Country-wide mapping of harvest areas and post-harvest forest recovery using Landsat time series data in Japan

Contrasting Forest Loss and Gain Patterns in Subtropical China Detected Using an Integrated LandTrendr and Machine-Learning Method

“…The single factor could result in larger bias for forest loss and gain detection considering the image quality issues [35]. Through considering more variables except for the normally used individual indices such as IFZ, NBR, TCW, EVI, and NDVI, machine or deep learning methods (such as RF, ANN, and RCNN) have been increasingly applied in forest loss and gain detection during the recent years, and these methods have been proved to more comprehensively and accurately reflect both slow and abrupt forest loss and gain events [18,33,[35][36][37][38]. Our study also proved that this integration method can be used to detect provincial scale forest loss and gain in China, where has more complex terrains, vegetation types, and disturbance events.…”

“…In addition, several other studies have also proved the better performance of this integrated approach. For example, [36] proposed an ensemble approach for predicting disturbance and subsequent recovery levels for each disturbed pixel, and argued that this approach could simultaneously describe the levels of both disturbance and subsequent recovery as a single hybrid pattern; [37] developed the integrated approach between the LandTrendr and gradient nearest neighbor (GNN) imputation method to detect large tree and snag distributions; [38] detected forest harvest and post-harvest recovery in Japan using the integrated RF and LandTrendr approach, and concluded that this approach can provide valuable information on a country-wide replantation activity in harvest areas. However, due to the more complex terrain, forest types, and climate conditions in China, our previous study [16] indicated that some other factors such as elevation and image textures could also be important factors for detecting forest loss and gain in China.…”