GPS Coordinates of 18,482 Crop Fields in East Africa with Improved Accuracy using Planet Imagery and Yolo v5 Object Detection Model

Georeferenced crop yield prediction is a valuable tool for agronomists and policymakers. One challenge with many existing datasets is that of location accuracy. GPS locations for fields can end up offset from the true location due to sensor inaccuracies or from locations being collected at the edges of fields rather than the field centers. This makes it harder to connect remote-sensed data to the yield values. The goal of this project was to produce a method that can help correct these location offsets by finding the most probable field center given an input location. We prepared and hosted a competition on Zindi (https://zindi.africa) where competitors model the problem using state-of-the-art data science techniques. We provided the competitors with satellite images of fields along with their corresponding manually annotated correct centers. Additionally, we also provided approximate plot size and measured yield in case these help with creating their solutions. Original positions are considered images' centers as (0,0) and a displacement vector for each field in the training set was provided. The goal of the competition was to predict these vectors for each vector in the test set. This dataset includes the locations of 18,481 crop fields across Kenya, Tanzania, and Rwanda, collected in 2016-2017 with mixed qualities and their error-corrected ones from the winning solution using Planet satellite imagery and the Yolo v5 object detection model.<br> <br>

地理配准作物产量预测（Georeferenced crop yield prediction）是农学家与政策制定者的重要辅助工具。当前多数现有数据集面临的核心挑战之一便是定位精度（location accuracy）问题。由于传感器存在误差，或是点位采集于田块边缘而非中心，农田的GPS定位结果往往会与真实位置产生偏移，这会大幅增加遥感数据（remote-sensed data）与产量数值进行关联的难度。

本项目旨在开发一种方法，通过给定的输入点位推算最可能的田块中心，以此修正上述定位偏移。我们在Zindi平台（https://zindi.africa）筹备并举办了一场竞赛，要求参赛者采用前沿数据科学技术对该问题进行建模。我们为参赛者提供了农田卫星影像及其对应的人工标注精准中心，此外还提供了近似地块面积（approximate plot size）与实测产量（measured yield），以辅助其构建解决方案。

训练集（training set）中的原始点位以影像中心作为(0,0)坐标，并为每个田块提供了位移向量（displacement vector）。本次竞赛的目标为预测测试集（test set）中每个样本对应的位移向量。

本数据集涵盖肯尼亚、坦桑尼亚与卢旺达境内总计18481块农田的点位信息，采集时间为2016至2017年，数据质量参差不齐；同时包含了基于Planet卫星影像与Yolo v5目标检测模型的获胜方案所生成的误差修正点位。