Skip to content
Technology技术

The math inside the engine.引擎背后的数学。

The Verdant Intelligence Platform runs a five-discipline AI stack — perception, retrieval, optimization, forecasting and coordination — that turns raw field data into demand-matched planting decisions, in real time.青源智能平台运行一套五大学科的 AI 栈——感知、检索、优化、预测与协同——将原始田间数据实时转化为与需求精准匹配的种植决策。

AI disciplinesAI 学科
5
Data sources / plot数据源 / 地块
6
Core models核心模型
3
Solver families求解器族
6
The pipeline流水线

Five stages, one decision五段流程,一个决策

Each discipline is a stage in a single pipeline. Perception reads the field, retrieval recalls what worked, optimization allocates planting to demand, forecasting points it at the market, and coordination keeps every tier in lockstep. The output is one decision the platform can act on.每个学科都是同一条流水线上的一段流程。感知读取田间,检索回溯有效经验,优化将种植分配到需求,预测将其对准市场,协同让每一层级步调一致。输出的是平台可直接落地执行的同一个决策。

  1. 01

    Perception

    CV

    Crop, disease, yield & grading from imagery.

  2. 02

    Retrieval

    CBR · kNN

    Reuse the proven crop mix of the most similar villages.

  3. 03

    Optimization

    MILP

    Allocate planting to the demand cap — never overproduce.

  4. 04

    Forecasting

    ML

    Forecast demand on live order data.

  5. 05

    Coordination

    MAS

    Orchestrate every tier of the supply chain.

Data & perception数据与感知

Six sources, fused — then seen六源融合,再被看见

Every plot is described by six fused data sources. On top of that, a computer-vision perception layer reads crop type, disease, yield and grading directly from satellite and UAV imagery — turning pixels into the ground-truth features the engine reasons over.每个地块都由六个融合数据源刻画。在此之上,CV 感知层直接从卫星与无人机影像中读取作物类型、病害、产量与分级——把像素转化为引擎用于推理的真实地表特征。

Climate
Soil
Agronomy
Price history
Live demand
Remote sensing

perception: crop · disease · yield · grading — from satellite & UAV感知:作物 · 病害 · 产量 · 分级 — 来自卫星与无人机

VerdantSense — Field Map · DatongVerdantSense — 田块地图 · 大同

Layers图层

  • Parcels地块
  • NDVI vegetationNDVI 植被
  • Satellite卫星
  • Demand heat需求热力
E 99.9621 · N 26.8901 · z14

Plot A-12

A
Crop作物
Forest herbs林下药材
Area面积
4.2 mu4.2 亩
Suitability适宜度
95%95%
Status状态
Contracted已签约
Suitability适宜度 High Good Moderate Low / fallow低 · 休耕 Jinsha River金沙江
Models & algorithms模型与算法

Three models do the heavy lifting三个模型挑起最重的活

The engine is small and legible by design: retrieve the most similar cases, allocate planting to the demand cap, and forecast where demand is heading. Here is the math, in the open.引擎在设计上小而可读:检索最相似的案例,将种植分配到需求上限,并预测需求的走向。这里把数学完全公开。

Case retrieval

CBR · kNN

Retrieve the k most similar past villages; reuse their proven crop mix as a warm start.

formula公式

sim(q, c) = 1 / (1 + Σ wᵢ (qᵢ − cᵢ)²)

Demand-driven allocation

MILP

Plant to the demand cap — never overproduce. π price · ρ yield · κ cost · A area · D demand.

formula公式

max Σ (E[πⱼ]ρₚⱼ − κₚⱼ) xₚⱼ s.t. Σ xₚⱼ ≤ Aₚ, Σ ρₚⱼxₚⱼ ≤ Dⱼ

Demand forecasting

ML

Forecast season-ahead demand and price from live order data and channel signals.

formula公式

D̂ⱼ,ₜ₊₁ = f(orders, price, season, channel)
The decision loop决策闭环

Decision loop决策闭环

The five disciplines compose into one closed loop — featurize, retrieve, reuse, forecast, optimize, validate, deploy, then retain the outcome so the next village starts warmer. It runs end to end, every season.五大学科组合成一个闭环——特征化、检索、复用、预测、优化、校验、部署,再沉淀结果,让下一个村庄起步时更有底气。它每一季都端到端运行。

