A Land Clearing Index for High-Frequency Unsupervised Monitoring of Land Development Using Multi-Source Optical Remote Sensing Images
Zhixin Qi,Anthony Gar-On Yeh,Xia Li,Xiaoping Liu
DOI: https://doi.org/10.1016/j.isprsjprs.2022.03.017
IF: 12.7
2022-01-01
ISPRS Journal of Photogrammetry and Remote Sensing
Abstract:Timely information on land development is critical for preventing illegal construction, which is an increasingly serious problem in rapidly urbanizing countries. Although remote sensing has been widely used to observe land cover change, monitoring land development at a high frequency (e.g., weekly or daily) remains a challenge because of the lack of images with sufficient temporal resolution and methods independent of human supervision. This study developed a novel unsupervised method, namely, land clearing index (LCI)-based method, for monitoring land development at high frequencies. The LCI was proposed by examining the characteristics of the spectral change caused by the land clearing in the initial stage of land development. It is calculated with two images acquired at different times and can be used with an appropriate threshold to detect land development areas. Top-of-atmosphere reflectance (TOAR) data were found to be better than surface reflectance (SR) data for the use with the LCI-based method because the atmosphere effect played a positive role in distinguishing developed land from undeveloped land and suppressing the false alarms caused by the season change in vegetation. The major innovations of the LCI-based method are as follows: 1) applicable to the land development detection between images acquired by different optical sensors to allow for high-frequency detection; 2) can detect land development without training data and human supervision; 3) can detect various types of land development, such as converting farmland for building purposes, urban redevelopment, and land reclamation; and 4) simple, straightforward, and easy to implement. The LCI-based method was tested with different satellite images (Landsat-5, Landsat-7, Landsat-8, and Sentinel-2) for land development detection in nine cities selected worldwide. The average detection accuracy (DA), false-alarm rate (FR), and overall accuracy (OA) were 88.62%, 1.68%, and 98.22%, respectively. Compared with the traditional post-classification comparison based on the random forest algorithm, the LCI-based method produced better results and avoided the time-consuming work on sample selection. The LCI was also tested in each of the cities to detect land development at the various time intervals (5 days, 1 month, 3 months, 6 months, and 1 year) in the different seasons (January to March, April to June, July to September, and October to December). Approximately all the DAs were higher than 90.00%, while the FRs were highly related to the detection interval and ranged from 0.00% to 71.60%. The false alarms were mainly caused by the seasonal change in croplands, natural vegetation, shadows of mountains and buildings, and suspended sediment in the water. An approach was developed to calibrate the detection results of the LCI-based method for false alarm suppression. The false alarms were reduced by as much as 99.98% by reinforcing the LCIbased method with this calibration approach. The testing indicated that the LCI-based method is applicable and reliable for monitoring land development at a short time interval; thus, it is capable of preventing illegal construction.