A Novel Algorithm for Detecting Convective Cells Based on H-Maxima Transformation Using Satellite Images

Abstract

Mesoscale convective systems (MCSs) play a pivotal role in the occurrence of severe weather phenomena, with convective cells constituting their fundamental elements. The precise identification of these cells from satellite imagery is crucial yet presents significant challenges, including issues related to merging errors and sensitivity to threshold parameters. This study introduces a novel detection algorithm for convective cells that leverages H-maxima transformation and incorporates multichannel data from the FY-2F satellite. The proposed method utilizes H-maxima transformation to identify seed points while maintaining the integrity of core structural features, followed by a novel neighborhood labeling method, region growing and adaptive merging criteria to effectively differentiate adjacent convective cells. The neighborhood labeling method improves the accuracy of seed clustering and avoids “over-clustering” or “under-clustering” issues of traditional neighborhood criteria. When compared to established methods such as RDT, ETITAN, and SA, the algorithm demonstrates superior performance, attaining a Probability of Detection (POD) of 0.87, a False Alarm Ratio (FAR) of 0.21, and a Critical Success Index (CSI) of 0.71. These results underscore the algorithm’s efficacy in elucidating the internal structures of convective complexes and mitigating false merging errors.

1. Introduction

Mesoscale Convective Systems (MCSs) are fundamental to the development of severe weather phenomena, such as intense precipitation and hail, with convective cells serving as their primary structural components [1,2]. MCSs are currently acknowledged as the predominant organizational form of deep moist convection in both tropical and mid-latitude regions, contributing between 30% and 70% of warm-season rainfall across the contiguous United States, as well as comparable proportions in East Asia, Africa, and South America [3]. These expansive cloud systems, which typically extend horizontally over distances exceeding 100 km and persist for several hours to more than a day, function as key mechanisms for latent heat release and the redistribution of atmospheric mass, moisture, and energy [4]. Within each MCS, individual convective cells—the fundamental building blocks—play a critical role. Each cell is characterized by a vigorous, nearly vertical updraft with a horizontal scale ranging from approximately ten to several tens of kilometers and vertical velocities frequently exceeding 10 m s⁻¹, often spanning from the boundary layer to the tropopause [5]. These strongly buoyant plumes entrain moist, conditionally unstable air, facilitating rapid ascent to cloud tops and ultimately leading to the generation of heavy precipitation, hail, damaging winds, and, in certain instances, tornadoes [6].

A convective cell is generally characterized by a robust updraft that extends horizontally over a scale ranging from approximately ten to several tens of kilometers, exhibiting vertical velocities on the order of 10 m/s and reaching nearly the full vertical extent of the troposphere [7,8,9,10]. These convective cells represent active regions within MCSs where convective dynamics occur, comprising air masses engaged in organized convective motion. The primary objective of convective cell detection methodologies is to accurately identify the most intense zones of convective activity. Satellite-based stereoscopic retrievals from the FY-4A and Himawari-8 rapid-scan imagers confirm that cloud-top cooling rates ≥ 1 K min⁻¹ (equivalent to vertical growth > 5 m s⁻¹) are reliable proxies for these intense cores [11,12]. Both Rapid Developing Convection (RDC) detection and Convection Initiation (CI) detection rely fundamentally on the identification of convective cells, as these processes inherently involve locating areas characterized by convective cell presence [13,14,15,16].

In recent years, numerous methodologies have been developed to identify convective cells utilizing satellite observations [17,18,19]. To address the highly irregular characteristics of convective clouds, data assimilation techniques have been employed for detection and tracking of these cells, although such approaches are computationally demanding [20]. The Source Apportionment (SA) method examines life cycle evolution of convective cells by leveraging sequences of satellite imagery to track and forecast MCSs [21]. The TOOCAN (Tracking of Organized Convection Algorithm Through a 3-D Segmentation) algorithm iteratively processes infrared satellite images to achieve three-dimensional segmentation of convective system volumes [22]. However, its application in real-time operational contexts remains challenging. A comprehensive study on the mechanisms underpinning the persistence and life cycle of MCSs was conducted, employing both conventional observational data and analytical simulations [23]. An optical flow technique was utilized to analyze the motion of deep convective clouds in GOES-16 ABI imagery, with the objective that a semi-Lagrangian model for cloud field advection not relying on the identification and tracking of individual cloud elements be developed [24]. The merging processes of convective cells and their influence on lightning activity during a severe squall line event over the Beijing Metropolitan Region were investigated, integrating data from the Beijing Broadband Lightning Network (BLNET), S-band Doppler weather radar, and additional meteorological datasets [8]. The latest Integrated Multisatellite Retrievals for Global Precipitation Measurement (IMERG) satellite precipitation products were applied to detect and monitor MCSs across Europe over a sixteen-year period [25]. The characteristics of convective cells and their association with both convective and stratiform precipitation were explored during a season-long Weather Research and Forecasting (WRF) simulation permitting convection over central Argentina [26].

