Integrate other large datasets.

[bw4sz]

This issue is meant for cataloging data that should eventually go into a deepforest baseline. Happy to have contributions from the community on this issue.

https://arxiv.org/pdf/2208.10607.pdf
https://github.com/jonathanventura/urban-tree-detection-data

A portion of this could be used.
https://google.github.io/auto-arborist/

More forest data

https://lila.science/datasets/forest-damages-larch-casebearer/

Along with this issue we need a strategy for updating the baseline model and potential tradeoffs to new model weights.

nightonion/yosemite-tree-dataset#2

Roadmap

1. Format train and test data to meet deepforest inputs.
2. Assess test, or portion of train, on existing backbone.
3. Document a training strategy on new data
4. Assess test on new model
5. Assess combined model on NeonTreeEvaluation benchmark to assess tradeoffs.

[bw4sz]

Keep an eye on https://www.fruitpunch.ai/challenges/ai-for-trees.

We have downloaded this dataset from Kenya, currently cannot get annotations to overlap.

In orange:

[bw4sz]

There my be data associated with
Methods Ecol Evol - 2021 - Tang - Large‐scale image‐based tree species mapping in a tropical forest using artificial.pdf

Ground labels of tree species are available for a 16-ha area, the
Luquillo Forest Dynamics Plot (LFDP), covering <1% of the area captured by the aerial images (Figure 1). All stems ≥1 cm in diameter in
the plot were censused in 2016, including information on species
taxonomy and stem diameters, and all trees are fully georeferenced
(see Thompson et al., 2002 for details). In this paper, we used ground
label information from all live stems ≥20 cm diameter in the 2016
census. This size class is expected to include most individuals with
visible canopies in the aerial photographs and provides more reliable
supervision information for the proposed method. We also ran experiments using 15 cm and 25 cm as the threshold, and results are
similar (Appendix C).

[bw4sz]

https://essd.copernicus.org/preprints/essd-2022-312/ germany tree species

[bw4sz]

Synthetic trees from ground view, useful in understanding generalization.

https://arxiv.org/pdf/2210.17424.pdf

[bw4sz]

https://besjournals.onlinelibrary.wiley.com/doi/10.1111/2041-210X.13860
drones. Part of https://openforestobservatory.org/dataset-details/ and the larger conversation with Derek Young at UC Davis. I wrote Derek directly about a larger collaboration.

[bw4sz]

https://essd.copernicus.org/articles/14/4967/2022/ siberia

[bw4sz]

Is there any overlapping RGB data here? Will need to contact each individually. https://open-research-europe.ec.europa.eu/articles/3-32

[bw4sz]

@henrykironde can we download these datasets and let's start listing here. I'll do some and you can do some as we prepare for GSOC students. For each dataset:

1. How many training trees?
2. How many training images?
3. What format are the annotations?
4. Drop a sample image with annotations overlaid (shapefiles?).

Overall, we can start moving them to

/blue/ewhite/DeepForest

In general my hope is to create a train and test split for each dataset where possible.

[bw4sz]

Summary of https://lila.science/datasets/forest-damages-larch-casebearer/.
which is now on /blue/ewhite/DeepForest

(base) [b.weinstein@login3 Radogoshi_Sweden]$ ls
Bebehojd_20190527  Bebehojd_20190819  Data_Set_Larch_Casebearer  Ekbacka_20190527  Ekbacka_20190819  images  Jallasvag_20190527  Jallasvag_20190819  Kampe_20190527  Kampe_20190819  Nordkap_20190527  Nordkap_20190819  test.csv  train.csv

There are 60,000 training trees.

(base) [b.weinstein@login3 Radogoshi_Sweden]$ cat train.csv | wc -l
60919

There are 41,148 test trees. That seems like too many test compared to train. We should reset that.

