A New Remote Sensing Index for the Detection of Multi-Type Forest Anomalies Based on Sentinel-2 Imagery

Abstract

Forest anomalies (e.g., pests, deforestation, and fires) are increasingly frequent phenomena on Earth’s surface. Rapid detection of these anomalies is crucial for sustainable forest management and development. On-orbit remote sensing detection of multi-type forest anomalies using single-temporal images is one of the most promising methods for achieving it. Nevertheless, existing forest anomaly detection methods rely on time series image analysis or are designed to detect a single type of forest anomaly. In this study, a Forest Anomaly Comprehensive Index (FACI) is proposed to detect multi-type forest anomalies using single-temporal Sentinel-2 images. First, the spectral characteristics of different forest anomaly events were analyzed to obtain potential band combinations. Then, the formulation of FACI was determined using imagery simulated by the LargE-Scale remote sensing data and image Simulation framework over heterogeneous 3D scenes (LESS) model. The thresholds for FACI for different anomalies were determined using the interquartile method and 90 in situ survey samples. The accuracy of FACI was quantitatively assessed using an additional 90 in situ survey samples. Evaluation results indicated that the overall accuracy of FACI in detecting the three forest anomalies was 88.3%, with a Kappa coefficient of 0.84. The overall accuracy of existing indices (NDVI, NDWI, SAVI, BSI, and TAI) is below 80%, with Kappa coefficients less than 0.7. In the end, a case study in Ji’an, Jiangxi Province, confirmed the ability of FACI to detect different stages of pest infection, as well as deforestation and forest fires, using single-temporal satellite images. The FACI provides a promising method for the on-orbit satellite detection of multi-type forest anomalies in the future.

1. Introduction

Forests are an important component of ecosystems, which provide a variety of functions for humans, including ecological, economic, social, and aesthetic benefits [1,2]. The forest damage caused by disturbances has increased due to the intensification of climate change and human activities [3,4]. For example, extreme climate events can lead to frequent forest wildfires and dieback [5,6,7]. Human activities, such as continuous deforestation, have led to a significant reduction in forest area [2]. Monitoring the various damages to the forest ecosystem is important for ecological management and protection [8,9]. Satellite remote sensing has become a crucial method for detecting forest anomalies since it has the advantages of wide imaging coverage and relatively low cost [10]. Nevertheless, traditional detection methods do not meet the requirements of timely monitoring. Most of these methods detect anomalies by comparing multi-temporal images, which requires lots of time to ensure data completeness [11,12]. Until now, the collection and processing of multi-temporal images have been time-consuming. What is more, after forest anomaly events occur, the long processing process generally takes one to four days, from task planning, satellite operation control, and ground data transmission to information extraction [13,14]. In recent years, on-orbit remote sensing detection of forest anomalies has received more and more attention [9,15,16]. By deploying the algorithm directly on the satellite, on-orbit real-time detection can be performed once the satellite image is captured. Then, the detected anomaly information can be transmitted to ground users, significantly shortening the remote sensing response time [9]. Plenty of progress has been made in on-orbit pre-processing technologies, such as geometric correction, radiometric correction, and atmospheric correction [17,18,19]. Therefore, establishing remote sensing indices for forest anomaly detection based on land surface reflectance is a potential way to achieve the on-orbit real-time monitoring of forest disturbances [13,20].

Currently, there are different remote sensing indices for detecting forest anomalies, such as the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) for detecting pine wilt diseases and pests [21,22,23], the Soil Adjust Vegetation Index (SAVI) and Bare Soil Index (BSI) for detecting deforestation, and the Tri-spectral Thermal Anomaly Index (TAI) for detecting forest fires [24,25]. For forest pests, Mantas et al. [26] employed Sentinel-2 data to identify pine (Pinus massoniana Lamb.) wood nematode (Bursaphelenchus xylophilus (Steiner and Buhrer) Nickle.) infection by comparing reflectance images in different seasons. Yu et al. [27] identified different stages of pine wood nematode infection by comparing NDVI variations in pine trees from May to October. Abdullah et al. [28] detected the early stages of bark beetle attack in European spruce using time series images of NDWI. For deforestation, it results in more information from the understory vegetation or bare soil [29]. Most studies use soil-related indices to detect deforestation pixels [30,31]. For example, Klaudia et al. [32] identified deforestation in Białowieza by comparing the SAVI difference between the pre-deforested and post-deforested images. In terms of forest fire, fire algorithms (related to the thermal infrared band) suitable for medium spatial resolution sensors are widely used [33,34]. For forest fires with small coverage, satellite observation with fine spatial resolution is needed [35]. The TAI was proposed to detect small-coverage fires [36,37]. As mentioned above, most indices are designed for specific forest anomaly events. If all indices are deployed on satellites for the comprehensive detection of multi-type forest anomaly events, it can easily result in conflicts in the detection results [13,38]. Therefore, it is valuable to develop a new remote sensing index for on-orbit comprehensive detection of multi-type forest anomalies.

