Dynamic Forest Change Detection Research Based On Deep Learning

Forests are one of the largest terrestrial ecosystems on Earth and play a vital role in the global biosphere.Monitoring changes in forests is a fundamental task in ecological forestry and an essential means of ensuring the authenticity,accuracy,and timeliness of forest resource data.The use of remote sensing technology to detect and map changes in forests is of significant importance for the protection,monitoring,and management of forest resources.The purpose of this paper is to use high-resolution imagery from GF-2 and multitemporal Sentinel-2 Level 1C remote sensing imagery to construct a forest extraction dataset and a multi-temporal forest change detection dataset.The paper proposes a forest extraction algorithm based on Fully Convolutional Network(FCN)combined with lowresolution labels,and a deep learning based forest change dynamic detection algorithm.These two algorithms are applied to the annual and quarterly forest change dynamic detection work in Xiangtan City to verify the feasibility of the proposed methods.The main conclusions of this study are as follows:(1)This paper uses the forest extraction algorithm to extract the forests in the 2020 and 2021 GF-2 images of Xiangtan City.The method achieved higher F1-scores(97.09%and 95.96%for 2020 and 2021,respectively)than other segmentation models,such as U-Net,FPN,and LinkNet.Furthermore,by comparing the forest extraction results of two time periods,the Precision,Recall,and F1-score of forest change detection were 69.98%,84.61%,and 76.60%,respectively.The method uses a small number of lowresolution labels to quickly and accurately extract the forest change area from highresolution remote sensing images,and provide data support for subsequent studies.(2)This paper proposes a forest change detection method that utilizes the U-Net++deep learning model and Cross Entropy Loss(CELoss)function.The proposed method has demonstrated significant improvements in detection performance on the Sentinel-2 remote sensing image dataset,achieving the Precision of 79.54%,Recall of 74.78%,and F1-score of 77.09%for forest change detection.Compared to other models,the U-Net++model utilized in this study has shown the most outstanding overall performance.(3)This paper combines U-Net++ change detection algorithm and Sentinel-2 remote sensing images to predict forest changes within the forest land area of Xiangtan City from 2017 to 2021.In the annual change detection results,the F1 score of each year is higher than 75%.In addition,the quarterly change detection results are used to map the dynamic forest change detection results from 2017 to 2021 in Xiangtan City,which provides a strong support for forest resource monitoring in the region.The dynamic detection of forest change based on multi-temporal remote sensing image data can precisely determine the time period of occurrence of each change patch with a short interval,which can better analyze the causes of change.This method is important for grasping the dynamic update status of forest resources and implementing management.