Temporal difference and density-based learning method applied for deforestation detection using ALOS-2/PALSAR-2

Remote sensing has established as key technology for monitoring of environmental degradation such as forest clearing. One of the state-of-the-art microwave EO systems for forest monitoring is Japan's L-band ALOS-2/PALSAR-2 which provides outstanding means for observing tropical forests due its cloud and canopy penetration capability. However, the complexity of the physical backscattering properties of forests and the associated spatial and temporal variabilities, render straightforward change detection methods based on simple thresholding rather inaccurate with high false alarm rates. In this paper, we develop a framework to alleviate problems caused by forest backscatter variability. We define three essential elements, namely "structures of density", "speed of change", and "expansion patterns" which are obtained by differential computing between two repeat-pass PALSAR-2 images. To improve both the detection and assessing of deforestation, a "deforestation behavior pattern" is sought through temporal machine learning mechanism of the three proposed elements. Our results indicate that the use of "structure of density" can introduce a more robust performance for detecting deforestation. Meanwhile, "speed of change" and "expansion pattern" are capable to provide additional information with respect to the drivers of deforestation and the land-use change.