Constructing a new Forest Anomaly Comprehensive Index (FACI) suitable for multi-type forest anomalies from these complex spectra requires addressing two main issues. (1) How to determine the form of the new index based on the common spectral characteristics and ensure the values of the new index follow a specific order for these multi-type forest anomaly events. (2) How to determine the thresholds for the new index in detecting different forest anomaly events. In this study, the form and coefficients of the FACI were determined with the help of the LargE-Scale remote sensing data and image simulation framework over heterogeneous 3D scenes (LESS) model [39,40,41] (see Section 2). The thresholds of FACI were determined based on in-site investigation samples (see Section 3). In addition, the FACI detection results were compared with those of existing indices. The advantages and limitations of FACI were summarized (see Section 4). The main conclusions are given in the end.

2. Method and Materials

2.1. Existing Indices and Potential Forms for Forest Anomalies Detection

Five widely used remote sensing indices for detecting forest anomalies are listed in

Table 1. Five remote sensing indices for forest anomaly detection.

Index	Equation	Anomaly Event	References
NDVI	$\left(\mathrm{B}8\mathrm{a}-\mathrm{B}4\right)/\left(\mathrm{B}8\mathrm{a}+\mathrm{B}4\right)$	pest	[27,42]
NDWI	$\left(\mathrm{B}8\mathrm{a}-\mathrm{B}11\right)/\left(\mathrm{B}8\mathrm{a}+\mathrm{B}11\right)$	pest	[21,28,42]
SAVI	$\left({\displaystyle \frac{\mathrm{B}8\mathrm{a}-\mathrm{B}4}{\mathrm{B}8\mathrm{a}+\mathrm{B}4+0.5}}\right)*\left(1+0.5\right)$	deforestation	[32,43]
BSI	${\displaystyle \frac{\left(\mathrm{B}11+\mathrm{B}4\right)-\left(\mathrm{B}8+\mathrm{B}2\right)}{\left(\mathrm{B}11+\mathrm{B}4\right)+\left(\mathrm{B}8+\mathrm{B}2\right)}}$	deforestation	[30,32]
TAI	$\left({\mathrm{B}}_{12}-{\mathrm{B}}_{11}\right)/{\mathrm{B}}_{8\mathrm{a}}$	fire	[36,37]

below, taking the Sentinel-2 sensor as an example. For pest infection detection, NDVI and NDWI are most commonly used [27,28]. SAVI and BSI are usually applied in detecting deforestation [30,32], while TAI has been specifically developed for fire detection [36]. Each one performs well for specific forest anomalies but may not yield satisfactory results for others.

In this study, the potential band combinations forms for detecting multi-type forest anomalies were determined by analyzing the Sentinel-2 spectral characteristics of different forest anomaly samples. Specifically, the forest sample points were extracted from Sentinel-2 imagery of Heyuan, Guangdong, China, which is a typical region characterized by extensive areas of Pinus massoniana severely affected by pine wood nematode (Bursaphelenchus xylophilus) infestations. These sample points cover healthy forests, early-stage pest infection, medium-stage pest infection, later-stage pest infection, deforested forested, and fire, with 200 sample points for each category. All abnormal reflectance values are removed using the quartile distance method [13] before conducting spectral analysis. Figure 1a shows the reflectance characteristics of pests, deforestation, fires, and healthy forests, while Figure 1b illustrates the spectral characteristics of different stages of pest infection (i.e., early, medium, and late stages).

