This is a scrollable orthomosaic composite of aerial photos taken at the beginning of 1958, covering much of the village of Yangambi and the ongoing agricultural research at the time. In addition a forest cover change map is presented using both the historical data and the Hansen et al. (2013) landsat based map as a reference. Use the button on the top right in the map to disable the forest cover map and bring up the underlying orthomosaic.

The final orthomosaic composition as shown below covers approximately 1.5 million pixels or 1325 km², with the extend of the scene being ~38 by 40 km. The overall accuracy of the Structure from motion orthomosaic composition was 0.88 m / pixel using the sparse cloud DEM for corrections at 45.8 m/pixel. The resulting georeferenced scene had a spatial accuracy of approximately 23m. Further georeferencing outside the Structure from motion (Sfm) workflow reduced the mean error to ~10m.

Methods

Data acquisition

Data for the central Congo Basin region, surrounding Kisangani, were collected in several flights during the dry season of 1958 and 1959 (from 8/01/1958 to 20/02/1958 and from 28/12/1958 to 9/01/1959 respectivelly) for the generation of (topographic) maps of the area. collection was supervised by the “Institut Géographique du Congo Belge” in Kinshasa (then Léopoldville).

Near-infrared images were gathered along flight paths running mostly from west to east, between 9 - 11h local time. Along a flight path continuous images were taken using a Wild Heerburgg RC5a (currently Leica Geosystems) with an Aviogon lens assembly (114.83mm / f 5.6, with a 90 degree view angle) resulting in square negative prints of 180 by 180 mm. Flights were flown at an average absolute altitude of ~5200 m above sea level, covering roughly 18 530 km² at an approximate scale of 1/40 000 (mm). The use of the integrated autograph system ensured timely acquisition of pictures with a precise overlap (~1/3) between images. This large overlap between images together with flight parameters would allow post-processing, using stereographs, to create accurate topographic maps.

Site selection

We prioritized flight paths which cover current day permanent sampling plots, larger protected areas, and past agricultural research facilities. This selection ensure the proper mapping of the Yangambi area and the life history of the forest surrounding it. Coverage of these areas was be prioritized in data recovery, selecting flight paths 1 through 11 for digitization. From this larger dataset we selected 71 survey images covering the larger Yangambi area for processing into an orthomosaic and further analysis. All the selected images stem from the flight campaign during January and February of 1958 and cover approximately 1325 km² (38 x 40 km). The area includes the Yangambi village, 20 contemporary permanent sampling plots, past and present agricultural experimental plots and large sections of Yangambi Man and Biosphere reserve surrounding to the west and east of the village.

Digitization and data processing

All selected images were contact prints as original negatives of the prints were not available. Images were scanned at a resolution exceeding their original resolution (or grain) at the maximal physical resolution of an Epson A3 flatbed scanner (i.e. 2400 dpi or 160MP per image) and saved as lossless tiff images. The raw image data was subsequently downsampled to half the original resolution (1200 dpi or ~45.5 px/mm, 81MP) and saved as lossless PNG format for processing purposes. This created a dataset with digital images at a resolution slightly above the grain of the photographs, while the reduced image size aided easier file handling and processing.

Data was processed into a georeferenced orthomosaic using a Sfm approach implemented in Agisoft Metashape (version 1.5.2, http://www.agisoft.com). An orthomosaic corrects remote sensing data to represent a perfectly downward looking image, free from perspective distortions due to topographic relief and camera tilt. Using the Sfm technique, relying on stereographic acquisitions and photogrammetry, we can reconstruct a 3D surface and relative camera positions in order to correct and mosaic the arial photographs together. During the Sfm analysis we masked out the edges of the images which include imaging meta-data, clouds and glare or large water bodies such as the Congo river.

We calculated the orthomosaic using a low resolution point cloud and digital elevation map (DEM). Additional ground control points were provided to assist in the referencing of image in space and constrain the optimization routine used in the Sfm algorithm. Ground control points consisted of rooftop edges of permanent structures which could be verified in both old and new aerial imagery (i.e. Google Maps / ESRI World Imagery). Although clouds were removed during the Sfm routine we did not mask all clouds in the final orthomosaic to maximize forest coverage. The orthomosaic was exported as a geotiff for further georeferencing.

Detailed geo-referencing was executed in QGIS (2018) using the georeferencer plugin (version 3.1.9) and ESRI World Imagery high resolution reference data. We used 3^th degree polynomial and 16 ground control points to correct the final image. Ground control points, raw image data and final processed image is provided in addition to measures of uncertainty such as mean root mean squared (RMSE), mean and median error across all ground control points. All subsequent analysis are executed on the final geo-referenced orthomosaic or subsets of it.

Long term changes in canopy cover & automatic forest classification

To map long term land-use and land-cover change we used the Global Forest Change version 1.6 data covering the historical survey data (Hansen et al. 2013, tile 10N-020E) to calculate the latest state of the forest with respect to the conditions at the start of 1958, 60 years earlier. In this analysis we included only forested pixels which recorded no loss throughout the whole 2000 - 2018 period.

We automatically delineated all natural forest in the historical data, thus excluding tree plantations, thinned or deteriorated forest stands, fields and buildings. We used a Keras based Unet Convoluted Neural Net (CNN) architecture implementation with an efficientnetb3 backbone running on TensorFlow to train a binary classifier (i.e. forest or non-forested). Training data were collected from the orthomosaic by randomly selecting 513 pixel square tiles from homogeneous forested or non-forested areas within the historical orthomosaic. Homogeneous tiles were combined in synthetic landscapes using a random gaussian field based binary mask. We generated 5000 synthetic landscapes for training, while 500 landscapes were generated for both the validation and the testing dataset. Source tiles did not repeat across datasets to limit overfitting. In order to limit stitch line misclassifications, along the seams of mosaicked images, we created synthetic landscapes with different forest tiles to mimick forest texture transitions. We applied this technique to 10% of the generated synthetic landscapes. The CNN model was trained for 100 epochs on a graphics processing unit (GPU) maximizing the Intersect-over-Union (IoU) using additional data augmentation. Data augmentation included random cropping to 320 pixel squares, random orientation, scaling, perspective, contrast and brightness shifts and image blurring. During final model evaluation we report the IoU of our out-of-sample test datasets. The optimized model was used to classify the complete orthomosaic using a moving window approach with a step size of 110 pixels and a majority vote across overlapping areas to limit segmentation edge effects.

References

Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, et al. 2013. “High-Resolution Global Maps of 21st-Century Forest Cover Change.” Science 342 (6160)