Machine vision analysis of harvested forestry sites using high resolution UAV data

The development of algorithms for building photo mosaics and 3D surface models over post-harvest forestry sites using high-resolution imagery captured by a low-flying UAV, is presented. The proposed system contains three databases with different resolution and height of shooting from forests and agricultural lands including: 1) Sentinel 2A/2B (resolution of 10, 20, and 60 m with 13 bands), 2) Airbus multispectral imagery dataset (resolution of 1.5 m with 4 bands), and 3) Multispectral imagery by SENSEFLY’S UAV (resolution of 11 cm with 4 bands). The proposed system conducts NDVI remote sensing on Airbus and SENSEFLY’S UAV multispectral imagery dataset as only one set of images are available. However, the proposed method uses sentinel 2A and 2B dataset form Trairao, Brazil. The result of the accuracy indicates that the accuracy of system is above 97.81%. However, the average accuracy of the proposed methods was found to be 98.13%. The precision of the classifier was found to be 97.75%. The response time of the system is tested, and it was found that the mode response time for this system is below 15 seconds. The mean value for the response time was found to be 7.56 seconds. The results indicate that the accuracy of the proposed algorithm can be maintained at 95% under 18% cloud coverage. The accuracy of system under 3% cloud coverage was found to be slightly lower than the average accuracy of proposed method under 0% cloud coverage. The accuracy drops by 0.40% from 0 to 3 percent cloud coverage. The average efficiency of system for cloud masking of 3%, 6%, 9%, 12%, 15% and 18% was found as 97.80%, 96.95%, 96.55%, 96.35%, and 95.85%, respectively.