For healthy vegetation, distinct absorption and reflection peaks are observed in the red band (i.e., B4) and green band (B3), respectively. The value of B4−B3 is always less than zero. However, for the anomaly forest resulting from deforestation and pest infection (medium stage), the reflectance differences between the red and green bands tend to be positive. Therefore, the B4B3 combination could be used to detect healthy forest, pest infection forest, and deforestation. For the fire detection, the shortwave infrared (SWIR) band (B12) is widely used. As shown in Figure 1a, the reflectance in B12 for fire events is significantly higher than that for other types of forest anomaly events. Furthermore, B4−B3 varied with different pest infection stages (later pest > medium pest > early pest > health), while B12 of different pest infection stages had the same order as B4−B3 (Figure 1b). It can be concluded that B4−B3 and B12 can, to some extent, distinguish abnormal forest events from healthy forests.

Here, the potential band combinations were determined base on three rules. The first rule was to select bands previously identified as sensitive to forest anomalies (i.e., B4−B3 term and B12 term). The second rule involved using the summation of B4−B3 and B12 to enhance anomaly information. The third rule required the use of ratio-based normalization to mitigate biases caused by different imaging times and enhance the comparability of the images. Finally, there are three potential band combination forms for comprehensive forest anomaly detection, including Index1: (B4 − B3)/(B4 + B3), Index2: (B4 − B3 + B12)/(B4 + B3 + B12), and Index3: (α × B4 – β × B3 + γ × B12)/(α × B4 + β × B3 + γ × B12). α, β, and γ represent the three coefficients of the red, green, and shortwave infrared bands, respectively. Among them, a larger α value indicates a higher sensitivity to pest infection information, a larger β value indicates a higher sensitivity to healthy forests, and a larger γ value indicates a higher sensitivity to fire.

2.2. Determining the Form of Forest Anomaly Comprehensive Index (FACI) by LESS

In order to evaluate the effectiveness of various potential band combinations, the LESS model was adopted to simultaneously simulate Sentinel-2 images with 12 bands. The simulated scene image contains healthy forests, pest-infested forests at different stages, deforested areas, and burning forests, which was derived from a series of reflectance obtained in spring. The scene (100 × 100 m) has a plane terrain filled with Pinus Nigra (height: 12.2 m, crown width: 5.9 m; see Figure 2a), comprising 2303 health trees, 113 trees in the early stages of infestation, 117 trees with medium pest infestation, and 120 trees in later stages of pest infestation (see Figure 2a). The exposed soil and 114 burning trees were used to characterize deforestation and fire, respectively. Firstly, the reflectance of multi-type forest anomalies and healthy forest could be simulated together within this 100 × 100 m scene using the LESS model. Then, the LESS-simulated images of different bands were used to calculate five existing remote sensing indices (Table 1) and three potential band combination forms, as shown in Figure 2b–i.

To further compare the discrimination of existing indices and different band combinations in identifying multi-type forest anomaly events, the mean and standard deviation of the index values were used for statistical analysis (Figure 3). Due to the continuous variation of index values at early-to-late pest infestation stages, the index values for other anomalies should not overlap with those of the pest infestation stages.

Figure 3a shows the capability of existing indices to distinguish different forest anomalies. The NDVI, SAVI, and BSI could not distinguish between deforestation and fire. The NDWI had difficulty differentiating between later pest infestations and deforestation. What is more, the TAI could only detect fires from other conditions.

Figure 3b shows the capability of three new band combination forms. For Index 1, fire was indistinguishable from deforestation and later-stage pest infestations. Index 2 can effectively distinguish between different degrees of pest infestations, deforestation, and fires, although the values for deforestation and late-stage pest infestations are similar. Index 2 shows great potential in detecting multiple types of forest anomalies by further improving its ability to distinguish between deforestation and later-stage pest infestations.