The Enhanced Thunderstorm Identification, Tracking, and Nowcasting (ETITAN) methodology employs a dual-threshold strategy to achieve more precise detection of convective cells relative to the original TITAN algorithm [27]. Nevertheless, due to the rapid morphological changes characteristic of convective cells, this approach may occasionally yield inaccurate identification outcomes. To preserve the internal structural integrity of thunderstorms while effectively differentiating adjacent storms within a cluster, the ETITAN method—grounded in mathematical morphology—addresses the issue of erroneously merging proximate thunderstorms through a sequence of procedures: (1) an initial identification is conducted using a single minimum threshold, which serves as the foundation for an erosion operation designed to separate merged storms into two distinct entities, thereby mitigating false mergers and clearly delineating neighboring thunderstorm units; (2) a secondary, elevated threshold is applied to detect more intense thunderstorms; and (3) erosion and dilation operations are subsequently performed on each identified thunderstorm, facilitating the effective separation of adjacent storms. Consequently, this process yields two independent thunderstorms within a cluster while maintaining their internal structural features. In contrast, the Rapid Developing Thunderstorm (RDT) method, developed in France, utilizes statistical decision-making frameworks and empirical criteria to identify robust convective systems [28]. Recognized as a leading technique for convective core detection, the RDT method leverages convective monitoring data collected across Europe at 15-min intervals to support short-term forecasting. This approach determines infrared channel thresholds based on variations in the vertical extent of convective fluid and employs an adaptive thresholding mechanism. The detection of convective cells within this framework is contingent upon the satisfaction of five specific conditions: (1) T_cold, the coldest bright temperature threshold, typically set at −55 °C; (2) T_warm, the warmest temperature threshold for the convective fluid, usually ranging from −10 to 5 °C; (3) ΔT, the temperature contour interval between the two fluids in the vertical direction, set at 1 °C; (4) ΔT_tower, the minimum temperature difference required to identify fluid extremes, which is 3 °C; (5) A_min, the minimum area threshold for fluid discrimination, generally corresponding to the size of a single image element in the infrared satellite cloud image. The ETITAN approach utilizes mathematical morphology operations, including erosion and dilation, to effectively distinguish closely situated thunderstorms. Conversely, the RDT method employs an adaptive thresholding strategy to identify convective systems in a straightforward and efficient manner. The convective cell detection technique proposed in this study integrates these two methodologies, leveraging both mathematical morphology and thresholding to accurately discriminate between adjacent convective cells.

In summary, the detection of convective cells entails isolating individual convective cells from larger convective cloud systems while excluding cirrus cloud components. A commonly employed method relies on two parameters: brightness temperature (Tb) and spatial area, utilizing a simple thresholding technique for segmentation. Nevertheless, this thresholding approach is highly sensitive to the chosen threshold value, often leading to unreliable extraction of core convective profiles. Several limitations are associated with this method: (1) it is challenging to distinguish adjacent convective cells, which may be erroneously merged and identified as a single entity; (2) accurately delineating individual convective cells within fragmented cloud regions is difficult, frequently resulting in the inclusion of extensive portions of the cloud along with neighboring clouds exhibiting lower brightness temperatures. If the brightness temperature threshold is set excessively high, the detected convective cell may conflate with surrounding colder cloud regions; conversely, if the threshold is too low, the detected cell may be incomplete. Given that average Tb values vary across different convective cells, employing a uniform threshold for detection proves insufficient for accurately discriminating between adjacent cells.

To address these challenges, this study proposes a novel algorithm named Convective Cell detection utilizing H-maxima Transformation (CCHT). The approach commences with the identification of seed points for convective cells through the application of the H-Maximum transformation technique. Subsequently, proximate seed points are systematically grouped and labeled according to a connectivity domain neighborhood labeling method, which effectively distinguishes adjacent seed clusters. Finally, convective cells are delineated using a region growing approach, incorporating a merging criterion that permits seed points to expand or merge with neighboring seed points, thereby forming the definitive convective cell structures.

