Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Satellite Band Selection Framework for Amazon Forest Deforestation Detection Task

Eduardo Neto, Fabio A. Faria, Amanda A. S. de Oliveira, Álvaro L. Fazenda

TL;DR

The paper addresses Amazon deforestation detection by selecting an optimal subset of Landsat-8 spectral bands via UMDA to improve both segment classification and semantic segmentation performance. The framework uses Haralick texture features, SLIC superpixels, SVM fitness, and a DeepLabv3+ segmentation model, demonstrating that carefully chosen band subsets can outperform using all bands or commonly used baselines. Key findings show that four-band combinations derived from UMDA (e.g., B4B3B1B6 or B4B3B1B7) can surpass All+NDVI baselines, achieving higher IoU and balanced accuracy while reducing data dimensionality. This challenges the notion that more spectral information always yields better results and points to practical gains in efficiency and generalization for large-scale forest monitoring.

Abstract

The conservation of tropical forests is a topic of significant social and ecological relevance due to their crucial role in the global ecosystem. Unfortunately, deforestation and degradation impact millions of hectares annually, necessitating government or private initiatives for effective forest monitoring. This study introduces a novel framework that employs the Univariate Marginal Distribution Algorithm (UMDA) to select spectral bands from Landsat-8 satellite, optimizing the representation of deforested areas. This selection guides a semantic segmentation architecture, DeepLabv3+, enhancing its performance. Experimental results revealed several band compositions that achieved superior balanced accuracy compared to commonly adopted combinations for deforestation detection, utilizing segment classification via a Support Vector Machine (SVM). Moreover, the optimal band compositions identified by the UMDA-based approach improved the performance of the DeepLabv3+ architecture, surpassing state-of-the-art approaches compared in this study. The observation that a few selected bands outperform the total contradicts the data-driven paradigm prevalent in the deep learning field. Therefore, this suggests an exception to the conventional wisdom that 'more is always better'.

A Satellite Band Selection Framework for Amazon Forest Deforestation Detection Task

TL;DR

The paper addresses Amazon deforestation detection by selecting an optimal subset of Landsat-8 spectral bands via UMDA to improve both segment classification and semantic segmentation performance. The framework uses Haralick texture features, SLIC superpixels, SVM fitness, and a DeepLabv3+ segmentation model, demonstrating that carefully chosen band subsets can outperform using all bands or commonly used baselines. Key findings show that four-band combinations derived from UMDA (e.g., B4B3B1B6 or B4B3B1B7) can surpass All+NDVI baselines, achieving higher IoU and balanced accuracy while reducing data dimensionality. This challenges the notion that more spectral information always yields better results and points to practical gains in efficiency and generalization for large-scale forest monitoring.

Abstract

The conservation of tropical forests is a topic of significant social and ecological relevance due to their crucial role in the global ecosystem. Unfortunately, deforestation and degradation impact millions of hectares annually, necessitating government or private initiatives for effective forest monitoring. This study introduces a novel framework that employs the Univariate Marginal Distribution Algorithm (UMDA) to select spectral bands from Landsat-8 satellite, optimizing the representation of deforested areas. This selection guides a semantic segmentation architecture, DeepLabv3+, enhancing its performance. Experimental results revealed several band compositions that achieved superior balanced accuracy compared to commonly adopted combinations for deforestation detection, utilizing segment classification via a Support Vector Machine (SVM). Moreover, the optimal band compositions identified by the UMDA-based approach improved the performance of the DeepLabv3+ architecture, surpassing state-of-the-art approaches compared in this study. The observation that a few selected bands outperform the total contradicts the data-driven paradigm prevalent in the deep learning field. Therefore, this suggests an exception to the conventional wisdom that 'more is always better'.
Paper Structure (22 sections, 7 equations, 4 figures, 6 tables)

This paper contains 22 sections, 7 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Fundamental Concepts
  3. Landsat Satellite
  4. PRODES
  5. Superpixels with SLIC
  6. Semantic Segmentation with DeepLabv3$+$
  7. Univariate Marginal Distribution (UMDA)
  8. Homogeneity Rate (HoR)
  9. Evaluation Metrics
  10. Precision:
  11. Recall:
  12. F1-Score:
  13. Accuracy:
  14. Intersection over Union (IoU):
  15. The Landsat Band Selection Framework
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Rondônia's thematic image for the PRODES year of $2016$. Extracted from TerraBrasilis deterrabrasilis. In (a) Thematic image of Rondônia in the year $2016$ and in (b) Legend for the thematic image where "d" means deforestation of the given year and "r" is the residue. The residue is deforestation that was not detected but occurred in previous years macielreasoning.
  • Figure 2: The pipeline of the Landsat-8 band selection framework based on UMDA for deforestation detection is depicted as follows: (a) displays an image from the Landsat-8 imaging satellite with its seven spectral bands, (b) represents the false-color image formed by three principal components of PCA PCA, (c) illustrates the output of the SLIC SLIC superpixel generation with various identified segments, (d) showcases the resulting image from the selection of segments indicated by white dots, (e) illustrates the Haralick HARALICK feature vectors extracted from the selected segments, (f) displays the feature vectors separated into three sets (training, validation, and test) according to their regions for use in training the SVM classifier svm during the evolutionary process of the UMDA algorithm muhlenbein1996recombination, (g) exhibits the best individual found by the UMDA algorithm (bands B4, B6, B1, and B7), (h) is an illustrative example of an image composed by the best band combination. Finally, in (i), the image segmented by the DeepLabv3$+$ architecture chen2018encoderdecoder is shown, where white pixels correspond to deforested regions in the input image.
  • Figure 3: In (a) is the region $3$ from the test set composed of the RGB bands. In (b) is the ground truth of the region $3$. In (c) is the semantic segmentation result of the baseline composition from the paper torres_2021. Finally, in (d), is our the best composition band found by UMDA-based framework.
  • Figure 4: In (a) is the region $4$ from the test set composed of the RGB bands. In (b) is the ground truth of the region $4$. In (c) is the semantic segmentation result of the baseline composition from the paper torres_2021. Finally, in (d), is our the best composition band found by UMDA-based framework.