non-spherical data distributions better than the rigid K-Means.
  3. Segmentation: They used Mean Shift segmentation for an object-based approach, which groups neighboring pixels into "objects" (like a stand of trees) to reduce noise and retain spatial context.
  4. Dynamic Time Warping (DTW): This is the paper's standout technique. DTW is an algorithm typically used in speech recognition to compare two sequences that may vary in speed or time. Here, it was used to match the observed NDVI (Normalized Difference Vegetation Index) profile of a cluster against "pure" ground-truth profiles of specific species (Teak, Sadad, and Mixed forest).
  5. Comparison: The DTW approach was benchmarked against standard supervised learners: Random Forest (RF) and Decision Trees (DT).

Datasets

The study focused on Ahwa village in the Dang district of Gujarat, India, a hilly region dominated by Tropical Moist Deciduous forests.

  • Satellite Data: High-resolution (5m) multispectral temporal images from four dates (Dec 2018, Jan 2019, Feb 2019, and March 2019).
  • Ancillary Data: A forest mask from the Forest Survey of India (FSI) was used to filter out non-forest areas like roads and crops.
  • Ground Truth: Six ground truth samples were used to generate "pure" profiles for three classes: Teak (Tectona grandis), Sadad (Terminalia tomentosa), and Mixed forest.

Key Insights and Ideas

  • Phenology as a Bio-Signature: The "aha moment" is that Teak and Sadad have distinct seasonal behaviors. Teak is a deciduous species that sheds leaves in a specific window, while Sadad has a unique ability to hold water in its bark during dry seasons. These biological traits manifest as distinct curves in NDVI plots over time.
  • DTW's Robustness: Traditional classifiers often struggle if a season is slightly delayed due to weather. Because DTW measures the shape of the temporal curve rather than just comparing values at fixed timestamps, it is much more resilient to year-to-year climatic shifts.
  • Object-Based > Pixel-Based: The results reinforce that considering a group of pixels (an "object") provides better results than looking at pixels in isolation, as it filters out "salt and pepper" noise.

Key Results

The GMM + DTW model was the clear winner, accurately reflecting the actual species distribution found on the ground.

Model Overall Accuracy (OA) Kappa Score
GMM + DTW 94.54% 0.88
Object-Based + RF 87.67% 0.72
Object-Based + DTW 91.11% 0.81
Pixel-Based + DT 85.91% 0.69

The GMM+DTW combination achieved a high Kappa score (0.88), indicating a very strong agreement between the model's predictions and the actual ground truth, far exceeding the performance of standard Random Forest or Decision Tree implementations.

Limitations and Future Work

While the results are impressive, the paper notes a few areas for further exploration:

  • Temporal Window: The study only used four months of data. Expanding this to a full year could capture even more nuanced phenological stages (like flowering seasons in June/August).
  • Scalability: The current method relies on high-resolution (5m) imagery. Future work could investigate if similar results are achievable with free, lower-resolution data like Sentinel-2 (10m).
  • Species Diversity: The study focused primarily on Teak and Sadad. Testing the model in more biodiverse "Mixed Forest" zones with dozens of species would be the ultimate stress test for the DTW approach.

Summary updated

Abstract

Tree species classification is an important step towards forest monitoring and biodiversity conservation. This research study evaluates several multispectral image classification techniques for tree species over Ahwa village in Dang district, South Gujarat, India. Multispectral images consisting of 4 bands-R, G, B and NIR collected over 4 months was used. Object-based segmentation using mean shift, cluster-based using K-Means and Gaussian Mixture Model (GMM) and pixel-based methods have been analyzed. Additionally, a new method of classification has been described using the Dynamic Time Warping (DTW) algorithm. It outperformed supervised classification techniques with accuracy over 95%. The GMM+DTW model accurately reflected the actual species distribution found in the ground truth.
BibTeX
@inproceedings{saha2020evaluation,
  author = {Arnav Saha and Srikumar Sastry and Viral Dave and Rajib Ghosh},
  title = {{EVALUATION OF TREE SPECIES CLASSIFICATION METHODS USING MULTI-TEMPORAL SATELLITE IMAGES}},
  booktitle = {˜The},
  year = {2020},
  doi = {10.5194/isprs-archives-xlii-3-w12-2020-79-2020},
}
Figure