Automatic deforestation detectors based on frequentist statistics and their extensions for other spatial objects

TL;DR

This paper introduces two novel statistical methods for detecting forest and non-forest areas in satellite images, including a parametric approach that extends to natural object detection and anomalies, with practical implementation and comparison to machine learning.

Contribution

The paper presents a new parametric statistical method for spatial object detection, extending the scope of traditional approaches and providing algorithms with practical implementation details.

Findings

The parametric method is novel and effective for natural object detection.

Algorithms outperform some standard machine learning methods in certain scenarios.

The methods are applicable to a broader class of spatial detection problems.

Abstract

This paper is devoted to the problem of detection of forest and non-forest areas on Earth images. We propose two statistical methods to tackle this problem: one based on multiple hypothesis testing with parametric distribution families, another one -- on non-parametric tests. The parametric approach is novel in the literature and relevant to a larger class of problems -- detection of natural objects, as well as anomaly detection. We develop mathematical background for each of the two methods, build self-sufficient detection algorithms using them and discuss practical aspects of their implementation. We also compare our algorithms with those from standard machine learning using satellite data.