The final index in identifying different forest anomalies was determined by continuously adjusting the coefficients of the various bands in Index 2. Finally, we found that when α was set to 2.5, β to 3, and γ to 1, the established FACI (i.e., Index 3) had the best performance in distinguishing between forest anomalies (in addition to fire, FACI can also effectively differentiate between varying degrees of pest infection and deforestation from healthy forest), see Equation (1). The FACI values for different forest conditions were in the order fire > deforestation > later pests > medium pests > early pests > healthy forest (Figure 3b).

$$\mathrm{FACI}={\displaystyle \frac{\left(2.5\times B_4+B_{12}\right)-3\times B_3}{\left(2.5\times B_4+B_{12}\right)+3\times B_3}}$$

(1)

2.3. Satellite and In Situ Data

Satellite and in situ data were collected for the Huagai Mountain, Qingyuan District, Ji’an City, Jiangxi Province (Figure 4a). The forest coverage rate in this region reaches 65.5%, predominantly consisting of natural secondary forests. Among these, coniferous forests constitute a significant portion, primarily dominated by Masson pine (Pinus massoniana) and slash pine (Pinus elliottii Engelm). The Sentinel-2 images (Figure 4b,c) of 25 December 2022 (including the healthy forests, pest-infected forests, and deforested area) and 28 February 2022 (including forest fire) were used to calculate the indices for forest anomaly detection. The in situ survey was conducted on 25 December 2022. Serious pine wilt disease occurred in this study area in 2022 (the related trees were cut down thereafter). The sample points for healthy forests, pest infestations, and deforestation were collected through in situ surveys (Figure 4d,e). The fire samples in Figure 4f were identified by visual interpretation of the Sentinel-2 image. In the end, there were 180 healthy, 180 pest-infested, 180 deforested, and 180 burning forest samples. Half of the samples (90 points) for each status were used for threshold determination in Section 3.1, whereas the remaining samples were used for accuracy assessment in Section 3.2.

2.4. Accuracy Assessment Indicators

In this study, a quantitative evaluation was conducted for existing forest anomaly indices and FACI using a confusion matrix. The evaluation indicators include Producer Accuracy (PA), User Accuracy (UA), Overall Accuracy (OA), and Kappa coefficient [44,45,46]. They could be calculated using Equations (2)–(5) as below.

$$P{A}_i={\displaystyle \frac{x_{ii}}{x_{+i}}}\times 100\%$$

(2)

$$U{A}_i={\displaystyle \frac{x_{ii}}{x_{i+}}}\times 100\%$$

(3)

$$OA={\displaystyle \frac{{{\displaystyle \sum }}_{i=1}^k{x}_{ii}}{N}}\times 100\%$$

(4)

$$Kappa={\displaystyle \frac{N{{\displaystyle \sum }}_{i=1}^k{x}_{ii}-{{\displaystyle \sum }}_{i=1}^k(x_{i+}\times x_{+i})}{N^2-{{\displaystyle \sum }}_{i=1}^k(x_{i+}\times x_{+i})}}$$

(5)

where x_ii represents the number of the diagonal elements of the confusion matrix; x_+i indicates the number of true samples in category i; x_i+ represents the number of predicted samples classified as category i; N represents the total number of samples; and k represents the total number of categories [45].

3. Results

3.1. Threshold Determination

Reliable thresholds are essential for quantitatively evaluating the performance of both existing indices and the proposed FACI. However, current forest anomaly indices do not have thresholds for detecting different types of forest anomaly events. For that, this study tried to determine the thresholds based on the separability in identifying various forest anomaly samples. At first, different forest anomaly indices (FACI, NDVI, NDWI, SAVI, BSI, and TAI) were calculated using Sentinel-2 images.

The thresholds were then determined by analyzing the quartiles of the indices for samples under different forest conditions (health, pest, deforestation, and fire). Taking the FACI in Figure 5a as an example, boxplots of different forest conditions were sorted according to ascending medians. The average of the 3rd quartile of healthy forests and 1st quartile of pests, average of the 3rd quartile of pests and 1st quartile of deforestation, and average of the 3rd quartile of deforestation and 1st quartile of fires was selected to distinguish these conditions. Finally, 3 FACI thresholds were obtained for identifying the four categories of forest conditions: 0.054, 0.268, and 0.44, which achieve a Kappa coefficient of 0.878.

