Identification of coffee agroforestry systems using remote sensing data: a review of methods and sensor data

Figures & data

Figure 1. Process of the systematic review of literature for studies of coffee production areas.

Figure 2. Types of coffee agroforestry systems defined for the review.

Table 1. Classification of AFS mapping approaches.

Criteria	Category	Attributes
Learning technique	Supervised classification	Land use classes are defined by the user. There are enough reference data that are used as training samples for the generation of thematic maps.
	Unsupervised classification	The information within the images is grouped into various spectral classes. The user labels and merges the generated classes according to the object of study.
Analysis unit	Per subpixel	The value of each pixel is considered a linear or non-linear combination of different spectral response and there is a proportional membership of the pixel to different classes.
	Per pixel	The algorithms combine the pixel information according to a defined characteristic, ignoring mixed pixel conflicts.
	Object-oriented	Images are segmented by aggregating pixels into objects, and classification is done based on these objects instead of classifying individual pixels.
Classifier types	Parametric	They use statistical parameters (covariance matrix, mean vector, etc.) from training sites. A Gaussian distribution is generally assumed, for example, Maximum Likelihood.
	Non-parametric	No assumptions are made about the data and it does not use statistical parameters, for example, KNN, CART, RF, ISODATA, and SVM.

Figure 3. Map of the studies selected by country.

Table 2. Case studies by type of AFS found.

Type of AFS	Rustic	Polyculture (shade $>$45%)	Polyculture (shade $<$45%)	Monoculture (shade $>$45%)	Monoculture (shade $>$45%)
Number of case studies	6	91	96	49	47

Figure 4. Algorithms identified for the classification of coffee agroforestry systems. ML = Maximum likelihood; RF = Random Forest; KNN = K-Nearest neighbor; CART = classification and regression tree; SMA = Spectral Mixture Analysis; ANN = artificial neuronal network; ISOSEG = per-field clustering classifier; OBIA = object-based Image Analysis; SVM = Support Vector Machine; LMNL = Multinomial Logit; ISODATA = iterative Self-Organizing data Analysis; ECHO = extraction and classification of homogeneous objects; MD = Minimum distance.

Table 3. Case studies by type of mapping approaches.

Criteria	Category	Number of case studies
Learning technique	Supervised classification	138
	Unsupervised classification	23
Analysis unit	Per subpixel	7
	Per pixel	121
	Object-oriented	33
Classifier types	Parametric	46
	Non-parametric	115

Table 4. Frequency of identified satellite images.

Type of satellite images	Sensor	Number of case studies
Low-resolution imagery ($>$20 m)	IRS - Resourcesat LISS IV	1
	IRS – LISS III	1
	Landsat series	118
	MODIS	7
Medium-resolution imagery (5-20 m)	SPOT series	12
	Sentinel 2	11
	Terra ASTER	4
High-resolution imagery ($<$5 m)	UAV / Aerials	19
	Quickbird	11
	WorldView 1 y 2	2
	IKONOS	15
	Cartosat 1	1
	GeoEye	5
Radar imagery (C-Band)	Sentinel 1	138
	RADARSAT	16

Table 5. Auxiliary data used in case studies.

Type of auxiliary data	Number of case studies
Land use maps	10
Soil type maps	3
Topographic maps, DEM, relief maps	30
Other maps	9
Statistical data, producer interviews, and field information	32
Texture descriptors	44
Vegetation indices	36
Without auxiliary data	94

Table 6. Case studies classified by reported accuracy evaluation metrics.

Reported accuracy evaluation metrics	Global accuracy only	Global accuracy and Kappa index	Overall accuracy, Kappa index and class accuracy
Number of studies	6	12	143

Figure 5. Distribution of accuracies reported by type of agroforestry system. (a) Global accuracy, (b) producer accuracy and (c) user accuracy.

Table 7. Analysis of variance by type of AFS and producer accuracy reported.