2. Data and Methods

2.1. FY-2F Satellite Images

Experimental dataset was sourced from the China Meteorological Administration (CMA) utilizing the FY-2F meteorological satellite. This satellite was launched on 13 January 2012, from the Xichang Satellite Launch Center (XSLCC) and is geostationarily positioned above the equator at 112° E longitude. FY-2F satellite is equipped with five spectral channels spanning various wavelength ranges; among these, four infrared channels provide a spatial resolution of 5 km, whereas the visible channel offers dual spatial resolutions of 1 km and 5 km, with the 5 km resolution specifically employed for daytime meteorological observations. Given that natural disasters such as storms predominantly occur in China during summer months, experimental data selected for analysis correspond to this season, with measurements recorded at six-minute intervals.

2.2. H-Maximum Transformation Technique

The proposed CCHT algorithm primarily utilizes H-Maximum transformation to assess the merging and splitting of paired seed points and to decide whether the paired seed clusters should be combined, with the objective of distinguishing adjacent convective cells. The H-Maximum transformation is a mathematical morphology technique intended to eliminate insignificant structures while preserving the fundamental shape characteristics of an image [29,30]. This method offers advantages such as rapid processing and is commonly applied in image segmentation and feature extraction tasks. Specifically, the technique suppresses all local maxima in a grayscale image that are less than or equal to a specified threshold h, which corresponds to the brightness temperature in infrared imagery. The H-Maximum transformation is formally defined as follows:

$$H_{\max }(X)=R_{ \mathrm{f}}^{\delta }(X-h)$$

(1)

where X represents pixel value in original image, and h signifies threshold, $R_{ \mathrm{f}}^{\delta }$ denotes dilate reconstruction. The H-Maximum transformation technique aids in preventing the separation of extreme points situated within the same convective cell. In the visible spectrum, extreme points within a convective cell can be manifested as cloud shadows on the cloud top or the brightly illuminated regions. In the IR band, despite the quick saturation of the IR optical depth, cloud edges often exhibit distinct characteristics due to the temperature gradients, which can be considered as extreme points in the context of convective cell analysis. Consequently, employing the H-Maximum transformation method for detection of convective cells can effectively avoid the fragmentation of extreme value points and enhance the differentiation between adjacent extreme value points.

A regional extremum can be identified by computing the difference between the pixel value of original image and the corresponding value derived from H-Maximum transformation. The regional extremum is formally defined as follows:

$$R_{\max }(X)=X-R_{ \mathrm{f}}^{\delta }(X-h)$$

(2)

where X denotes original image, and $R_{ \mathrm{f}}^{\delta }(X-h)$ represents H-Maximum transformation.

The identification of seed points by the H-Maximum transformation method commences with the determination of a threshold value, denoted as h. This threshold is established through extensive experimentation involving repeated testing across all five spectral channels of FY-2F satellite imagery. The results indicate that the most precise detection of convective cell seed points is achieved at a threshold value of 0.03, thereby designating 0.03 as the optimal parameter. Subsequently, the seed points corresponding to convective cells are derived utilizing the H-Maximum transformation approach.

2.3. Neighborhood Labeling Method

The 8-connected neighborhood criterion is proposed to cluster adjacent seed points and mark them in order. The marked seed points are called seed clusters. Pixel markers in the same connected domain are the same, so that different seed clusters are distinguished. Suppose that the labeled image is represented by the matrix LA, each element LA (p) represents label of the pixel p, the foreground pixel set of image is represented by F, and LA (p) satisfies two conditions: (1) If $p\in F^c$, $\mathit{LA}\left(p\right)=0$; If $p\in F$, $\mathit{LA}\left(p\right)=\infty$. (2) If and only if $p\in F$ and $q\in F$. In the same connected domain, $\mathit{LA}\left(p\right)=\mathit{LA}\left(q\right)$; else, $\mathit{LA}\left(p\right)\ne \mathit{LA}\left(q\right)$.

The labeling process is as follows. First, the initialization matrix is set to

$$L{A}^0(p)=\{\begin{array}{cc}\infty & \forall p\in F\\ 0& \forall p\in F^c\end{array}$$

(3)

Then, at the iteration time t > 0, the iterative expression of the matrix is

$$L{A}^{\left(t\right)}(p)=\min \left\{L{A}^{\left(t\right)}\right(p);\lambda ;L_{\min }(p\left)\right\}$$

(4)

in which

$$L_{\min }(p)=\underset{\mathit{kl}}{\min }\{L{A}^{(t-1)}(q)|q=(x_p+k,y_p+l)\in F\}$$

(5)

where $\lambda$ denotes a connected domain counter, and integers k and l are defined in the neighborhood of the connected domain.

