Use AngleCam and how to contribute
Animation of the AngleCam input (image frames) and output (leaf angle estimates) for a Acer pseudoplatanus crown. Note that here the output (leaf angle distribution) is for simplicity integrated to average leaf angles. The animation shows that during a sunny day Tilia cordata tends to oscilate its leaf angles. The estimates show a relatively high variance, mostly caused by small changes in leaf orientation due to wind. Despite considerable variation in illumination conditions, the predictions show a relatively stable course during the day.
AngleCam is based on Convolutional Neural Networks (at current stage with TensorFlow and the EfficientNet backbone). We trained the networks with several thousands reference samples that were generated from visual interpretation of invidiual image frames. For each image frame, we sampled 20 leaves, which were then converted to a leaf angle distribution (beta distribution). The CNN models were, hence, trained to predict a leaf angle distribution for each individual image. The model accuracy was estimated from independent holdouts. Additionally, we performed a independent validation using terrestrial laser scanning and the TLSLeAF method by Atticus Stovall.
Model evaluation based on training data, test data and terrestrial laser scanning. A manuscript describing the method and its evaluation is currently in review.
Comparison of AngleCam and TLSLeAF for predicting leaf angle distributions.
-
R- and Python-Scripts for running AngleCam can be found in code_run_AngleCam.
-
The mandatory model object (hdf5) and example data can be downloaded from
-
Further details on the method and its evaluation are published (2022) in Methods in Ecology and Evolution
The code requires a running TensorFlow installation (see script for some help ressources). Please contact me if you find any bugs or have problems getting the models running: https://rsc4earth.de/authors/tkattenborn/ https://twitter.com/TejaKattenborn
Current evaluations indicate the transferability of the approach across scence conditions, species and plant forms. However, we cannot eventually state how well the models perform on your datasets (which may be composed of very different species, leaf forms or cameras). Additional labels (reference data) may be helpful to tune the model towards you application scenario. A R-script for producing new reference data is included in this repository. We would be very thankful if you would share these samples with us, so we can continuously improve the model performance and transferability. In return, we provide you a model object that was optimized for your data. AngleCam is truly in a alpha-phase and it success also depends on your help. Contributors will ofcourse be involved in upcoming analysis and scientific output.
-
Many thanks for funding to the German Centre for Integrative Biodiversity Research (iDiv) and Leipzig University.
-
Thanks to Marco Bobbinger for setting up the example.