DeepTreeAttention) [b.weinstein@login3 Images]$ rio info B06_0051.JPG
/home/b.weinstein/miniconda3/envs/DeepTreeAttention/lib/python3.8/site-packages/rasterio/__init__.py:277: NotGeoreferencedWarning: Dataset has no geotransform, gcps, or rpcs. The identity matrix will be returned.
  dataset = DatasetReader(path, driver=driver, sharing=sharing, **kwargs)
{"bounds": [0.0, 1500.0, 1500.0, 0.0], "colorinterp": ["red", "green", "blue"], "compress": "jpeg", "count": 3, "crs": null, "descriptions": [null, null, null], "driver": "JPEG", "dtype": "uint8", "height": 1500, "indexes": [1, 2, 3], "interleave": "pixel", "mask_flags": [["all_valid"], ["all_valid"], ["all_valid"]], "nodata": null, "photometric": "ycbcr", "res": [1.0, 1.0], "shape": [1500, 1500], "tiled": false, "transform": [1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0], "units": [null, null, null], "width": 1500}

The resolution of the image is unknown, but looks like be about 10cm.

[bw4sz]

Here is a report for https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit

I have sent the maintainers an email. Things didn't quite line up.

Hello,
I maintain a small python tool for object detection in forested landscapes. 
https://deepforest.readthedocs.io/en/latest/
https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13472

The tree detection tool has become quite popular in the scientific community and we are looking to bolster it with images around the world. I've been collecting links to tree annotations and imagery and have a student this summer who will start training new models. I went to inspect the dataset linked from https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit, but I became a bit confused. 
I downloaded the .shp and overlaid it on the imagery. I had initial trouble because in the document here: https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit, the geotiff that is linked does not spatially overlap with the .shp. 

Looking around https://map.openaerialmap.org/#/-175.26188850402832,-21.1463924204947,13/square/20002233030/5a82999d5a9ef7cb5d5ae685?_k=ofo82l I found a geotiff that does overlap the area. However, zooming in, it looks as if there isn't strong georeferencing, the tree points don't correspond to trees. They seem more or less randomly distributed. I'm assuming that this is not the tile that was intended to be georeferenced to this shapefile.

All help appreciated.
Ben Weinstein

There are 13,000 labeled trees. We may need to clean up non-coconut trees with new annotations.

[bw4sz]

I emailed the authors from https://www.researchgate.net/profile/Mertcan-Nazar/publication/369093360_Deep_Learning_Based_Poplar_Tree_Detection_and_Counting_Using_Multispectral_UAV_Images/links/640994725749505945793fce/Deep-Learning-Based-Poplar-Tree-Detection-and-Counting-Using-Multispectral-UAV-Images.pdf

looks like drone images from turkey.

[bw4sz]

https://zenodo.org/record/7528566

https://www.mdpi.com/2072-4292/15/5/1463

This is a private dataset and cannot be shared or re-used in additional capacity besides improving the model baseline. I can see the images but haven't checked the .xml annotation overlap. The image quality looks dark but acceptable.

[bw4sz]

Report on

We present a training dataset of tropical Northern Australia savanna woodland tree species that was generated using RPAS and on-ground surveys to confirm species labels. RPAS-derived imagery was annotated, resulting in 2547 polygons representing 36 tree species.

This dataset is now on /blue/ewhite/DeepForest/Australia_Savannah

Figure is from paper, trained model, but gives a sense for the resolution and annotation format. Data is still downloading from zenodo.

[bw4sz]

I wrote the corresponding author:

https://www.sciencedirect.com/science/article/pii/S0303243422000903

A small number of polygon annotations from central park NY.

[bw4sz]

Here is a report on https://github.com/jonathanventura/urban-tree-detection-data

from article https://arxiv.org/pdf/2208.10607.pdf

Nice summary on the README. Excellent paper.

https://github.com/jonathanventura/urban-tree-detection-data#readme

Annotations are made in point format! This may require more trouble than its worth, but could be generated in a semi-supervised manner using existing deepforest baseline. As a way of improving resolution and urban scenes. NAIP imagery is in high demand from deepforest users. Atleast to be used in an evaluation benchmark against NAIP.

[bw4sz]

Exploring https://doi.pangaea.de/10.1594/PANGAEA.933263?format=html#download. Siberia trees from.

