Research On Forest Land Change Monitoring Based On Spatial Downscaling

With the development of human activities and the world economy,forest resources have significantly decreased,and forest coverage has changed rapidly.Thus,it is crucial to obtain information on forest cover changes quickly and accurately.The Sentinel-2 remote sensing satellite includes two satellites,2A and 2B.Due to its high spatial and spectral resolution,it has great research value in monitoring forest cover changes.However,the 10-60-meter spatial resolution of Sentinel-2 images is affected by mixed pixels in fine-scale forest change detection,resulting in monitoring accuracy that does not meet actual requirements.The spatial downscaling mixing pixel decomposition technology can quantitatively solve the problem of mixed pixels by solving the spatial ratio occupied by each type of land cover from the spectrum.Through spatial downscaling and location technology,mixed pixels can be precisely decomposed to sub-pixel level,thereby more accurately estimating their internal spatial arrangement.Based on this,forest change can be monitored at a smaller scale to better meet actual needs.This study used Sentinel-2 and domestic Jilin-1 satellite images to conduct forest change monitoring by employing the spatial downscaling method.The main research contents are as follows:(1)Comparing the results of five pixel-based classification monitoring methods,namely decision tree,random forest,artificial neural network,maximum likelihood,and support vector machine,it is experimentally proven that the random forest algorithm performs the best in terms of overall accuracy,forest mapping accuracy,forest user accuracy,Kappa coefficient,etc.The four accuracy indicators reached 87.59%,87.34%,88.03%,and 0.841,respectively.However,compared with JL-1’s classification monitoring results with higher resolution,there is still a large gap in classification monitoring accuracy.The existence of "same spectrum different substance" and "same substance different spectrum" mixed pixels seriously affects the accuracy of pixel-level classification monitoring,resulting in differences from submeter classification monitoring.(2)The shortcomings and deficiencies of the traditional PSO endmember extraction algorithm are analyzed,and the QPSO endmember optimization extraction method is introduced,which is more suitable for high-resolution remote sensing images.Combining the endmember extraction of the QPSO algorithm and the abundance inversion method based on unconstrained least squares,the Sentinel-2 image’s land object endmembers on January 12,2021,in Qingshanpu Town were decomposed into mixed pixels,including forest land,buildings,water bodies,unused land,grassland,and farmland,and precision verification was conducted.The experimental results show that compared with the traditional endmember extraction algorithms VC A,N-Findr,and PSO algorithms,the endmember spectra extracted by the QPSO optimization algorithm are clear in meaning.In the two items of root-mean-square error RMSE and mean absolute error MAE,they are smaller than PSO and VCA.In terms of spectral angle parameter SAD,the QPSO algorithm performs the best,with an average SAD value of 0.726,which is higher in accuracy and stability than the other three algorithms.The average SAD of the PSO algorithm is second,at 0.794,slightly higher than that of the QPSO algorithm.The average SAD of the VCA and N-FINDR algorithms is relatively high,at 0.861 and 0.906,respectively.The QPSO algorithm has increased by 8.5%,15.6%,and 19.8%over the traditional three algorithms in terms of average SAD,and the accuracy is higher than that of the traditional endmember extraction algorithm.(3)Based on the SPSAM spatial downscaling location algorithm,the pixels are segmented into 16 finer sub-pixels according to the gravity relationship between pixels and sub-pixels according to the downscaling factor,and spatial information on the land cover of each sub-pixel is assigned according to the abundance inversion results.This successfully breaks through the 10m spatial resolution of Sentinel-2 images,downsizing the forest monitoring results to 2.5m and achieving sub-pixel-level extraction of the study area’s forest.(4)Combined with the Jilin-1 satellite image with sub-meter resolution and the downscaling correction algorithm based on edge matching,the edge contour details of the downscaled forest monitoring results are corrected,making the forest extent closer to actual results.The results show that the evaluation indicators of the corrected downscaled location results approach the sub-meter monitoring results,especially the forest user accuracy,which reaches 94.28%.Based on the corrected downscaled location algorithm,five experiments were carried out to extract forest changes in Qingshanpu Town from 2019 to 2022.The number of sub-pixels monitored by the downscaled location results was counted.It showed that the forest coverage rate had steadily increased in recent years,with the portion of forest land increasing from 37.81%in 2019 to 40.15%in 2022,and the forest area increasing from 1751.74 hectares in 2019 to 1856.24 hectares in 2022.However,from 2022 to 2023 it drastically dropped to 1643.67 hectares,with a negative forest change rate of 11.51%,resulting in a direct loss of forest land of 213.57 hectares.