The neighborhood labeling method can ensure that the extracted seed points are not only continuous but also distinguish different convective cells. The proposed method primarily utilizes H-Maximum transformation to assess the merging and splitting of paired seed points, as well as to decide if the paired seed clusters should be combined to distinguish between adjacent convective cells.

Liu et al. (2014) [31] pioneered the use of H-maxima Transform for cloud-core detection. However, the present study is not a simple replication. We have made three substantive innovations:

1.

Optimization of Region Labeling and Merging Mechanisms: Introduction of “Novel Neighborhood Labeling Method” and “Adaptive Merging Criterion”.

(1) Liu et al. (2014) [31] follows a relatively simple workflow: “extended maxima transform for seed generation → counting merger/split times → fixed merger criterion for merging decisions”. The rigid merging criterion lacks flexibility, leading to merging errors in complex convective scenarios (e.g., dense adjacent cells).

(2) The present study optimizes this workflow by adding a “novel neighborhood labeling method”, which improves the accuracy of seed clustering and avoids “over-clustering” or “under-clustering” issues of traditional neighborhood criteria. Additionally, it upgrades the merging mechanism to an “adaptive merging criterion”, dynamically adjusting merging decisions based on the actual features of convective cells (e.g., core temperature differences, boundary distances, structural similarity) instead of relying on fixed rules. This significantly enhances the ability to distinguish adjacent convective cells and effectively solves Liu et al. (2014)’s [31] core problem of “poor distinction between adjacent cells under anvil cloud interference”.

2.

Expansion of Data Dimension: From “Single-Channel (Infrared)” to “Multi-Channel”.

(1) Liu et al. (2014) [31] explicitly relies only on geostationary satellite infrared data for detection. The single data dimension makes it vulnerable to the inherent limitations of the infrared band (e.g., interference from anvil clouds, blurred boundaries), restricting the accuracy of distinguishing adjacent cells.

(2) The present study incorporates multi-channel data from the FY-2F satellite. By fusing complementary information from multiple bands (e.g., infrared, visible light), it enriches the feature dimensions of convective cells-capturing more comprehensive details such as structure, temperature, and texture. This upgrade at the data source level enhances detection robustness and compensates for the information deficiency of single-channel data in the Liu et al. (2014) [31].

3.

More Comprehensive Performance Validation and Comparison: Multi-Algorithm Benchmarking and Quantitative Metrics.

(1) Liu et al. (2014) [31] only qualitatively demonstrates “effective distinction of adjacent convective cells” through “case studies in China” and “experimental results on infrared images”. It lacks comparisons with mainstream algorithms and quantitative performance indicators, limiting its persuasiveness.

(2) The present study provides quantitative performance metrics (Probability of Detection (POD) = 0.87, False Alarm Ratio (FAR) = 0.21, Critical Success Index (CSI) = 0.71) and conducts horizontal comparisons with three mainstream convective cell detection algorithms (RDT, ETITAN, SA), clearly proving its superior performance over existing methods. Furthermore, it quantifies the algorithm’s advantages by concluding that it “reduces false merging errors” and “clearly reveals the internal structure of convective complexes”, making the validity of its innovations more rigorous and intuitive compared to Liu et al. (2014)’s [31] qualitative validation.

The procedure of the CCHT algorithm is given below (Algorithm 1):

Algorithm 1: Convective Cell detection utilizing H-maxima Transformation (CCHT)

Input:       Satellite image data X Output:       Convective cells C Initialization:    Set Tb threshold T = 241 K, i is adaptive threshold.    while Tb > T          Tb = 0    end while Detection:    Compute the seed-points according to (1). Threshold is set to 0.03. for I = 1 to i do    threshold=threshold + 0.01    Accumulate other seed points and compute the number of merger m. end Cluster the adjacent seed points utilizes neighborhood labeling method. Decide pair-wise seed clusters whether should be merged.

3. Results and Discussion

3.1. Satellite Data Preprocessing

The dataset employed in this study consists of FY-2F satellite observations acquired over a nine-month period from June to August 2013–2015, totaling 6180 files. The satellite records data at hourly intervals, featuring a spatial resolution of 5 km and encompassing five spectral channels: Infrared (IR)1, IR2, IR3 (Water vapor), IR4, and Visible (VIS) (See

Table 1. Wavelength and Spatial Resolution of FY-2F satellite imager radiometric channels.