The same steps were applied to determine the thresholds of the other indices (see Figure 5b–f): 0.667, 0.46, and 0.287 for NDVI; 0.259, 0.006, and −0.214 for NDWI; 0.333, 0.185, and 0.136 for SAVI; −0.153, 0.043, and 0.248 for BSI; and −0.333, −0.277, and 1.381 for TAI. These thresholds for each index were fairly obtained using the same quartile rule, which facilitated subsequent accuracy comparisons between FACI and existing indices. It can be observed from Figure 5a–f that FACI differentiated the forest conditions more significantly than other indices.

3.2. Accuracy Evaluation and Comparison

The remaining samples for each forest condition were used to assess the accuracy of the indices using the different determination thresholds. As shown in

Table 2. Accuracy evaluation of indices.

	Type	HH *	PT	DF	FE	Total	UA (%)
FACI OA (%): 88.3 Kappa: 0.84	HH	89	3	0	0	92	96.7
	PT	1	70	16	0	87	80.5
	DF	0	17	74	5	96	77.1
	FE	0	0	0	85	85	100
	Total	90	90	90	90	360
	PA (%)	98.9	77.8	82.2	94.4
NDVI OA (%): 71.4 Kappa: 0.61	HH	85	6	1	1	93	91.4
	PT	5	59	23	0	87	67.8
	DF	0	20	44	20	84	52.4
	FE	0	5	22	69	96	71.9
	Total	90	90	90	90	360
	PA (%)	94.4	65.6	48.9	76.7
NDWI OA (%): 76.7 Kappa: 0.68	HH	88	6	0	0	94	93.6
	PT	2	54	41	1	98	55.1
	DF	0	26	49	4	79	62.0
	FE	0	4	0	85	89	95.5
	Total	90	90	90	90	360
	PA (%)	97.8	60.0	54.4	94.4
SAVI OA (%): 66.7 Kappa: 0.56	HH	86	6	1	0	93	92.5
	PT	4	48	40	5	97	49.5
	DF	0	23	33	12	68	48.5
	FE	0	13	16	73	102	71.6
	Total	90	90	90	90	360
	PA (%)	95.6	53.3	36.7	81.1
BSI OA (%): 75.83 Kappa: 0.67	HH	89	9	0	1	99	89.9
	PT	1	47	13	0	61	77.0
	DF	0	34	75	27	136	55.1
	FE	0	0	2	62	64	96.9
	Total	90	90	90	90	360
	PA (%)	98.9	52.2	83.3	68.9
TAI OA (%): 74.2 Kappa: 0.66	HH	55	37	1	0	93	59.1
	PT	20	40	3	0	63	63.5
	DF	15	13	86	4	118	72.9
	FE	0	0	0	86	86	100
	Total	90	90	90	90	360
	PA (%)	61.1	44.4	95.6	95.6

* HH, PT, DF, and FE represent healthy forests, pests, deforestation, and fires, respectively. UA, PA, and OA represent user accuracy, producer accuracy, and overall accuracy, respectively.

, the OA of these indices was in the order of FACI > NDWI > BSI > TAI > NDVI > SAVI.

Table 2 shows that the UA and PA of FACI in detecting healthy forests and fires exceeded 94%. For pest and deforestation detection, FACI had acceptable accuracy, with a UA and PA of approximately 80%. Although the UA and PA of the NDVI for healthy forests exceeded 90%, the UA and PA were below 60%, 70%, and 80% for deforestation, pests, and fires, respectively. The UA and PA of the NDWI for healthy forests and fires exceeded 90% but were both ~60% for pest and deforestation detection. The UA and PA of the SAVI for healthy forests exceeded 90%, while they were 71.6%–81.1% for fire detection and below 60% for pests and deforestation. For BSI, the UA for healthy forests and fire exceeded 89%, and the PA for healthy forests and deforestation exceeded 83%. Nevertheless, both the UA for pest detection and the PA for deforestation were below 60%. The TAI performed well in detecting fires and deforestation but poorly in detecting healthy forests and pests (both PA and UA were 44.4%–63.5%).

In addition, the calculated Kappa coefficients indicated a similar order: FACI > NDWI > BSI > TAI > NDVI > SAVI. It is worth noting that only the Kappa coefficient value for FACI exceeds 0.8, while all other existing indices are below 0.7 (i.e., from 0.56 for SAVI to 0.68 for NDWI).

