数学 · 机器学习 · 作物表型
Mathematics · Machine Learning · Crop Phenotyping
数学与应用数学专业本科生，就读于中国农业大学理学院。日常大多和图像有关：训练检测与分割模型去读玉米的根和叶片，也为科研写那些别人真的跑得起来的工具。几年数学建模竞赛留下一个习惯——先问数据从哪来，再问结论凭什么站得住。工作之外是另一套节奏：魔术社的纸牌、圆号与钢琴、还没整理的田间照片，和一场始终不见起色的篮球。这个网站也是作品之一——它不解决什么问题，但每一行都在。
Applied-mathematics undergraduate at China Agricultural University. Most days are about images: detection and segmentation models that read maize roots and leaves, plus the tooling that lets someone else actually run them. Years of modelling competitions left one habit — ask where the data came from before asking whether the conclusion holds. Off duty, a different rhythm: card magic, horn and piano, unsorted field photographs, and a basketball game that never improves. This website is one of the works too — it solves nothing, but every line of it is here.
把一次真实试验从田间记录一路带到站得住的结论的育种工作台——采、选、洗、算、用五个环节，一条证据链，版本从 V1.0 迭代到了 V13.0。
A breeding workbench that carries a real experiment from field record to defensible conclusion — five stages, one evidence chain, and a version history that now runs from V1.0 to V13.0.
2026 年 6 月下旬起在中国农业大学理学院投入使用的校园智能体：课表、作业、通知、校园事务收进同一个对话框，每个回答都能点回出处。
A campus agent in production at CAU's College of Sciences since June 2026: one conversation box for timetables, deadlines, notices and campus business, with every answer traceable to its source document.
从田间视频里逐株数清气生根的层数。真实遮挡下，单帧检测的准确率只有 36.6–45.1%；本文的遮挡感知多帧融合把株级准确率提到 88.9%。
Counting aerial root tiers per plant from field video. Single-frame detection sits at 36.6–45.1% accuracy under real occlusion; the occlusion-aware multi-frame fusion described here takes plant-level accuracy to 88.9%.
Aerial roots are counted by hand, and counting them from video is unreliable: the roots overlap, occlude each other, and change appearance between frames. This work treats a sequence of captures as evidence rather than as one noisy observation. A cascade — EfficientNet-B3 for tier classification, YOLOv10 for detection, and an occlusion-aware multi-frame fusion step — lifted accuracy from under 50% to 88.9% over a structured set of more than ten thousand plants.
Stripe rust presents as thin, low-contrast lesions against a cluttered canopy — the case where a plain U-Net loses the boundary first. ECV-UNet keeps the U-Net backbone and changes two things: an attention module in the skip connections, and a Transformer block at the bottleneck.
A vanishing element is one on which some irreducible character of the group takes the value zero, and the set of them constrains the group surprisingly tightly. This paper uses that set, together with the prime graph built from it, to give new characterisations of the Chevalley groups G2(3^m).
育种家：一个零代码、拖拽式的育种数据工作台。做了三年，已过校赛，目前在省赛阶段。
育种家 — a zero-code, drag-and-drop workbench for breeding data. Three years of work; through the school round and into the provincial round.
玉米气生根层级识别，准确率从不到 50% 提到 88.9%。第一作者的论文正在按审稿意见修改。
Tier detection for maize aerial roots, from under 50% accuracy to 88.9%. The first-author paper is being revised after review.
用特征标的消失元刻画零散单群与 Chevalley 群。并列第一作者，目前在审。
Characterising sporadic and Chevalley groups through the vanishing elements of their characters. Joint first author; under review.
2026 年推免申请，方向是统计与科学数据。
Applying for 2026 admission, toward statistics and scientific data.
第三周就会学到的那个分解，以及其中没人提的一点：三项里有一项不归你管。
The decomposition everyone learns in week three, and the part of it that nobody mentions: one of the three terms is not yours to control.
一行凸性，三个有用的不等式。写在这里，好让我不用每次做优化论证时都从头搭一遍。
One line of convexity, three useful inequalities. Written down here so I stop rebuilding it from scratch at the start of every optimisation argument.
一本讲测量误差的书的笔记，以及其中一个改变了我做实验方式的观点。
Notes from a book about measurement error, and the one idea from it that changed how I set up experiments.
纸牌魔术是我唯一被允许说谎的地方。它教给我的关于注意力和时机的东西，比任何一门课都多。
Card magic is the only place where I am allowed to be dishonest. It taught me more about attention and timing than any lecture did.
继续读Read more拍的几乎都是植物，而且几乎都在奇怪的时间。下地是带相机最好的借口。
Mostly plants, mostly at odd hours. Field work is a good excuse to carry a camera.
继续读Read more买望远镜的时候还没意识到，这件事有一大半是天气。多数时候看月亮，多数时候是给别人看。
I bought the telescope before I understood how much of it is weather. Mostly the moon, mostly showing it to other people.
继续读Read more圆号是问起来会先报的那一个，钢琴是没人听的时候才弹的那一个。两样都没产出什么值得归档的东西——这也是它们没被算进项目的主要原因。
Horn is the one I will name first if asked; piano is the one I play when nobody is listening. Neither has produced anything worth filing, which is most of why they are not in the project list.
继续读Read more邮件是最可靠的。如果和研究或某个项目有关，把主题写清楚，你会收到认真的回复。
Email is the reliable one. If it is about research or a project, put that in the subject line and you will get a real answer.
