10:00 AM - 10:15 AM
[HTT19-05] Trial of "visualization" of deforestation in Japan using deep learning
Keywords:Environmental change, Forest monitoring, AI
In this study, we report an attempt to detect deforestation sites throughout Japan between the spring and summer of 2021 using deep learning and satellite images, and to open the results to the public.
GLAD [1], and JJ-FAST [2,3] are two of the systems for capturing wide-area forest changes using satellite data and disclosing the results to the public in near real time. GLAD uses optical satellites (LANDSAT 7, 8, Sentinel-2 satellite) and C-band synthetic aperture radar satellite (Sentinel-1) to capture forest changes near the equator (+- 30 degrees) at intervals of 5 to 12 days. JJ-FAST uses the L-band synthetic aperture radar satellite (PALSAR-2) to provide information on forest changes that occurred in tropical forests in 78 countries 3 to 4 days after observation. But there is no system that can easily know if such a forest change is occurring in Japan. In the forestry industry in Japan, the number of unmanaged forests is increasing due to sluggish timber prices and the aging of forest owners, and illegal deforestation has been reported in some areas. Sustainable forest management is also required from the perspective of SDGs, but it has been considered that it would be costly to monitor wide-area forests and open logging detection points to the public.
In this study, forest change detection was performed using a 10 m resolution Sentinel-2 data obtained free of charge from the satellite data cloud service (Google Earth Engine). From the data observed from the summer of 2019 to the summer of 2021, images with few clouds were selected, covered with cloud masks, and summarized by season. Of these, using time-series data obtained in the five seasons from summer 2019 to summer 2020 in the Kanto region, a deep learning model was trained to capture the forest changes that occurred during the period from spring to summer of 2020. Recently, the use of deep learning is increasing even for satellite images, but in this research, we are improving the detection accuracy by using time-series images. Using this model, we detected all areas where forest changes of 0.25 ha or more occurred in 47 prefectures from the spring to summer of 2021. As a result, deforestation of 6472.9 ha was confirmed at the 6582 sites. The most frequently detected forest change sites were 0.25 to 0.5 ha in size, accounting for 46% of the total.
The detection accuracy largely depends on the cloud removal rate of the end-of-term images. In the Tohoku region where most of the clouds could be removed, it was 70.5% (Fukushima prefecture) to 93.5% (Yamagata prefecture). On the other hand, in the Chugoku region where many clouds remained in the end-of-term image, the detection accuracy was 2% (Okayama prefecture) to 48.3% (Hiroshima prefecture). As a result of applying deep learning to remove sites falsely detected by clouds to the detection result polygons in Okayama Prefecture, which had the worst detection accuracy, the accuracy improved to 76.1%.
The obtained forest change location information is open to the public for free using Google map, and the status of deforestation in each prefecture were reported by using YouTube (channel name: AI wo tsukatte uchukara shinrinhenka mitsukechaimashita). Since Google Map has a function to display your current location, you can easily access the detected forest change location and visit the site while looking at the forest change information published on your smartphone etc. This time, it was confirmed that after granting Google's editing authority to the collaborators, it is possible to go to the site and upload photos of forest changes. It was confirmed that such existing functions can be used to encourage interest in forest management and surveillance and active participation. In the presentation, we will also report on the computer cost, time cost, etc. required when "visualizing forest change points of 0.25 ha or more in Japan with free satellite data and deep learning".
[1] Hansen, M.C., Krylov, A., Tyukavina, A., Potapov, P.V., Turubanova, S., Zutta, B., Ifo, S., Margono, B., Stolle, F., Moore, R., 2016. Humid tropical forest disturbance alerts using Landsat data. Environ. Res. Lett. 11. Web: http://www.glad.umd.edu/index.php (accessed on 14 Feb. 2022).