This Individual-labelled trees dataset is a part of the SiDroForest data collection (https://www.pangaea.de/?q=keyword%3A%22SiDroForest%22) and contains spatial data in the form of points and polygons of 872 trees and shrubs that were recorded in Siberia during a 2-month fieldwork expedition in 2018 by the Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research in Germany (Kruse et al., 2019). During the fieldwork, 15 m radius vegetation surveys were performed during which information such as height, species, crown diameter were recorded for individual trees. I

The files are too large to download, so I don't know which orthomosaics go with the crown polygons. I downloaded "Kruse_et_al_SiDroForest_RGB_Orthomosiac.zip", but there are two others.

From this image it make it seem like there are small image chips, not one giant tile, but I don't see that in the manifest.

There are also synthetic trees that I have not investigated: https://doi.pangaea.de/10.1594/PANGAEA.932795

[bw4sz]

Deepforest was used to preprocess training data in Ecuador. Important to use only in train, since it is weakly annotated and already touched our system. https://arxiv.org/pdf/2201.11192.pdf

https://github.com/gyrrei/ReforesTree

[bw4sz]

We have very little data from south asia. May be interested in annotating some tiles from https://www.mdpi.com/1999-4907/14/3/586

[bw4sz]

9,000 trees from TEAK and SOAP neon sites. with alive dead labels

https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2022JG007234
TEAK_trees.zip

available on github

data will need to be cropped

[bw4sz]

We’ve received approval by our field partner (LEAD Foundation) too that they are willing to share the datasets that we collected together.

We have drone imagery of 17 villages in Tanzania where we implement an agroforestry program. For 13 of these villages, we have data over 3 consecutive years (2018 , 2019, 2020) in this cloud bucket:

https://console.cloud.google.com/storage/browser/justdiggit-drone

Some of the data has been labeled. The file ‘justdiggit-drone/label_sample/Annotations_trees_only.json’ contains the most complete annotations. There is also another file 'label_sample_COCO_RDP - Tree annotation.json' which also contains a class 'messy vegetation' in

addition to trees. However, there are very few annotations for this class compared to the trees.

We have some more recent data too, but that hasn’t been added to the cloud bucket yet. Let me know if you would need that too.

For some of the villages we also have 50 cm resolution satellite images (Planet SkySat) captured around the same time as we have the drone flights (as we want to use the detected trees on our drone imagery to detect trees on this satellite imagery too). They are in this bucket:

https://console.cloud.google.com/storage/browser/justdiggit-skysat

Has been downloaded to

/blue/ewhite/DeepForest/justdiggit/

very large, over 300GB.

[bw4sz]

South Asia
https://zenodo.org/record/7843154#.ZGUQD-zMKDU

https://zslpublications.onlinelibrary.wiley.com/doi/10.1002/rse2.332

https://catalogue.ceda.ac.uk/uuid/dd4d20c8626f4b9d99bc14358b1b50fe

[bw4sz]

https://ai-climate.berkeley.edu/scale-mae-website/ scale training

and scale based pretraining

https://github.com/sustainlab-group/SatMAE/issues

[bw4sz]

I had a meeting with Josh Veitch Michaelis from ETH about https://restor.eco/?lat=26&lng=14.23&zoom=3. OpenStreetMap data, alot of labels, not sure the status.

[bw4sz]

https://www.biorxiv.org/content/10.1101/2023.08.03.548604v1.full.pdf polygon annotations, 14 species

The dataset created here (23,000 segmented individual tree crowns) includes 13 species and genera,
which have an environmental and an economic importance in Northeastern North America.
Moreover, these trees are segmented on seven different dates, totalling almost 161,000 annotated
tree crowns. This dataset is available online for use by others. 

I emailed the authors. Nice paper.

Dataset is out with RGB and LiDAR imagery. https://zenodo.org/records/8148479

[bw4sz]

Check back in on this preprint Data Availability Statement: Crown data used for model training and all M. polymorpha data products are openly available in Figshare at xxxxx.Acknowledgments: The Global Airborne Observatory (GAO) is managed by the Center for Global Discovery and Conservation Science at Arizona State University. The GAO is made possible by support from private foundations, visionary individuals, and Arizona State University.

