← Knowledge Hub

Artificial Intelligence

Choosing a model by the data you actually have.

Overview

AI in precision agriculture is mostly supervised machine learning: a model learns from labelled examples. The practical question is rarely "which is the newest model" but "how much labelled data and computing do I have".

Key concepts

Train and test split

Always evaluate on images the model has never seen; otherwise accuracy is overstated.

Classical machine learning

SVM, random forest, gradient boosting and KNN work on feature vectors. In our weed study random forest led with fewer than 40 images per class and gradient boosting with more.

Deep learning and YOLO

CNN-based detectors such as YOLO find and locate objects directly in images. They need annotated data, but augmentation can multiply a small annotated set, as in our 800-tile thistle model.

Learning curves

Plot performance against the number of training images to see where more data stops paying off.

Code examples

import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import recall_score

rng = np.random.default_rng(0)

def split_per_class(y, n):
    train = np.concatenate([rng.choice(np.where(y == c)[0], n, replace=False) for c in np.unique(y)])
    return train, np.setdiff1d(np.arange(len(y)), train)

# X: feature matrix (one row per plant image), y: class labels
for n in (5, 10, 20, 40, 80, 160):
    tr, te = split_per_class(y, n)
    clf = RandomForestClassifier(n_estimators=300, random_state=0).fit(X[tr], y[tr])
    print(n, "images/class -> recall", round(recall_score(y[te], clf.predict(X[te]), average="macro"), 3))

Case study

Case studies from my work: a 94.56 % recall weed classifier from 160 images per class, and a YOLO thistle detector.

References from my work