Channel	Wavelength (μm)	Spatial Resolution (km)	Used
IR1	10.3–11.3	5	√
IR2	11.5–12.5	5	√
IR3	6.3–7.6	5	√
IR4	3.5–4.0	5
VIS	0.55–0.90	1, 5	√

). The FY-2F satellite’s full-disk observation is not a continuous high-frequency scan; instead, a complete image of the Earth is captured every hour during the non-flood season and every half hour during the flood season. However, the regional rapid-scan mode offers one image every 6 min, which is stored as 6-min interval files.

Following extraction of IR1 data from FY-2F satellite dataset, we enhanced its quality through a series of procedures including image denoising, enhancement, and the detection and correction of missing lines. The enhancement process entails modifying the minimum and maximum grayscale values observed. This adjustment improves the visibility of particular features in the satellite images, like the edges of convective cells. The refined IR1 data subsequently served as input for the algorithm. Each input image measures 200 × 200 pixels and represents cloud top brightness temperature. Figure 1 illustrates an IR1 image (2148 UTC on 5 August 2013) from the dataset utilized by the proposed CCHT algorithm, a number of convective cells can be seen in the image, which contains different phases of MCSs lifecycle including genesis, maturity and dissipation.

3.2. Case Study

A sequence of experiments was performed to validate the algorithms utilizing FY-2F satellite data acquired from June to August 2013–2015. Validation set was built by combining Doppler radar composite reflectivity ≥35 dBZ at 6-min resolution. As illustrated in Figure 2, the results indicate that the adaptive detection algorithm for mesoscale convective cells successfully identifies erroneous linkages between convective cells. Moreover, the algorithm demonstrates a strong capacity to differentiate individual convective cells within a convective complex, thereby offering a more detailed characterization of the complex’s internal structure. As shown in Figure 2b, blue indicates convection without lightning (35–40 dBZ), white indicates small-scale convection with lightning (40–45 dBZ), and red indicates convection with lightning (≥45 dBZ). All colored regions in Figure 2b represent convective areas (radar reflectivity ≥ 35 dBZ). Colors distinguish lightning association and convective scale, not non-convective regions.

The coldest cloud tops correspond to the minimum Tb values observed in IR imagery, and these minimum Tb pixels closely resemble the seed points identified through the proposed CCHT algorithm. To group adjacent seed points, an eight-connected neighborhood labeling criterion was employed. Additionally, a merging criterion was established to determine whether pairs of seed clusters should be combined. The results of experiments conducted on real meteorological datasets demonstrate the efficacy of the proposed algorithm in accurately characterizing the core regions of observed convective cells.

Figure 3 presents an illustration of application of the CCHT algorithm for identification of convective cells, with each image having dimensions of 200 × 200 pixels. The figure comprises four images, depicting convective cells highlighted in red, as determined by the CCHT method. Seed clusters identified within the images may either expand or merge with adjacent clusters. The previously described merger criterion is employed to ascertain the occurrence of either expansion or merging. Notably, the region growing criterion, which governs the decision-making process regarding the merging of pairwise seed clusters, is fundamental to the effective growth and amalgamation of these clusters.

To illustrate the appearance of correctly identified convective cells, false positive cells, and radar-undetected cells, we use Figure 3a to generate Figure 4a, which highlights correctly identified cells in red and false positives in dark green. Additionally, Figure 4b (based on Figure 3b) shows radar events that the CCHT algorithm failed to detect, marked in green.

3.3. Comparison of CCHT with Other Methods

This section employs three quantitative metrics—Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI)—to evaluate the performance of various methods. POD, FAR and CSI are contingency-table skill scores that originated in meteorological forecast verification [32]. Metrics (POD, FAR, CSI) were computed using CINRAD radar data (1 km spatial resolution, 6 min temporal resolution) as the reference, where a convective storm was defined as radar echo intensity ≥35 dBZ at 1 km altitude. The mathematical expressions defining these evaluation metrics are presented below.

$$\mathrm{POD}={\displaystyle \frac{h}{h+m}}$$

(6)

$$\mathrm{FAR}={\displaystyle \frac{f}{f+h}}$$

(7)

$$\mathrm{CSI}={\displaystyle \frac{h}{h+m+f}}$$

(8)

where h represents count of correctly identified convective cell, f indicates the number of non-convective cell mistakenly identified as convective, and m refers to number of actual convective cell that were not detected. The metrics POD, FAR, and CSI range from 0 to 1, elevated values of POD and CSI correspond to increased predictive accuracy, whereas a reduced FAR value similarly signifies enhanced prediction reliability [33,34]. The ideal detection performance is characterized by a high probability of detection coupled with a low probability of false alarms.