Solver lineup求解器阵容

exact MILP (small) · GA · SA · PSO · Tabu Search (large) · learned policy

Exact MILP for small instances; metaheuristics and a learned policy scale the same objective to large ones.小规模实例用精确 MILP;元启发式与学习型策略将同一目标扩展到大规模实例。

verdant://engine/decision_loop
01q = featurize(village)          # data layer02C = Retrieve top-k sim(q,·)     # CBR03ŷ = Reuse(C) -> crop priors04D = Forecast demand             # ML05x* = argmax margin(x; ŷ,D,s,A)  # MILP06while not Validate(x*): Revise  # feedback07deploy order-based planting(x*)08Retain(village, x*, outcome)    # learn
Forecasting · ML预测 · ML

From pixels to a demand curve从像素到一条需求曲线

Remote-sensing vegetation indices and live order data drive season-ahead forecasts — so planting tracks where the market is heading, with anomalies flagged early.遥感植被指数与实时订单数据驱动提前一季的预测——让种植紧跟市场走向,并尽早标记异常。

VerdantForecast — Vegetation & demand analyticsVerdantForecast — 植被与需求分析

NDVI avgNDVI 均值

0.72+0.05+0.05

Demand index需求指数

+18%QoQ环比

Forecast horizon预测窗口

6 moMLML

Anomalies异常

3flagged已标记

NDVI vs demand · season 2026NDVI vs 需求 · 2026 季

NDVIDemand需求

NDVI calendar · 24 weeksNDVI 日历 · 24 周

remote-sensing vegetation index遥感植被指数

lowhigh
Coordination协同

The platform is the broker平台即枢纽

Four agents wrap the chain — plot, aggregator, demand and quality. Each senses its node and streams state to the platform; the platform queries the engine and routes the decision back. Every link runs both ways, in real time.四个智能体环绕整条链路——地块、聚合商、需求与质量。各自感知所在节点并将状态实时回传平台;平台查询引擎,再把决策路由回去。每条链路都双向实时运行。

农户 · 地块

Farmer / plot agent

what each plot can grow

集货 · 分级

Aggregator agent

grading, batching, logistics

The broker枢纽

Verdant Intelligence Platform青源智能平台

Senses every node, queries the engine, routes the decision back — two-way, real-time.感知每个节点,查询引擎,再把决策路由回来——双向、实时。

4 agents · whole-chain coordination4 个智能体 · 全链协同
需求 · 买家

Demand / buyer agent

live orders & price signals

质量 · 溯源

Quality / trace agent

CV grading & traceability

From crop to cluster从单一作物到产业集群

How one crop becomes a whole industry一种作物如何长成一个产业

The engine doesn't stop at planting. Six fusion modes extend a single crop into a Big-Industry cluster and an eco-community — measured along six dimensions of value.引擎并不止步于种植。六种融合模式将单一作物延展为大产业集群与生态社区——并沿六个价值维度加以衡量。

01

Internal-agri fusion

农业内部融合

02

Value-chain extension

产业链延伸

03

Function expansion

功能拓展

04

New-tech infusion

新技术渗透

05

Multi-format composite

多业态复合

06

Agri-town integration

产城融合

Six dimensions of value六大价值维度

GreenClusterDigitalBrandInnovationIntegration
Grounded in research扎根于研究

Architecture grounded in Case-Based Reasoning & Multi-Agent Systems research for agri-food supply chains (Procedia Computer Science 232, 2024).

Research & perspectives研究与观点
Research & perspectives研究与观点
White Paper14 min

Demand-capped planting: an optimization view of rural overproduction

How a demand-cap constraint (MILP) reframes the smallholder planting decision and structurally removes synchronized glut.

Read阅读
Method9 min

Case-Based Reasoning as a moat: why every village compounds the next

Retrieve, reuse, revise, retain — and why a real case base does not copy the way software does.

Read阅读
Field Note11 min

From “one flower” to “one chain”: the Datong deployment

A first-deployment account of pairing AI coordination with on-the-ground operating capability.

Read阅读

拾千农 · 得万村 · Onboard a thousand farmers, reach ten thousand villages.

See the technology in action一套引擎,万千村庄

VerdantSource turns coordination into compounding returns — connecting fragmented supply to real demand, village by village.VerdantSource 青源把协调变成可复利的回报——把分散的供给与真实需求逐村连接起来。