PLANET/AI4COPERNICUS

This data set provides sample input and output data of the developed PLANET tool (hyPer Local climAte driven Tool) for the three use cases investigated (Kenya, Papua New Guinea, Puerto Rico.) funded by the AI4COPERNICUS project. The capabilities of the tool include A Generative Adversarial Network (GAN) that downscales the climate information and a feedforward neural network that classifies the land suitability for specific crops based on the refined input data. There are four steps, each with a dedicated folder: STEP1: Input data set (seasonal climate data, Reanalysis data, Digital elevation model data). STEP2: Downscaled seasonal climate data sets for each region with GANs. Each region is accompanied by the Generator Loss, the Peak signal to noise ratio (PSNR) in html formats, the output images during training by the generator (training.png) and a png with the downscaled meteorological fields. STEP3: Sample input data sets (FAO STAT) of yield data for maize and NDVI that are imported to a LSTM network for yield prediction. Output are yield prediction in CSV format, the lstm structure in png and the Mean Square Error of LSTM training over the epochs. Caution is needed as this step is not mandatory to Land Suitability in STEP4 but act as layer to investigate relationships of NDVI with yield prediction. FAO STAT data only provide crude yield data and user input to the model are more appropriate. STEP4: Land Suitability input and output data sets. Besides the meteorological data, SoilGrids data are utilised and a Feed forward network makes a classification ranging from not suitable at all to highly suitable. Output for each region include the confusion matrix in png and the suitability in CSV format for each region. The accuracy of the model is found in the file fnn_acc.html and the structure of the model in FNN_structure.png.本数据集为AI4COPERNICUS项目资助开发的PLANET工具（hyPer Local climAte driven Tool，超局地气候驱动工具）提供了针对三个研究案例（肯尼亚、巴布亚新几内亚、波多黎各）的样本输入输出数据。该工具具备两项核心功能：一是通过生成对抗网络（Generative Adversarial Network，GAN）实现气候信息降尺度，二是借助前馈神经网络（Feedforward Neural Network，FNN）基于精细化输入数据对特定作物的土地适宜性进行分类。本数据集分为四个步骤，每个步骤设有专属文件夹：步骤1：输入数据集，涵盖季节气候数据、再分析数据与数字高程模型数据。步骤2：基于生成对抗网络生成的各区域降尺度季节气候数据集。每个区域配套包含生成器损失、峰值信噪比（Peak Signal to Noise Ratio，PSNR）的HTML格式文件，生成器训练过程输出图像（training.png）以及展示降尺度气象场的PNG图像。步骤3：玉米产量与归一化植被指数（Normalized Difference Vegetation Index，NDVI）的样本输入数据集（FAO STAT），将此类数据导入长短期记忆网络（Long Short-Term Memory，LSTM）以完成产量预测。本步骤的输出结果包括CSV格式的产量预测数据、PNG格式的LSTM网络结构文件，以及各训练轮次下LSTM训练的均方误差相关数据。需注意，本步骤并非步骤4土地适宜性分析的强制环节，仅作为探究NDVI与产量预测关系的辅助分析图层。此外，FAO STAT仅提供粗略的产量数据，用户自行向模型输入的自定义数据更为适配。步骤4：土地适宜性输入与输出数据集。除气象数据外，本步骤还使用了SoilGrids土壤格点数据集，通过前馈神经网络完成分类任务，分类等级覆盖从完全不适宜到高度适宜的区间。各区域的输出结果包含PNG格式的混淆矩阵，以及各区域土地适宜性的CSV格式文件。模型精度信息存储于文件fnn_acc.html中，模型网络结构则存储于FNN_structure.png文件中。