Multiple comparison after Kruskal-Wallis test
Chi-square = 82.524
P-value = $<$2.2$\times e^{-16}$
AFS compared	Observed difference	Critical difference	Difference
Monoculture (shade $<$45%) – Monoculture (shade $>$45%)	99.02	73.14	Significant
Monoculture (shade $<$45%) – Polyculture (shade $>$45%)	59.40	67.51	No significant
Monoculture (shade $<$45%) – Polyculture (shade $>$45%)	41.82	68.46	No significant
Monoculture (shade $<$45%) – Rustic	147.28	143.72	Significant
Monoculture (shade $>$45%) – Polyculture (shade $>$45%)	158.41	55.71	Significant
Monoculture (shade $>$45%) – Polyculture (shade $>$45%)	140.84	56.86	Significant
Monoculture (shade $>$45%) – Rustic	48.26	138.56	No significant
Polyculture (shade $<$45%) – Polyculture (shade $>$45%)	17.57	49.40	No significant
Polyculture (shade $<$45%) – Rustic	206.67	135.68	Significant
Polyculture (shade $>$45%) – Rustic	189.10	136.15	Significant

Figure 6. Producer accuracy for mapping shaded monocultures (shade $<$45%) by algorithm.

Figure 7. Analysis of variance and producer accuracy by (a) whether it use training data, (b) analysis unit and (c) whether it use statistical parameters for mapping monocultures (shade $<$45%).

Figure 8. Producer accuracy for mapping shaded monocultures (shade $<$45%) by sensor type.

Figure 9. Auxiliary information used for the classification of monocultures (shade $<$45%).

Table 8. Analysis of variance by spatial resolution and producer accuracy achieved for mapping of monocultures (shade $<$45%).

Multiple comparison after Kruskal-Wallis test
Chi-square = 1.98
P-value = 0.5766
AFS compared	Observed difference	Critical difference	Difference
High resolution – low resolution	1.70	14.43	No significant
High resolution – medium resolution	4.36	20.50	No significant
Low resolution – medium resolution	2.66	19.52	No significant

Figure 10. Reported producer accuracy for coffee monoculture (shade $>$45%) by algorithm type.

Figure 11. Analysis of variance and producer accuracy by a) type of classification, b) analysis unit and c) whether it use statistical parameters for mapping coffee monocultures (shade $>$45%).

Figure 12. Reported producer accuracy for monocultures (shade $>$45%) by sensor type.

Table 9. Analysis of variance by spatial resolution of the images used and produces accuracy achieved for monocultures (shade $>$45%).

Multiple comparison after Kruskal-Wallis test
Chi-square = 5.0634
P-value = 0.1672
AFS compared	Observed difference	Critical difference	Difference
High resolution – low resolution	1.99	27.70	No significant
High resolution – medium resolution	35.57	46.27	No significant
High resolution – radar	2.70	30.38	No significant
Low resolution – medium resolution	33.58	40.29	No significant
Low resolution – radar	0.71	20.15	No significant
Medium resolution – radar	32.88	42.18	No significant

Figure 13. Auxiliary information used for the classification of monocultures (shade $<$45%).

Figure 14. Distribution of producer accuracy values reported for polyculture mapping by algorithm.

Figure 15. Distribution of producer accuracy values reported for polyculture mapping by algorithm type: a) by supervised or unsupervised classification, b) by analysis unit, and c) whether it is possible to adjust algorithm parameters.

Figure 16. Kruskal-Wallis analysis of variance for supervised and unsupervised classification for mapping polycultures with different shade coverage percentages.

Figure 17. Kruskal-Wallis analysis of variance for object-oriented and per-pixel classifications for mapping polycultures with different shade coverage percentages.

Figure 18. Kruskal-Wallis analysis of variance for parametric and non-parametric algorithms for mapping polycultures with different shade coverage percentages.

Figure 19. Producer accuracy reported for polycultures by sensor type.

Figure 20. (a) Producer accuracy reported by type of auxiliary data used and (b) number of case studies by type of auxiliary data used for polycultures (shade $<$45%); and (c) global accuracy reported by type of auxiliary data used and (d) number of case studies by type of auxiliary data used for polycultures (shade $>$45%).

Figure 21. Producer accuracy reported for mapping of rustic systems by a) algorithm, b) by supervised or unsupervised classification, c) by analysis unit and d) whether it is possible to adjust algorithm parameters.

Table 10. Analysis of variance of parametric and non-parametric algorithms and producer accuracy reported for mapping of rustic AFS mapping.

Multiple comparison after Kruskal-Wallis test
Chi-square = 11.345
P-value = 0.0007563
AFS compared	Observed difference	Critical difference	Difference
Non-parametric – parametric	35.46	29.53	Significant

Figure 22. Producer accuracy for mapping rustic systems by type of auxiliary data used.