Ground Cover Change (GCC) model

The Ground Cover Change (GGC) model's dataset is a collection of primary input and output data sources to classify regional land cover change maps. These datasets were essential to formulate a simple and automatic label supervision model to secure reliable land cover samples from multisource remote sensing data and predict accurate land cover maps, aiming to evaluate land change interactions prompted by post-disaster resettlement sites. The affected areas after the 2010 Merapi volcano eruptions in Java, Indonesia, were targeted for this research. Steps to perform data processing pipelines consist of: (1) Data collected from open-sourced datasets including the Ice, Cloud, and land Elevation Satellite 2 (ICESat-2) canopy height datasets, a land cover product, Forest Watch Canopy Heigths, and ancillary DEM-derived subsets of data. (2) These datasets were converged to perform a three-dimensional supervision of land cover samples derived from the pre-existent land cover product. (3) Highly-reliable samples were processed to redeem validation and metric sample units for multitemporal and static amplitude metric sets. (4) A filtering code was parsed with a regionally-consistent metric set to classify candidate training samples with levels of reliability based on a three-point scoring system. (5) A Random Forest model was learnt with refined training data to classify 2016-2021 Sentinel-2 satellite images over Java's tropical rainforest. (6) Land change interactions in localized areas of the Cangkringan district, Yogyakarta, Java, Indonesia, were analyzed to quantify land change interactions caused by urban resettlement sites. We found that agricultural development near resettlement sites was trending upwards despite the overall declining tendency. However, we also saw 25% forest loss, 39% built-up growth, and 43% cropland growth as a product of resettlement sites. The GCC model applies structural height vetting filters and converges statistical feature metrics to automatically detect spurious labels and classifies the data by levels of reliability. The training dataset can be nurtured annually with additional high-quality samples providing all the necessary tools (training and reference datasets) to train learning models in real-time. Given the time constraints posed by emergencies, this model becomes a valuable asset to the recovery phases of natural disasters, where our understanding of how to rebuild affected areas is limited to the traditional urban planning schemes of development. By analyzing post-disaster resettlement in the rural periphery of urban areas, resilient communities can be built with the objective of strengthening rural environments and diminishing the overall human footprint on Earth.