Using Deep Learning to Identify Costa Rican Native Tree Species From Wood Cut Images.

Tree species identification is critical to support their conservation, sustainable management and, particularly, the fight against illegal logging. Therefore, it is very important to develop fast and accurate identification systems even for non-experts. In this research we have achieved three main results. First, we developed-from scratch and using new sample collecting and processing protocols-an dataset called CRTreeCuts that comprises macroscopic cross-section images of 147 native tree species from Costa Rica. Secondly, we implemented a CNN for automated tree species identification based on macroscopic images of cross-sections of wood. For this CNN we apply the fine-tuning technique with VGG16 as a base model, pre-trained with the ImageNet data set. This model is trained and tested with a subset of 75 species from CRTreeCuts. The top-1 and top-3 accuracies achieved in the testing phase are 70.5% and 80.3%, respectively. The Same-Specimen-Picture Bias (SSPB), which is known to erroneously increase accuracy, is absent in all experiments. Finally, the third result is Cocobolo, an Android mobile application that uses the developed CNN as back-end to identify Costa Rican tree species from images of cross-sections of wood.

Instance segmentation for the fine detection of crop and weed plants by precision agricultural robots.

PREMISE: Weed removal in agriculture is typically achieved using herbicides. The use of autonomous robots to reduce weeds is a promising alternative solution, although their implementation requires the precise detection and identification of crops and weeds to allow an efficient action. METHODS: We trained and evaluated an instance segmentation convolutional neural network aimed at segmenting and identifying each plant specimen visible in images produced by agricultural robots. The resulting data set comprised field images on which the outlines of 2489 specimens from two crop species and four weed species were manually drawn. We adjusted the hyperparameters of a mask region-based convolutional neural network (R-CNN) to this specific task and evaluated the resulting trained model. RESULTS: The probability of detection using the model was quite good but varied significantly depending on the species and size of the plants. In practice, between 10% and 60% of weeds could be removed without too high of a risk of confusion with crop plants. Furthermore, we show that the segmentation of each plant enabled the determination of precise action points such as the barycenter of the plant surface. DISCUSSION: Instance segmentation opens many possibilities for optimized weed removal actions. Weed electrification, for instance, could benefit from the targeted adjustment of the voltage, frequency, and location of the electrode to the plant. The results of this work will enable the evaluation of this type of weeding approach in the coming months.

A new fine-grained method for automated visual analysis of herbarium specimens: A case study for phenological data extraction.

PREMISE: Herbarium specimens represent an outstanding source of material with which to study plant phenological changes in response to climate change. The fine-scale phenological annotation of such specimens is nevertheless highly time consuming and requires substantial human investment and expertise, which are difficult to rapidly mobilize. METHODS: We trained and evaluated new deep learning models to automate the detection, segmentation, and classification of four reproductive structures of Streptanthus tortuosus (flower buds, flowers, immature fruits, and mature fruits). We used a training data set of 21 digitized herbarium sheets for which the position and outlines of 1036 reproductive structures were annotated manually. We adjusted the hyperparameters of a mask R-CNN (regional convolutional neural network) to this specific task and evaluated the resulting trained models for their ability to count reproductive structures and estimate their size. RESULTS: The main outcome of our study is that the performance of detection and segmentation can vary significantly with: (i) the type of annotations used for training, (ii) the type of reproductive structures, and (iii) the size of the reproductive structures. In the case of Streptanthus tortuosus, the method can provide quite accurate estimates (77.9% of cases) of the number of reproductive structures, which is better estimated for flowers than for immature fruits and buds. The size estimation results are also encouraging, showing a difference of only a few millimeters between the predicted and actual sizes of buds and flowers. DISCUSSION: This method has great potential for automating the analysis of reproductive structures in high-resolution images of herbarium sheets. Deeper investigations regarding the taxonomic scalability of this approach and its potential improvement will be conducted in future work.

Using a Convolutional Siamese Network for Image-Based Plant Species Identification with Small Datasets.

