PhenoNet Platform

To make PhenoNet available to a wide community, a user-friendly and extensible web platform was developed. Four key online functions were proposed, including:

1. wheat phenophase prediction using PhenoNet
2. phenology prediction of other kinds of crops using self-uploaded networks
3. training new networks based on PhenoNet architecture but private datasets using transfer learning or not
4. remote prediction through PhenoNet APIs.

🍂 Wheat phenophase prediction using PhenoNet

This function enables automatic and accurate phenophase prediction from uploaded wheat images by utilizing the powerful capabilities of PhenoNet or other SOTA DL networks. The web platform proposes an interactive page for users to upload wheat images and extract corresponding phenology stages. The prediction results are saved in the database and can be viewed online or downloaded in CSV format to the local disk. This function greatly simplifies the process of analyzing wheat phenology information and saves time and resources.

🍁 Phenology prediction of other kinds of crops using self-uploaded networks.

This function enables users to upload networks that were trained on other datasets to predict the phenology stage of other crops. The platform constructed a web form for uploading DL networks with ONNX format and testing images. The self-uploaded network will be assigned an individual key and can be selected during application. Overall, integrating a self-uploaded network into the PhenoNet platform can improve the platform's applicability in more crops.

🌳 Training new networks based on PhenoNet architecture but private datasets using transfer learning or not

This function enables new network training using the architecture of PhenoNet or other SOTA networks with the self-uploaded dataset, and the network can be initialized randomly (by default) or initialized with weights pre-trained on the WheatPheno (i.e., transfer learning). The PhenoNet platform presents an interactive form for users to submit a training task by uploading datasets and selecting a network for training. After submitting a training task, it is added to the processing queue. Once the training task is complete, the platform automatically saves the best weight with the lowest validation loss to the database and generates a link for users to download it to the local disk. Overall, this function simplifies the application process of PhenoNet and WheatPheno on crop phenology research. Users can adopt the flexible and powerful PhenoNet platform for DL-based phenology analysis of more crops.

🌾 Remote prediction through PhenoNet APIs

This function enables users to invoke the above functions without accessing web pages after applying for an authorization key in the web platform. For example, to predict the phenology stage by invoking the API through Python, users can package their image data and authorization key in a dictionary format. Then, they can upload the dictionary to the API endpoint using the HTTPS POST method to submit the prediction task. Once the prediction is completed, the status code and phenology stage will be returned in JSON format. Detailed interface documentation and demo code were proposed for users to employ PhenoNet APIs (https://help.phenonet.org/phenoapi/). PhenoNet APIs provide more flexible ways of extending their capabilities. For example, it can be integrated into resource-constrained devices in the field to achieve real-time phenology analysis.