1-km pantropical shifting cultivation map

Pantropical Shifting Cultivation Map — Model Weights and Example Data This deposit contains the trained model weights and example data for the paper: A pantropical map of shifting cultivation at 1 km resolution using deep learning and PlanetScope imagery [Author list, Journal, Year] Contents bestF1.pkl — Trained EfficientNet-B1 classifier weights (PyTorch). The model was trained on 4-band (R, G, B, NIR) PlanetScope NICFI basemap patches at 234×234 pixels to classify five tropical land-use types: high tree-cover woodland, shifting cultivation, conventional agriculture, mixed woody plantation and others. Weights correspond to the epoch with the highest validation macro-F1 score. samples_example.zip — A set of example PlanetScope training patches in .npy format (channel-first, 4×234×234, float32) illustrating the input data format expected by the model. sample_csv_demo.csv — Example label CSV showing the required column format for training: path (file path to patch), label (land-use class string), useCase (train/test), augProb (per-sample augmentation probability for weakly supervised learning), id (site identifier for spatial cross-validation splitting). Related resources Code: https://github.com/ywant/shifting_cultivation_1km Model architecture: EfficientNet-B1 with ImageNet pre-trained weights, 4-channel input, 4-layer dense classification head Input imagery: NICFI PlanetScope basemaps accessed via the Planet NICFI programme Map outputs: to be prepared Reproduce the results Download bestF1.pkl, place it at saved_models/bestF1.pkl, and follow the instructions in the GitHub repository README to run inference or reproduce the accuracy assessment.泛热带刀耕火种地图——模型权重与示例数据 本数据集存档包含对应下述论文的训练模型权重与示例数据： 《基于深度学习与PlanetScope(PlanetScope)影像的1公里分辨率泛热带刀耕火种地图》[作者列表、期刊、年份] ## 内容清单 1. bestF1.pkl：基于PyTorch(PyTorch)框架训练完成的EfficientNet-B1(EfficientNet-B1)分类器权重文件。该模型以234×234像素的4波段（红、绿、蓝、近红外）PlanetScope(PlanetScope) NICFI(NICFI)底图斑块作为训练数据，可对5类热带土地利用类型进行分类：高树木覆盖林地、刀耕火种、常规农业、混合木质种植园及其他类型。此权重对应验证集宏F1评分最高的训练轮次。 2. samples_example.zip：一组.npy格式的示例PlanetScope(PlanetScope)训练斑块数据，采用通道优先布局，维度为4×234×234，数据类型为float32，用于展示模型所需的输入数据格式。 3. sample_csv_demo.csv：示例标签CSV文件，展示了模型训练所需的列格式规范："path"（斑块文件路径）、"label"（土地利用类别字符串）、"useCase"（训练/测试集划分）、"augProb"（弱监督学习场景下的单样本增强概率）、"id"（用于空间交叉验证拆分的站点标识符）。 ## 相关资源 - 代码仓库：https://github.com/ywant/shifting_cultivation_1km - 模型架构：采用预训练于ImageNet数据集的EfficientNet-B1(EfficientNet-B1)，输入为4通道，搭配4层全连接分类头 - 输入影像：通过Planet NICFI计划获取的NICFI(NICFI) PlanetScope(PlanetScope)底图 - 地图产出：待筹备 ## 结果复现 下载bestF1.pkl并将其放置于`saved_models/bestF1.pkl`路径下，随后按照GitHub仓库README文档中的说明运行推理任务或重现精度评估流程。