[2] JJ-FAST, https://www.eorc.jaxa.jp/jjfast/jj_index.html, accessed on 14 Feb. 2022
[3] Manabu Watanabe, Christian Koyama, Masato Hayashi, Izumi Nagatani, Masanobu Shimada, “Refined Algorithm for Forest Early Warning System with ALOS-2/PALSAR-2 ScanSAR Data in Tropical Forests Regions”, Remote sensing of environment, 265, 112643, 2021.8, https://www.eorc.jaxa.jp/jjfast/topics/20210817.html
GLAD [1], and JJ-FAST [2,3] are two of the systems for capturing wide-area forest changes using satellite data and disclosing the results to the public in near real time. GLAD uses optical satellites (LANDSAT 7, 8, Sentinel-2 satellite) and C-band synthetic aperture radar satellite (Sentinel-1) to capture forest changes near the equator (+- 30 degrees) at intervals of 5 to 12 days. JJ-FAST uses the L-band synthetic aperture radar satellite (PALSAR-2) to provide information on forest changes that occurred in tropical forests in 78 countries 3 to 4 days after observation. But there is no system that can easily know if such a forest change is occurring in Japan. In the forestry industry in Japan, the number of unmanaged forests is increasing due to sluggish timber prices and the aging of forest owners, and illegal deforestation has been reported in some areas. Sustainable forest management is also required from the perspective of SDGs, but it has been considered that it would be costly to monitor wide-area forests and open logging detection points to the public.
In this study, forest change detection was performed using a 10 m resolution Sentinel-2 data obtained free of charge from the satellite data cloud service (Google Earth Engine). From the data observed from the summer of 2019 to the summer of 2021, images with few clouds were selected, covered with cloud masks, and summarized by season. Of these, using time-series data obtained in the five seasons from summer 2019 to summer 2020 in the Kanto region, a deep learning model was trained to capture the forest changes that occurred during the period from spring to summer of 2020. Recently, the use of deep learning is increasing even for satellite images, but in this research, we are improving the detection accuracy by using time-series images. Using this model, we detected all areas where forest changes of 0.25 ha or more occurred in 47 prefectures from the spring to summer of 2021. As a result, deforestation of 6472.9 ha was confirmed at the 6582 sites. The most frequently detected forest change sites were 0.25 to 0.5 ha in size, accounting for 46% of the total.
The detection accuracy largely depends on the cloud removal rate of the end-of-term images. In the Tohoku region where most of the clouds could be removed, it was 70.5% (Fukushima prefecture) to 93.5% (Yamagata prefecture). On the other hand, in the Chugoku region where many clouds remained in the end-of-term image, the detection accuracy was 2% (Okayama prefecture) to 48.3% (Hiroshima prefecture). As a result of applying deep learning to remove sites falsely detected by clouds to the detection result polygons in Okayama Prefecture, which had the worst detection accuracy, the accuracy improved to 76.1%.
The obtained forest change location information is open to the public for free using Google map, and the status of deforestation in each prefecture were reported by using YouTube (channel name: AI wo tsukatte uchukara shinrinhenka mitsukechaimashita). Since Google Map has a function to display your current location, you can easily access the detected forest change location and visit the site while looking at the forest change information published on your smartphone etc. This time, it was confirmed that after granting Google's editing authority to the collaborators, it is possible to go to the site and upload photos of forest changes. It was confirmed that such existing functions can be used to encourage interest in forest management and surveillance and active participation. In the presentation, we will also report on the computer cost, time cost, etc. required when "visualizing forest change points of 0.25 ha or more in Japan with free satellite data and deep learning".
[1] Hansen, M.C., Krylov, A., Tyukavina, A., Potapov, P.V., Turubanova, S., Zutta, B., Ifo, S., Margono, B., Stolle, F., Moore, R., 2016. Humid tropical forest disturbance alerts using Landsat data. Environ. Res. Lett. 11. Web: http://www.glad.umd.edu/index.php (accessed on 14 Feb. 2022).
[2] JJ-FAST, https://www.eorc.jaxa.jp/jjfast/jj_index.html, accessed on 14 Feb. 2022
[3] Manabu Watanabe, Christian Koyama, Masato Hayashi, Izumi Nagatani, Masanobu Shimada, “Refined Algorithm for Forest Early Warning System with ALOS-2/PALSAR-2 ScanSAR Data in Tropical Forests Regions”, Remote sensing of environment, 265, 112643, 2021.8, https://www.eorc.jaxa.jp/jjfast/topics/20210817.html