Data collected from June to August 2023, with a temporal resolution of six minutes and a spatial resolution of five kilometers, were employed to evaluate the performance of the methods under investigation. The three-month (June–August) evaluation period was selected to align with the peak convective season in East Asia, ensuring the dataset captures >80% of annual convective events and provides a statistically robust sample for algorithm validation. The dataset used to validate the methodology is the FY-2F geostationary satellite S-VISSR (Stretched Visible and Infrared Spin Scan Radiometer) dataset [35]. The FY-2F S-VISSR dataset is generated through a standardized preprocessing pipeline for geostationary satellite data. A total of 6280 samples were selected for this assessment. A temporal resolution of six minutes facilitates a more detailed capture of temporal dynamics, whereas a spatial resolution of five kilometers provides sufficient spatial coverage and accuracy.

The imager IR pixel resolution is an important factor for our algorithm. While the FY-2F data with a 5 km IR pixel resolution served our initial research needs, the FY-4 series AGRI imager, with its 4 km IR resolution and more frequent scans, has the potential to improve the detection of smaller convective cells. Future research could explore the application of FY-4 data to further enhance the performance of our algorithm.

The comparative results of the different methods are presented in

Table 2. Assessment results from different methods.

Methods	h	m	f	POD	FAR	CSI
RDT	20,215	8157	7213	0.71	0.26	0.57
ETITAN	24,627	7865	8021	0.76	0.25	0.61
SA	32,548	5613	9457	0.85	0.23	0.68
CCHT	34,782	5197	9246	0.87	0.21	0.71

. The proposed CCHT algorithm incorporates H-Maximum transformation technique within its morphological framework and introduces a neighborhood labeling method to distinguish between adjacent convective cells. In contrast, the RDT, ETITAN, and SA methods do not utilize such a criterion to improve detection accuracy. The neighborhood criteria of RDT, ETITAN, and SA often fail to account for the spatial heterogeneity of convective cells (e.g., edge temperature gradients, core compactness), leading to over-clustering or under-clustering. Consequently, the CCHT method exhibits superior performance relative to other approaches, achieving a POD of approximately 0.87, a FAR near 0.21, and a CSI around 0.71.

Collectively, these findings highlight that the CCHT method not only improves the probability of accurately identifying genuine convective activity but also decreases the incidence of false alarms, thereby establishing it as a more dependable instrument for operational convective monitoring and nowcasting.

For the specific case study shown in Figure 3a, the contingency values are: hits (h) = 84, misses (m) = 16, and false alarms (f) = 19. This results in a POD of 0.84, FAR of 0.18, and CSI of 0.70 for this individual snapshot, illustrating the typical performance captured in the aggregate statistics in

Table 3. Assessment result of CCHT method for Figure 3.

Method	h	m	f	POD	FAR	CSI
CCHT	84	16	19	0.84	0.18	0.70

. These values are roughly consistent with the overall performance of CCHT in the sample dataset. The slight difference is due to the missed weak cell, which highlights future optimization directions (e.g., dynamic adjustment of Tb thresholds based on regional atmospheric conditions).

4. Conclusions

This study introduces CCHT, an innovative H-maxima transformation-based algorithm developed for detection of convective cells, addressing the limitations inherent in traditional threshold-based approaches, such as their sensitivity to parameter selection and inability to distinguish adjacent cells effectively. By integrating H-maxima transformation with a novel neighborhood labeling method, region growing and adaptive merging criteria, CCHT effectively isolates convective cells while preserving their spatial integrity. Evaluation using FY-2F satellite data from June to August 2013 and August 2023 revealed that CCHT outperforms existing methods, including RDT, ETITAN, and SA, in identifying individual cells within mesoscale convective systems, as evidenced by improved POD and CSI metrics, attaining a POD of approximately 0.87, a FAR close to 0.21, and a CSI of approximately 0.71. Future work will aim to adapt CCHT for real-time operational use and to evaluate its applicability with forthcoming geostationary satellite platforms.