Overall, the existing indices demonstrated acceptable accuracy in detecting certain forest anomalies but performed poorly in detecting others. Only the FACI exhibited reliable accuracy in detecting multi-type forest anomalies, with an OA of 88.3% and a Kappa coefficient of 0.84.

3.3. Application of FACI-Based Forest Anomaly Detection

In this section, the effectiveness of FACI in detecting multi-type forest anomalies at the regional scale was evaluated. The input images are cloud-free images from July to December 2022, as shown in Figure 6(a1–f1). Figure 6(a1) shows that no significant pest infection or widespread deforestation occurred on this day (taken on 13 July 2022). During the peak growing season in August, Figure 6(b1) shows that some pest infections began to spread within the red ellipse in the image. In September (Figure 6(c1)), sporadic infections have developed into massive infections (as shown in red ellipse). From October to November (Figure 6(d1,e1)), localized infections developed into large areas with severe pest infestations (later stages of pest infestation). In December, the area with severe pest infestations on Huagai Mountain had been deforested, as shown in the red ellipse of Figure 6(f1).

The corresponding FACI-based detection results are shown in Figure 6(a2–f2). The results revealed sporadic pest infestations (ellipse in Figure 6(a2)) in July, with some pest infections (ellipses in Figure 6(b2)) beginning to spread in August. Figure 6(c2–e2) shows that localized pest infestations spread across various regions, rapidly progressing from early to severe stages, which is consistent with the natural color image (Figure 6(c1–e1)). Furthermore, the deforested areas in December were effectively detected by FACI (Figure 6(f2)).

In addition, the effectiveness of forest fire (details in Section 2.3) detection was also evaluated, as shown in Figure 6(g1–g2). It is worth noting that active fire ranges are usually interpreted through false-color images (R: Band 12; G: Band 11; B: Band 8A) [35], as shown in Figure 6(g3). The detection results in Figure 6(g2) have good correspondence to Figure 6(g3), which indicates that FACI accurately identified the extent of forest fires, and it was able to distinguish between deforested areas and pest-infested forests around the fire, as shown in the ellipse of Figure 6(g2).

Meanwhile, existing remote sensing indices were also evaluated in detecting multi-type forest anomalies comprehensively. Taking the image acquired on 25 December 2022 with severe pest infestations and deforestation as an example (Figure 7(a1)), the detection results are shown in Figure 7(a2–a6) (the thresholds were derived in Section 3.1). Figure 7(a2) shows that NDVI misidentified some pest infestation and deforestation areas as fires, as indicated by the red ellipse. The NDWI tended to misidentify deforestation as pest infestations and a small portion of severe pest infestations as fires (Figure 7(a3)). The SAVI performed poorly, misidentifying some areas of pest infestations and healthy forests as forest fires (Figure 7(a4)). The BSI misidentified some pest-infested forests as deforestation areas (Figure 7(a5)). TAI exhibited the worst performance in forest anomaly detection, misidentifying large areas of healthy forests as deforestation (Figure 7(a6)).

In the other example of fire scenario (Figure 7(b1–b6)), NDVI, SAVI, and BSI misidentified deforestation as fires, while NDWI and TAI misidentified large areas of healthy forest as pest infestations. The regional-scale detection results indicate that the existing indices are effective for identifying specific forest anomaly events; however, they are unable to comprehensively detect multi-type forest anomalies.

3.4. The Reliability of the Determined Threshold

To further evaluate the reliability of the selected three FACI thresholds, new thresholds were established by applying ±10% fluctuations around the values determined in Section 3.1. For example, for the threshold1 (0.054) between pest and deforestation, applying a ±10% fluctuation results in two new thresholds of 0.0326 and 0.0754 (i.e., 0.054% ± 10%). Similarly, for Threshold2 (0.268) between pest and deforestation, the ±10% fluctuation results in new thresholds of 0.2508 and 0.2852. For Threshold3 (0.44) between deforestation and fire, the ±10% fluctuation leads to new thresholds of 0.422 and 0.457.

