TREE DETECTION IN ORCHARDS FROM VHRSATELLITE IMAGES USING SCALE-SPACE

THEORY

MILAD MAHOURVALENTYN TOLPEKIN

ALFRED STEIN

(M.MAHOUR@UTWENTE.NL)SEPTEMBER 26, 2016

PRECISION AGRICULTURE AND ORCHARDS § Precision agriculture management

v Fieldv Orchard

Ø Individual trees

§ Precision irrigationv Save water, energy and moneyv Optimal management strategyv Efficient decision makingv RS technologies

§ Challenges v Environment: Efficient use of water

v Science: Tree crown boundary detection2

IRRIGATION ORCHARD MANAGEMENT

§ Crop Water requirement: § Crop coefficient

§ Tree type§ Tree cover fraction

§ Remote sensing images:§ Coarse and fine spatial resolution§ Spatial pattern: shape, location, tree species, tree crown size

3

STUDY AREA AND MATERIAL

§ The orchards:• Orchard 1: Peach• Orchard 2: Walnut

§ RS images:§ Worldview-2§ UltraCam digital image

4

INDIVIDUAL TREE DETECTION FROM RS IMAGES

§ Remote sensing images:• Coarse resolution: Improper to detect tree crown boundary• Fine resolution: Not sufficient spectral and geometrical detail• Very High Resolution (VHR): Adjacent tree canopies interlocks

§ Image processing techniques:• Region-based image segmentation• Template matching: Not the same size all trees• Valley following: Local minima pixels surrounding tree crowns• Local maxima: trees and their background with similar contrast

5

OBJECTIVE

§ To detect individual trees from VHR satellite images on complex orchards integrating scale-space theory.

§ What kind of information do we need?• Shape• Size• Spatial pattern

§ How to use results to investigate the application of scale-space theory for tree detection from RS data?

§ How to validate the results?

6

TREE DETECTION

7

SCALE-SPACE THEORY

§ Gaussian scale-space representation (Lindeberg, 2008), as:𝐿(𝑥,𝑦;𝑠)= 𝑔(𝑥,𝑦;𝑠)∗𝑓(𝑥,𝑦)

Input image: 𝑓(𝑥,𝑦)Gaussian kernel: 𝑔 𝑥, 𝑦; 𝑠 = *

+,-𝑒/ 01231

145

If 𝑠 = 0 à 𝐿 𝑥7, 𝑦7;0 = 𝑓(𝑥, 𝑦)

§ Grey-level blob:• Bell shaped intensity profile

• A pair of local maximum and minimum

8

SCALE-SPACE BLOB DETECTION

§ The grey-level blobs at each scale level§ Linking grey-level blobs as higher order objects:

• Blob selection with detℋ as blob detector

detℋ𝐿 𝑥, 𝑦; 𝑠 = 𝑠+ 𝐿>>𝐿?? − 𝐿>?+

§ Local maximum in scale-space• Point with 𝑥, 𝑦, 𝑠 is maximum along three directions• Position 𝑥, 𝑦• Size of blob (𝑟) à 𝑟 = 2𝑠

9

HOW DOES SCALE-SPACE WORK?

§ How to treat scale levels?

• Min 𝑠 = Min diameter• Max 𝑠 = Max diameter

10

Min diameter

(m)

Max diameter

(m)

Orchard 1 1.0 3.2Orchard 2 3.4 16.4

UNCERTAINTY ASSESSMENT§ Object accuracy indicator:

• True positives• Type I error: False positives• Type II error: False negatives

§ Total detection error• Overestimated• Underestimated

§ Positional accuracy (E)

11

SCALE-SPACE RESULTS OF ORCHARD 1

§ Total: 519§ Detected: 498 (95.7%)§ TP: 468 (90.0%)§ FP: 29 (5.5%)§ FN: 47 (9.0%)

§ Overestimated: 0.45§ Underestimated: 0.07§ Total error: 0.49§ E: 0.24 m

12

SCALE-SPACE RESULTS OF ORCHARD 2

§ Total: 500§ Detected: 462 (92.4%)§ TP: 408 (81.6%)§ FP: 55 (11.0%)§ FN: 59 (12.0%)

§ Overestimated: 0.21§ Underestimated: 0.24§ Total error: 0.25§ E: 1.03 m

13

DISCUSSION

§ Automatic tree detection à supervised Vs. unsupervised§ Controlling scale variation: If scale increases§ Benefit from traditional image analysis à user contribution§ Commercial VHR satellite images Vs. UAV§ Importance of tree detection§ Integration with:

• Crop health • Crop water requirement• Monitoring trees

§ Shape of blobs

14

CONCLUSIONS

§ Meaningful delineation of individual trees§ Qazvin is a semi-arid region

• Traditional irrigation network• Agricultural product for Tehran (70%)

§ Scale-space theory provide automatic way for tree detection• Different tree sizes• Spatial variation• Limited prior information

§ Results are in good agreement as compared with reference data§ Further improvements:

• Saddle point study: prevent multiple detection of same tree§ Flexibility of modeling: scientists, farmers and decision makers

15

THANK YOU!