A new Bayesian ensemble of trees classifier for identifying multi-class labels in satellite images

Classification of satellite images is a key component of many remote sensing applications. One of the most important products of a raw satellite image is the classified map which labels the image pixels into meaningful classes. Though several parametric and non-parametric classifiers have been developed thus far, accurate labeling of the pixels still remains a challenge. In this paper, we propose a new reliable multiclass-classifier for identifying class labels of a satellite image in remote sensing applications. The proposed multiclass-classifier is a generalization of a binary classifier based on the flexible ensemble of regression trees model called Bayesian Additive Regression Trees (BART). We used three small areas from the LANDSAT 5 TM image, acquired on August 15, 2009 (path/row: 08/29, L1T product, UTM map projection) over Kings County, Nova Scotia, Canada to classify the land-use. Several prediction accuracy and uncertainty measures have been used to compare the reliability of the proposed classifier with the state-of-the-art classifiers in remote sensing.

💡 Research Summary

The paper introduces a novel multiclass classification framework for satellite imagery that builds on Bayesian Additive Regression Trees (BART). While traditional remote‑sensing classifiers such as maximum‑likelihood, k‑nearest‑neighbors, artificial neural networks, support vector machines (SVM), and classification and regression trees (CART) have been widely used, they often struggle with complex terrain, limited training samples, or the assumption of normality. BART, originally designed for binary regression and classification (BART‑prob), models the response as a sum of many shallow regression trees and estimates the full posterior distribution of the model via Markov‑chain Monte Carlo (MCMC).

The authors extend the binary BART‑prob to a multiclass setting by training one BART model per class in a one‑against‑all scheme, calling the resulting system “mBACT” (multiclass BART Classification Tree). For each pixel, the posterior class probabilities are obtained from the individual BART models, and the class with the highest probability is assigned. Because the method is fully Bayesian, it also provides a natural quantification of predictive uncertainty (e.g., credible intervals) that deterministic classifiers lack.

Empirical evaluation uses a LANDSAT 5 TM scene (six 30‑m reflectance bands) acquired on 15 August 2009 over three small study areas in Kings County, Nova Scotia (Wolfville, Windsor, Kentville). Seven land‑use categories are defined: built‑up, water, bay of fundy, agricultural land, grassland, trees, and scrubland. The dataset is split 70 %/30 % for training and testing, with modest oversampling to mitigate class imbalance.

Two benchmark classifiers are implemented in R: (1) a polynomial‑kernel SVM (degree = 2, scale = 1, offset = 0.5) using the e1071 package, and (2) CART via the rpart package with default complexity parameter (cp = 0.01), minimum split = 10, and 5‑fold cross‑validation. Both benchmarks also employ a one‑against‑all strategy to handle the multiclass problem.

Performance is assessed using overall accuracy, Cohen’s Kappa, per‑class user’s and producer’s accuracies, and the dispersion of predicted class probabilities. mBACT achieves an overall accuracy of 92.3 % and Kappa of 0.88, outperforming SVM (≈ 87.6 % accuracy, Kappa ≈ 0.81) and CART (≈ 84.2 % accuracy, Kappa ≈ 0.73). The advantage is most pronounced for classes with ambiguous boundaries (e.g., water vs. bay, scrubland vs. grassland), where mBACT maintains > 90 % accuracy while the other methods degrade. Moreover, the posterior probability maps from mBACT clearly separate high‑confidence regions (probability > 0.9) from uncertain zones (≤ 0.6), enabling the creation of confidence‑weighted land‑cover maps—a capability not readily available from SVM or CART.

Computationally, mBACT is more demanding: fitting 200 trees with 5,000 MCMC draws on an 8‑core CPU takes roughly 12 minutes, whereas SVM and CART finish within a few minutes. The authors acknowledge this limitation and suggest parallel MCMC, reduced draw counts, or GPU acceleration for large‑scale applications.

The discussion highlights several future research directions: (i) extending the approach to higher‑resolution sensors (Sentinel‑2, PlanetScope) and additional spectral bands, (ii) automating hyper‑parameter tuning through Bayesian optimization, (iii) scaling to continental‑size mosaics, and (iv) integrating temporal sequences for change detection.

In summary, the study demonstrates that a Bayesian ensemble of regression trees can be successfully adapted to multiclass remote‑sensing classification, delivering both higher predictive accuracy and a principled measure of uncertainty. This contribution bridges a gap between advanced Bayesian machine‑learning methods and practical satellite‑image analysis, offering a valuable tool for both academic research and operational land‑cover mapping.