**4. Discussion**

#### *4.1. Overcoming Observation Gaps in Mangrove Monitoring*

In the critical review by Younes Cárdenas et al. (2017) based on 55 recent peer-reviewed journal articles using Landsat or ASTER images to monitor mangrove forests, the authors conclude that the majority of multi-temporal studies focus on only a fraction of available satellite imagery with on average 7–11 years between scenes of multitemporal analysis [13]. Yet, high temporal change detection of mangrove forests is vital in mapping threats from aquaculture expansion and coastal development as well as to understand cyclic processes such as logging in production areas and seasonal biomass fluctuations [15]. In our study, 446 unique scenes were processed into annual median composites to study the temporal dynamics of mangrove forests in one of Vietnam's most prominent regions for mangrove conservation, Ngoc Hien district in the Ca Mau province. Despite a large number of scenes per composite (15.4 scenes on average), it still resulted in a high number of pixels with no observations and/or illogical land use transitions for di fferent years between 2001–2019 (See Table 1). Following temporal gap-filling and post-classification optimization, the resulting optimized time series allowed us to better monitor the state of mangrove forests in Ngoc Hien with spatial and temporal continuity and logical consistency in transitions.

#### *4.2. Post-Classification Temporal Optimization*

Accurate land use classification remains a challenging undertaking in landscapes dominated by aquaculture and mangrove forest land use [13]. In our study, the accuracy assessment for the reference year 2015 yielded assuring accuracy confiners (Table 3). We used 514 single-year reference data points observed in-situ for training the classifier and validation of the results. Ideally, we would have such data available for multiple years; however, this is di fficult to acquire and organize. With limited availability of ground-truth data, reducing classification uncertainties and increasing temporal consistency is key to provide high-quality land use maps.

The use of yearly median composites allowed for fast computational processing and a comprehensible annual time series output for policy and decision-makers. Nevertheless, the application of median composites also poses disadvantages. Adequate composites still require the presence of sufficient high-quality observations. Moreover, a year-round even temporal distribution of scenes is required to facilitate an adequate median composite that is representative for the entire year. Knowledge regarding yearly phenology, the impact of tides of reflectance signals, trends in biophysical parameters (functional traits) of mangrove ecosystems is for a large part still lacking to make appropriate judgments on possible biases in median composites [13]. In other words, gaining more understanding regarding these temporal processes, also within yearly cycles, will help gain insight into the robustness of median composites for mangrove forest ecosystems.

Further challenges in accurate classifications of mangrove forests are raised by; (1) the fine-grained landscape mosaic with mangrove plots and aquaculture ponds often sized at sub-pixel (30 m × 30 m) measures, (2) the unknown implications of tidal e ffects on spectral signals and the high level of water vapor observed in these coastal regions, and (3) recent trends towards integrated mangrove-shrimp farming production systems which have made discrimination between mangrove, aquaculture paddy land use classes more ambivalent [9,13,15,52,53]. These challenges highlight the importance of making the most of the temporal information available to lower uncertainties in the final classification product. This is particularly important when noise in remote sensing signals is high—which is commonly the case in cloud-covered mangrove regions—and when multi-annual validation/training data are scarce. Several situations can cause illogical changes in a land use time series, such as classification errors, reflectance signal noise, and imperfect image co-registration.

Here, we demonstrated how the exploitation of information in the temporal domain allowed for additional optimization and a means of cross-validation of the GEE classification outputs. Studying temporal patterns and cross-validating land use changes in relation to the previous and following years help increase the robustness to noise and credibility of classification efforts. Specifically, the temporal information of land cover maps in a time series has been used to detect illogical land use changes and improve classification results [48]. The approach for temporal mangrove monitoring outlined is relatively easy to implement using GEE output and post-classification optimization. At the same time, it provides valid spatial classification (Table 3; 94–96% accuracy) and temporal interpolation (Table 4; 87 to 92% accuracy). Temporal interpolation follows a simple discrete inverse distance weighting algorithm, however other and more advanced statistical learning approaches can potentially be interesting alternatives [41,49,50,54–56].

Figure 5 illustrates the further implementation of our gap-free time series in studying multi-temporal mangrove land use trends in Ngoc Hien. Bi-temporal approaches risk highlighting observations that result from isolated instances that can introduce inconsistencies and uncertainties in classification, especially considering the, at times, unfavorable signal-to-noise ratios found in satellite remote sensing [13]. Instead of comparing two single timestamps, the integration of yearly gap-free land use classification enables temporal cross-validation along logical land use transitions and gap-filling based on temporally neighboring information. This temporally dense analysis allows us to fully assess the direction and frequency of land use changes affecting mangrove forests [57,58]. This is also important to ensure that the unchanged forest between two time points has remained undisturbed in the years in between. Moreover, the multi-annual approach allows us to assess and quantify observed land use changes, e.g., forest disturbances, at a high temporal frequency, thereby opening venues to better monitor and study mangrove forest disturbance regimes and mangrove degradation processes as compared to bi-temporal land use change.

The land use trend map (Figure 5) facilitates the detection of hotspots for mangrove forest change; deforestation, degradation, and regrowth. While the map is informative for deforestation and the land use change drivers behind it, the assessment of forest degradation remains arbitrary inherent to the operationalization of the land use classification scheme. This ties in with the challenges and complexity of defining forest degradation [59]. A large variety of existing definitions of forest degradation require different methods for assessment based on the objectives of the intended study [60]. The gap-free land use change maps presented here may help flag changes in mangrove forest extent. However, to overcome the arbitrariness of classes, complimentary maps on quantitative canopy/leaf traits and biomass can further enhance our ability to assess forest degradation and forest regrowth along a spectrum of ecologically relevant indicators [61–63].