The application of deep learning techniques may prove difficult when datasets are small. Recently, techniques such as one-shot learning, few-shot learning, and Siamese networks have been proposed to address this problem. In this paper, we propose the use a convolutional Siamese network (CSN) that learns a similarity metric that discriminates between plant species based on images of leaves. Once the CSN has learned the similarity function, its discriminatory power is generalized to classify not just new pictures of the species used during training but also entirely new species for which only a few images are available. This is achieved by exposing the network to pairs of similar and dissimilar observations and minimizing the Euclidean distance between similar pairs while simultaneously maximizing it between dissimilar pairs. We conducted experiments to study two different scenarios. In the first one, the CSN was trained and validated with datasets that comprise 5, 10, 15, 20, 25, and 30 pictures per species, extracted from the well-known FLAVIAmathsizesmall dataset. Then, the trained model was tested with another dataset composed of 320 images (10 images per species) also from FLAVIAmathsizesmall. The obtained accuracy was compared with the results of feeding the same training, validation, and testing datasets to a convolutional neural network (CNN) in order to determine if there is a threshold value t for dataset size that defines the intervals for which either the CSN or the CNN has better accuracy. In the second studied scenario, the accuracy of both the CSN and the CNN-both trained and validated with the same datasets extracted from FLAVIAmathsizesmall-were compared when tested on a set of images of leaves of 20 Costa Rican tree species that are not represented in FLAVIAmathsizesmall.

An expert study on hierarchy comparison methods applied to biological taxonomies curation.

Comparison of hierarchies aims at identifying differences and similarities between two or more hierarchical structures. In the biological taxonomy domain, comparison is indispensable for the reconciliation of alternative versions of a taxonomic classification. Biological taxonomies are knowledge structures that may include large amounts of nodes (taxa), which are typically maintained manually. We present the results of a user study with taxonomy experts that evaluates four well-known methods for the comparison of two hierarchies, namely, edge drawing, matrix representation, animation and agglomeration. Each of these methods is evaluated with respect to seven typical biological taxonomy curation tasks. To this end, we designed an interactive software environment through which expert taxonomists performed exercises representative of the considered tasks. We evaluated participants' effectiveness and level of satisfaction from both quantitative and qualitative perspectives. Overall quantitative results evidence that participants were less effective with agglomeration whereas they were more satisfied with edge drawing. Qualitative findings reveal a greater preference among participants for the edge drawing method. In addition, from the qualitative analysis, we obtained insights that contribute to explain the differences between the methods and provide directions for future research.

Toward a large-scale and deep phenological stage annotation of herbarium specimens: Case studies from temperate, tropical, and equatorial floras.

PREMISE OF THE STUDY: Phenological annotation models computed on large-scale herbarium data sets were developed and tested in this study. METHODS: Herbarium specimens represent a significant resource with which to study plant phenology. Nevertheless, phenological annotation of herbarium specimens is time-consuming, requires substantial human investment, and is difficult to mobilize at large taxonomic scales. We created and evaluated new methods based on deep learning techniques to automate annotation of phenological stages and tested these methods on four herbarium data sets representing temperate, tropical, and equatorial American floras. RESULTS: Deep learning allowed correct detection of fertile material with an accuracy of 96.3%. Accuracy was slightly decreased for finer-scale information (84.3% for flower and 80.5% for fruit detection). DISCUSSION: The method described has the potential to allow fine-grained phenological annotation of herbarium specimens at large ecological scales. Deeper investigation regarding the taxonomic scalability of this approach is needed.

Going deeper in the automated identification of Herbarium specimens.

BACKGROUND: Hundreds of herbarium collections have accumulated a valuable heritage and knowledge of plants over several centuries. Recent initiatives started ambitious preservation plans to digitize this information and make it available to botanists and the general public through web portals. However, thousands of sheets are still unidentified at the species level while numerous sheets should be reviewed and updated following more recent taxonomic knowledge. These annotations and revisions require an unrealistic amount of work for botanists to carry out in a reasonable time. Computer vision and machine learning approaches applied to herbarium sheets are promising but are still not well studied compared to automated species identification from leaf scans or pictures of plants in the field. RESULTS: In this work, we propose to study and evaluate the accuracy with which herbarium images can be potentially exploited for species identification with deep learning technology. In addition, we propose to study if the combination of herbarium sheets with photos of plants in the field is relevant in terms of accuracy, and finally, we explore if herbarium images from one region that has one specific flora can be used to do transfer learning to another region with other species; for example, on a region under-represented in terms of collected data. CONCLUSIONS: This is, to our knowledge, the first study that uses deep learning to analyze a big dataset with thousands of species from herbaria. Results show the potential of Deep Learning on herbarium species identification, particularly by training and testing across different datasets from different herbaria. This could potentially lead to the creation of a semi, or even fully automated system to help taxonomists and experts with their annotation, classification, and revision works.