Therefore, there are 27 possible threshold combinations. Threshold1, Threshold2, and Threshold3 for FACI, since each threshold category has three values, as shown in Figure 8. The meanings of Threshold1, Threshold2, and Threshold3 in the threshold combinations correspond to those shown in Figure 5a, representing the distinction between health and pest, pest and deforestation, and deforestation and fire, respectively. These different threshold combinations were used to identify multi-type forest anomaly events (samples in Section 3.1) for the evaluation of the reliability of the selected FACI thresholds. It is noted that the threshold combination 14 is the same as that in Figure 5a.

Finally, the accuracy of 27 threshold combinations was evaluated using the Kappa coefficient, as shown in Figure 9. It can be observed that selecting either the minimum or maximum value for the three thresholds in the combinations does not yield the highest Kappa coefficient. The highest accuracy is achieved when the middle values are selected, with the Kappa coefficient reaching its peak value of 0.878 for threshold combinations 13 and 14. The difference between threshold combinations 13 and 14 lies in the value of Threshold3, which is 0.423 and 0.44, respectively. Considering that the FACI values for fires are significantly higher than those for deforestation, threshold combination 14 may be more universally applicable than threshold combination 13 in practical applications. Threshold combination 14 corresponds to the optimal threshold determined in Section 3.1 of this study, demonstrating that the threshold determined using the quartile method was reliable.

3.5. The Effectiveness of FACI in Other Regions

To test the effectiveness of FACI in detecting forest anomalies in other regions, we used severe pine wilt disease and pest areas in Heyuan, Guangdong, China (Figure 10(a1)), severe pest-infection areas in the Czech Republic (Figure 10(b1)), and fire and pest-infection areas in Chongqing, China (Figure 10(c1)) as the experimental sample areas [47,48,49,50]. The FACI-detected results are shown in Figure 10(a2–c2).

Related studies have indicated that the peak period for the outbreak of pine wood nematodes in the Heyuan region occurs in late September and early October each year [50]. The test image (26 September 2019) shows sporadic pine wood nematode infections on the central island of Heyuan, with some pine trees exhibiting reddening and discoloration (Figure 10(a1)). The FACI accurately identified most of the pest-infected areas (Figure 10(a2)). In the Czech forest anomaly scene, the pine wood nematode infection areas involved both pest infestation and logging (i.e., deforestation), as indicated by the yellow ellipses in Figure 10(b1). It can be observed that the large-area pest infestation and deforestation in Figure 10(b2) were accurately identified by FACI. In the Chongqing scene, the FACI successfully identified the actual fire zones and detected the surrounding pest infestation areas (Figure 10(c1,c2)). Overall, the test results show that FACI had an acceptable performance in other regions.

4. Discussion

4.1. The Advantages of FACI in Multi-Type Forest Anomaly Detection

The FACI demonstrates superior performance compared to existing forest anomaly indices. For pest infection detection, existing remote sensing indices (e.g., NDVI and NDWI) exhibit a serious missed detection, whereas the FACI has a better accuracy with an UA of 80.5%. This improvement may be attributed to the adoption of a pest-sensitive band [42]. In fact, B12 has been employed to investigate vegetation water stress induced by pest infection [51]. The regional-scale test result (Figure 6) indicated that FACI can identify different degrees of pine wilt disease. The primary reason is that chlorophyll in the leaves decreases as pest infestation progresses, resulting in reduced reflectance in the green band and increased reflectance in the red and shortwave infrared bands [26,52]. These reflectance spectral characteristics of different stages of the pine wood nematode were also reported in Japanese pine (Pinus densiflora) and Chinese red pine (Pinus massoniana) [53,54].

For deforestation detection, the FACI achieved a UA of 77.1%, which shows less misidentification than BSI and SAVI. The accuracy of the FACI still requires further improvement when compared to time series studies detection on deforestation. For example, Ang et al. [55] used time series images from Landsat, Sentinel-2, and Sentinel-1 to detect deforestation in the Amazon, which achieved a UA of 90.9% ± 3.3%. Decuyper et al. [56] tracked and quantified deforestation in Peru using Landsat time series satellite images and an anomaly vegetation change detection algorithm, attaining a precision of over 90%. In contrast, FACI utilizes only single-temporal Sentinel-2 optical imagery for deforestation detection, which simplifies the process and facilitates on-orbit remote sensing applications. This method inevitably results in slightly lower detection accuracy due to the limited input of remote sensing information.