…

On Wed, Aug 9, 2023 at 2:40 PM Benjamin Weinstein ***@***.***> wrote: https://www.biorxiv.org/content/10.1101/2023.08.03.548604v1.full.pdf polygon annotations, 14 species The dataset created here (23,000 segmented individual tree crowns) includes 13 species and genera, which have an environmental and an economic importance in Northeastern North America. Moreover, these trees are segmented on seven different dates, totalling almost 161,000 annotated tree crowns. This dataset is available online for use by others. I emailed the authors. Nice paper. — Reply to this email directly, view it on GitHub <#340 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJHBLDKXRUOYAAGTBEH5ALXUP7UTANCNFSM57ZN6MTA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Ben Weinstein, Ph.D. Research Scientist University of Florida http://benweinstein.weebly.com/

[bw4sz]

https://zenodo.org/record/8008028

https://www.mdpi.com/2072-4292/15/14/3599

Netherlands

Data is corrupt as far as I can tell? Data was downloaded, provided by authors.

[bw4sz]

data was not made available, but probably could ask authors. Netherlands. https://www.mdpi.com/2072-4292/15/17/4128#

[bw4sz]

LiDAR benchmark, as Puliti if there is overlapping RGB data?

[bw4sz]

There is a LiDAR dataset, but one figure has a ortho basemap. I emailed the authors.

https://essd.copernicus.org/articles/14/2989/2022/essd-14-2989-2022.pdf

[bw4sz]

An additional siberia dataset? https://doi.pangaea.de/10.1594/PANGAEA.957253

[bw4sz]

Need to write https://www.nature.com/articles/s41598-020-79653-9

[bw4sz]

Need to write

[bw4sz]

France LiDAR paper, is there orthos? https://www.mdpi.com/2072-4292/14/5/1083

[bw4sz]

unlabeled https://datadryad.org/stash/dataset/doi:10.5061/dryad.21t1805 tropical forest liana

[tyfoan]

Here is a report for https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit

I have sent the maintainers an email. Things didn't quite line up.

Hello, I maintain a small python tool for object detection in forested landscapes.  https://deepforest.readthedocs.io/en/latest/ https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13472

The tree detection tool has become quite popular in the scientific community and we are looking to bolster it with images around the world. I've been collecting links to tree annotations and imagery and have a student this summer who will start training new models. I went to inspect the dataset linked from https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit, but I became a bit confused.  I downloaded the .shp and overlaid it on the imagery. I had initial trouble because in the document here: https://docs.google.com/document/d/16kKik2clGutKejU8uqZevNY6JALf4aVk2ELxLeR-msQ/edit, the geotiff that is linked does not spatially overlap with the .shp. 

Looking around https://map.openaerialmap.org/#/-175.26188850402832,-21.1463924204947,13/square/20002233030/5a82999d5a9ef7cb5d5ae685?_k=ofo82l I found a geotiff that does overlap the area. However, zooming in, it looks as if there isn't strong georeferencing, the tree points don't correspond to trees. They seem more or less randomly distributed. I'm assuming that this is not the tile that was intended to be georeferenced to this shapefile.

All help appreciated. Ben Weinstein

There are 13,000 labeled trees. We may need to clean up non-coconut trees with new annotations.

If you still need help then you should change the CRS to ESPG:3857 for .tif layer

[rlrognstad]

Ran across datasets for this paper and repo.
An interesting use case to complement RGB with thermal imagery for detecting trees in shadow.
Potential immediate use of RGB labeled images and future use if high-res thermal becomes more widely available (HotSat or similar)

[bw4sz]

Thanks @rlrognstad. We got this one, and are in contact with the authors. I appreciate you looking with us! I'm compiling a list below. This issue is pretty old and pre-dates any formal attempt in pulling data together. They actually used DeepForest for a portion of their training structure.

I suppose I can close this issue and point to the sheet.

https://docs.google.com/spreadsheets/d/1-Q6ekQNE7TZBHQnrbjGl_2tcfh_X9154Kbzn3YFFruM/edit?usp=sharing

Definitely let me know about absolutely anything you see!