For fire detection, FACI performed well, with a UA of 100% and a PA of 94.4%. This performance is comparable to that of the TAI (a specialized wildfire detection index), which has a UA of 100% and a PA of 95.6%. In a similar fire detection study, the Normalized Hotspot Index (NHI) based on Sentinel-2 imagery was utilized to detect active fires in the Arafo–Candelaria area with a precision of 97.2% [57]. It is evident that the FACI can detect active fires using optical reflectance information as accurately as the TAI and NHI. This is because the B12 band (shortwave infrared) is sensitive to high-temperature thermal anomalies. The reflected energy of B12 includes a substantial amount of thermal radiation emitted by high-temperature sources.

It is worth noting that existing studies have also explored the detection of multi-type forest anomaly events. These studies typically assess forest damage using time series imagery analysis. For instance, Candotti et al. [42] accurately utilized supervised classification with Sentinel-2 imagery from 2017 to 2022 to detect forest damage caused by windstorms and pest infestations. Schiller et al. [58] successfully combined transformers with Sentinel-2 time series data to automatically extract forest disturbance information. Ye et al. [59] innovatively used dense time series Landsat 8 images and the “random continuous change” algorithm to detect forest disturbances. These methods are well-suited for long-term forest dynamic change studies but less effective for the rapid on-orbit detection of forest anomalies using single-temporal imagery.

4.2. The Limitation of the FACI

Inevitably, the FACI-based detection method has some limitations for identifyling multi-type forest anomalies. First, the detection of forest pests has shown reliable accuracy for the pine wood nematode, but widespread testing has not yet been conducted for other types of forest pests. Although most pest infestations lead to varying degrees of increased reflectance in the red band and decreased reflectance in the green band [60,61,62], the corresponding FACI values may vary due to the subtle spectral differences among different types of pest infestations. The sensitivity and accuracy of FACI for detecting forest anomalies caused by different types of pests need to be further evaluated in future studies. Second, both the PA and UA of FACI in detecting pests and deforestation were around 80%, indicating that these anomalies may be challenging to distinguish in certain situations. This can be explained by the fact that deforestation detection relies on the enhancement of soil spectral information after deforestation. The greater the intensity of deforestation, the more soil information (e.g., roads and tree crown gaps) is exposed in the deforested areas [63,64]. FACI demonstrates high detection accuracy for large-area deforestation, but it may easily confuse severe pest infestations with sporadic deforestation. Future studies should explore the spectral differences between severe pest infestations (including dead forests) and deforestation events to further improve the performance of FACI. Third, FACI cannot detect forest anomalies caused by natural disasters, such as floods and heavy snowfall, that lead to significant damage to forest resources [65,66]. When these forest anomaly events occur, the spectral characteristics of the forest could be dominated by those of water. For that, the FACI could be further improved by incorporating water-sensitive bands, which would enable the detection of floods.

5. Conclusions

The rapid detection of multi-type forest anomalies is crucial for sustainable forest management and development. Currently, on-orbit remote sensing detection using single-temporal images is one of the most promising methods for achieving it. For that, a new Forest Anomaly Comprehensive Index (i.e., FACI) was proposed for the detection of pests, deforestation, and fires using single-temporal Sentinel-2 images. The form of FACI was determined by spectral analysis and LESS simulation. Three thresholds were determined for FACI to identify healthy forests, pest infestations, deforestation, and fires with the help of in situ surveys. Meanwhile, the accuracy of FACI was evaluated and compared with the existing five remote sensing indices in identifying forest anomalies. The main conclusions are as follows. The FACI achieved an overall accuracy of 88.3% in detecting the three forest anomalies, with a corresponding Kappa coefficient of 0.84. A region-scale study confirmed the FACI’s ability to efficiently detect different stages of pest infection, as well as deforestation and forest fires, using single-temporal images. The FACI demonstrated better performance compared to existing indices that are primarily developed for specific forest anomalies. Overall, FACI is a promising method for detecting multi-type forest anomalies in future real-time on-orbit satellite applications.